Introduction to Data Compression?
Guy E. Blelloch Computer Science Department Carnegie Mellon University
blellochcs.cmu.edu
September 25, 2010
Contents
1
Introduction 2 Information Theory 2.1 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 The Entropy of the English Language . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Conditional Entropy and Markov Chains . . . . . . . . . . . . . . . . . . . . . . . 3 Probability Coding 3.1 Pre?x Codes . . . . . . . . . . . . . . . . . 3.1.1 Relationship to Entropy . . . . . . 3.2 Huffman Codes . . . . . . . . . . . . . . . 3.2.1 Combining Messages . . . . . . . . 3.2.2 Minimum Variance Huffman Codes 3.3 Arithmetic Coding . . . . . . . . . . . . . 3.3.1 Integer Implementation . . . . . . . 4 Applications of Probability Coding 4.1 Runlength Coding . . . . . . 4.2 MoveToFront Coding . . . . 4.3 Residual Coding: JPEGLS . . 4.4 Context Coding: JBIG . . . . 4.5 Context Coding: PPM . . . . .
?
3 5 5 6 7 10 10 11 13 15 15 16 19 22 25 26 27 28 29
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
This is an early draft of a chapter of a book I’m starting to write on “algorithms in the real world”. There are surely many mistakes, and please feel free to point them out. In general the Lossless compression part is more polished than the lossy compression part. Some of the text and ?gures in the Lossy Compression sections are from scribe notes taken by Ben Liblit at UC Berkeley. Thanks for many comments from students that helped improve the presentation. c 2000, 2001 Guy Blelloch
1
5 The LempelZiv Algorithms 32 5.1 LempelZiv 77 (Sliding Windows) . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5.2 LempelZivWelch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 6 Other Lossless Compression 37 6.1 Burrows Wheeler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 7 Lossy Compression Techniques 7.1 Scalar Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Vector Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Transform Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 41 41 43
8 A Case Study: JPEG and MPEG 44 8.1 JPEG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 8.2 MPEG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 9 Other Lossy Transform Codes 9.1 Wavelet Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Fractal Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3 ModelBased Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 50 52 55
2
1 Introduction
Compression is used just about everywhere. All the images you get on the web are compressed, typically in the JPEG or GIF formats, most modems use compression, HDTV will be compressed using MPEG2, and several ?le systems automatically compress ?les when stored, and the rest of us do it by hand. The neat thing about compression, as with the other topics we will cover in this course, is that the algorithms used in the real world make heavy use of a wide set of algorithmic tools, including sorting, hash tables, tries, and FFTs. Furthermore, algorithms with strong theoretical foundations play a critical role in realworld applications. In this chapter we will use the generic term message for the objects we want to compress, which could be either ?les or messages. The task of compression consists of two components, an encoding algorithm that takes a message and generates a “compressed” representation (hopefully with fewer bits), and a decoding algorithm that reconstructs the original message or some approximation of it from the compressed representation. These two components are typically intricately tied together since they both have to understand the shared compressed representation. We distinguish between lossless algorithms, which can reconstruct the original message exactly from the compressed message, and lossy algorithms, which can only reconstruct an approximation of the original message. Lossless algorithms are typically used for text, and lossy for images and sound where a little bit of loss in resolution is often undetectable, or at least acceptable. Lossy is used in an abstract sense, however, and does not mean random lost pixels, but instead means loss of a quantity such as a frequency component, or perhaps loss of noise. For example, one might think that lossy text compression would be unacceptable because they are imagining missing or switched characters. Consider instead a system that reworded sentences into a more standard form, or replaced words with synonyms so that the ?le can be better compressed. Technically the compression would be lossy since the text has changed, but the “meaning” and clarity of the message might be fully maintained, or even improved. In fact Strunk and White might argue that good writing is the art of lossy text compression. Is there a lossless algorithm that can compress all messages? There has been at least one patent application that claimed to be able to compress all ?les (messages)—Patent 5,533,051 titled “Methods for Data Compression”. The patent application claimed that if it was applied recursively, a ?le could be reduced to almost nothing. With a little thought you should convince yourself that this is not possible, at least if the source messages can contain any bitsequence. We can see this by a simple counting argument. Lets consider all 1000 bit messages, as an example. There are 21000 different messages we can send, each which needs to be distinctly identi?ed by the decoder. It should be clear we can’t represent that many different messages by sending 999 or fewer bits for all the messages — 999 bits would only allow us to send 2999 distinct messages. The truth is that if any one message is shortened by an algorithm, then some other message needs to be lengthened. You can verify this in practice by running GZIP on a GIF ?le. It is, in fact, possible to go further and show that for a set of input messages of ?xed length, if one message is compressed, then the average length of the compressed messages over all possible inputs is always going to be longer than the original input messages. Consider, for example, the 8 possible 3 bit messages. If one is compressed to two bits, it is not hard to convince yourself that two messages will have to expand to 4 bits, giving an average of 3 1/8 bits. Unfortunately, the patent was granted. 3
Because one can’t hope to compress everything, all compression algorithms must assume that there is some bias on the input messages so that some inputs are more likely than others, i.e. that there is some unbalanced probability distribution over the possible messages. Most compression algorithms base this “bias” on the structure of the messages – i.e., an assumption that repeated characters are more likely than random characters, or that large white patches occur in “typical” images. Compression is therefore all about probability. When discussing compression algorithms it is important to make a distinction between two components: the model and the coder. The model component somehow captures the probability distribution of the messages by knowing or discovering something about the structure of the input. The coder component then takes advantage of the probability biases generated in the model to generate codes. It does this by effectively lengthening low probability messages and shortening highprobability messages. A model, for example, might have a generic “understanding” of human faces knowing that some “faces” are more likely than others (e.g., a teapot would not be a very likely face). The coder would then be able to send shorter messages for objects that look like faces. This could work well for compressing teleconference calls. The models in most current realworld compression algorithms, however, are not so sophisticated, and use more mundane measures such as repeated patterns in text. Although there are many different ways to design the model component of compression algorithms and a huge range of levels of sophistication, the coder components tend to be quite generic—in current algorithms are almost exclusively based on either Huffman or arithmetic codes. Lest we try to make to ?ne of a distinction here, it should be pointed out that the line between model and coder components of algorithms is not always well de?ned. It turns out that information theory is the glue that ties the model and coder components together. In particular it gives a very nice theory about how probabilities are related to information content and code length. As we will see, this theory matches practice almost perfectly, and we can achieve code lengths almost identical to what the theory predicts. Another question about compression algorithms is how does one judge the quality of one versus another. In the case of lossless compression there are several criteria I can think of, the time to compress, the time to reconstruct, the size of the compressed messages, and the generality—i.e., does it only work on Shakespeare or does it do Byron too. In the case of lossy compression the judgement is further complicated since we also have to worry about how good the lossy approximation is. There are typically tradeoffs between the amount of compression, the runtime, and the quality of the reconstruction. Depending on your application one might be more important than another and one would want to pick your algorithm appropriately. Perhaps the best attempt to systematically compare lossless compression algorithms is the Archive Comparison Test (ACT) by Jeff Gilchrist. It reports times and compression ratios for 100s of compression algorithms over many databases. It also gives a score based on a weighted average of runtime and the compression ratio. This chapter will be organized by ?rst covering some basics of information theory. Section 3 then discusses the coding component of compressing algorithms and shows how coding is related to the information theory. Section 4 discusses various models for generating the probabilities needed by the coding component. Section 5 describes the LempelZiv algorithms, and Section 6 covers other lossless algorithms (currently just BurrowsWheeler).
4
2 Information Theory
2.1 Entropy
Shannon borrowed the de?nition of entropy from statistical physics to capture the notion of how much information is contained in a and their probabilities. For a set of possible messages S , Shannon de?ned entropy1 as, 1 H (S ) = p(s) log2 p (s ) s∈S where p(s) is the probability of message s. The de?nition of Entropy is very similar to that in statistical physics—in physics S is the set of possible states a system can be in and p(s) is the probability the system is in state s. We might remember that the second law of thermodynamics basically says that the entropy of a system and its surroundings can only increase. Getting back to messages, if we consider the individual messages s ∈ S , Shannon de?ned the notion of the self information of a message as i(s) = log2 1 . p (s )
This self information represents the number of bits of information contained in it and, roughly speaking, the number of bits we should use to send that message. The equation says that messages with higher probability will contain less information (e.g., a message saying that it will be sunny out in LA tomorrow is less informative than one saying that it is going to snow). The entropy is simply a weighted average of the information of each message, and therefore the average number of bits of information in the set of messages. Larger entropies represent more information, and perhaps counterintuitively, the more random a set of messages (the more even the probabilities) the more information they contain on average.
1
Technically this de?nition is for ?rstorder Entropy. We will get back to the general notion of Entropy.
5
Here are some examples of entropies for different probability distributions over ?ve messages. p(S ) = {0.25, 0.25, 0.25, 0.125, 0.125} H = 3 × 0.25 × log2 4 + 2 × 0.125 × log2 8 = 1.5 + 0.75 = 2.25 p(s) = {0.5, 0.125, 0.125, 0.125, 0.125} H = 0.5 × log2 2 + 4 × 0.125 × log2 8 = 0.5 + 1.5 = 2 p(s) = {0.75, 0.0625, 0.0625, 0.0625, 0.0625} H 4 = 0.75 × log2 ( ) + 4 × 0.0625 × log2 16 3 = 0.3 + 1 = 1.3
Note that the more uneven the distribution, the lower the Entropy. Why is the logarithm of the inverse probability the right measure for self information of a message? Although we will relate the self information and entropy to message length more formally in Section 3 lets try to get some intuition here. First, for a set of n = 2i equal probability messages, the probability of each is 1/n. We also know that if all are the same length, then log2 n bits are required 1 = log2 n. to encode each message. Well this is exactly the self information since i(Si ) = log2 p i Another property of information we would like, is that the information given by two independent messages should be the sum of the information given by each. In particular if messages A and B are independent, the probability of sending one after the other is p(A)p(B ) and the information contained is them is i(AB ) = lg 1 1 1 = lg + lg = i(A) + i(B ) . p(A)p(B ) p(A) p(A)
The logarithm is the “simplest” function that has this property.
2.2 The Entropy of the English Language
We might be interested in how much information the English Language contains. This could be used as a bound on how much we can compress English, and could also allow us to compare the density (information content) of different languages. One way to measure the information content is in terms of the average number of bits per character. Table 1 shows a few ways to measure the information of English in terms of bitspercharacter. If we assume equal probabilities for all characters, a separate code for each character, and that there are 96 printable characters (the number on a standard keyboard) then each character 6
bits/char bits ?log(96)? 7 entropy 4.5 Huffman Code (avg.) 4.7 Entropy (Groups of 8) 2.4 Asymptotically approaches: 1.3 Compress 3.7 Gzip 2.7 BOA 2.0 Table 1: Information Content of the English Language would take ?log 96? = 7 bits. The entropy assuming even probabilities is log 96 = 6.6 bits/char. If we give the characters a probability distribution (based on a corpus of English text) the entropy is reduced to about 4.5 bits/char. If we assume a separate code for each character (for which the Huffman code is optimal) the number is slightly larger 4.7 bits/char. Note that so far we have not taken any advantage of relationships among adjacent or nearby characters. If you break text into blocks of 8 characters, measure the entropy of those blocks (based on measuring their frequency in an English corpus) you get an entropy of about 19 bits. When we divide this by the fact we are coding 8 characters at a time, the entropy (bits) per character is 2.4. If we group larger and larger blocks people have estimated that the entropy would approach 1.3 (or lower). It is impossible to actually measure this because there are too many possible strings to run statistics on, and no corpus large enough. This value 1.3 bits/char is an estimate of the information content of the English language. Assuming it is approximately correct, this bounds how much we can expect to compress English text if we want lossless compression. Table 1 also shows the compression rate of various compressors. All these, however, are general purpose and not designed speci?cally for the English language. The last one, BOA, is the current stateoftheart for generalpurpose compressors. To reach the 1.3 bits/char the compressor would surely have to “know” about English grammar, standard idioms, etc.. A more complete set of compression ratios for the Calgary corpus for a variety of compressors is shown in Table 2. The Calgary corpus is a standard benchmark for measuring compression ratios and mostly consists of English text. In particular it consists of 2 books, 5 papers, 1 bibliography, 1 collection of news articles, 3 programs, 1 terminal session, 2 object ?les, 1 geophysical data, and 1 bitmap b/w image. The table shows how the state of the art has improved over the years.
2.3 Conditional Entropy and Markov Chains
Often probabilities of events (messages) are dependent on the context in which they occur, and by using the context it is often possible to improve our probabilities, and as we will see, reduce the entropy. The context might be the previous characters in text (see PPM in Section 4.5), or the neighboring pixels in an image (see JBIG in Section 4.3). 7
Date May 1977 1984 1987 1987 1987 . 1988 . Oct 1994 1995 1997 1999
bpc 3.94 3.32 3.30 3.24 3.18 2.71 2.48 2.47 2.34 2.29 1.99 1.89
scheme LZ77 LZMW LZH MTF LZB GZIP PPMC SAKDC PPM? BW BOA RK
authors Ziv, Lempel Miller and Wegman Brent Moffat Bell . Moffat Williams Cleary, Teahan, Witten Burrows, Wheeler Sutton Taylor
Table 2: Lossless compression ratios for text compression on Calgary Corpus The conditional probability of an event e based on a context c is written as p(ec). The overall (unconditional) probability of an event e is related by p(e) = c∈C p(c)p(ec), where C is the set of all possible contexts. Based on conditional probabilities we can de?ne the notion of conditional 1 selfinformation as i(ec) = log2 p(e of an event e in the context c. This need not be the same as c) the unconditional selfinformation. For example, a message stating that it is going to rain in LA with no other information tells us more than a message stating that it is going to rain in the context that it is currently January. As with the unconditional case, we can de?ne the average conditional selfinformation, and we call this the conditionalentropy of a source of messages. We have to derive this average by averaging both over the contexts and over the messages. For a message set S and context set C , the conditional entropy is H (S  C ) =
c∈C
p (c)
s∈S
p(sc) log2
1 . p (s  c)
It is not hard to show that if the probability distribution of S is independent of the context C then H (S C ) = H (S ), and otherwise H (S C ) < H (S ). In other words, knowing the context can only reduce the entropy. Shannon actually originally de?ned Entropy in terms of information sources. An information sources generates an in?nite sequence of messages Xk , k ∈ {?∞, . . . , ∞} from a ?xed message set S . If the probability of each message is independent of the previous messages then the system is called an independent and identically distributed (iid) source. The entropy of such a source is called the unconditional or ?rst order entropy and is as de?ned in Section 2.1. In this chapter by default we will use the term entropy to mean ?rstorder entropy. Another kind of source of messages is a Markov process, or more precisely a discrete time Markov chain. A sequence follows an order k Markov model if the probability of each message 8
P(bw) P(ww) Sw P(wb) Sb P(bb)
Figure 1: A two state ?rstorder Markov Model (or event) only depends on the k previous messages, in particular p(xn xn?1 , . . . , xn?k ) = p(xn xn?1 , . . . , xn?k , . . .) where xi is the ith message generated by the source. The values that can be taken on by {xn?1 , . . . , xn?k } are called the states of the system. The entropy of a Markov process is de?ned by the conditional entropy, which is based on the conditional probabilities p(xn xn?1 , . . . , xn?k ). Figure 1 shows an example of an ?rstorder Markov Model. This Markov model represents the probabilities that the source generates a black (b) or white (w ) pixel. Each arc represents a conditional probability of generating a particular pixel. For example p(w b) is the conditional probability of generating a white pixel given that the previous one was black. Each node represents one of the states, which in a ?rstorder Markov model is just the previously generated message. Lets consider the particular probabilities p(bw ) = .01, p(w w ) = .99, p(bb) = .7, p(w b) = .3. It is not hard to solve for p(b) = 1/31 and p(w ) = 30/31 (do this as an exercise). These probabilities give the conditional entropy 30/31(.01 log(1/.01) + .99 log(1/.99)) + 1/31(.7 log(1/.7) + .3 log(1/.3)) ≈ .107 This gives the expected number of bits of information contained in each pixel generated by the source. Note that the ?rstorder entropy of the source is 30/31 log(31/30) + 1/31 log(1/30) ≈ .206 which is almost twice as large. Shannon also de?ned a general notion source entropy for an arbitrary source. Let An denote the set of all strings of length n from an alphabet A, then the nth order normalized entropy is de?ned as Hn = 1 1 p(X ) log . n X ∈An p (X ) (1)
This is normalized since we divide it by n—it represents the percharacter information. The source entropy is then de?ned as H = lim Hn .
n→∞
In general it is extremely hard to determine the source entropy of an arbitrary source process just by looking at the output of the process. This is because to calculate accurate probabilities even for a relatively simple process could require looking at extremely long sequences. 9
3 Probability Coding
As mentioned in the introduction, coding is the job of taking probabilities for messages and generating bit strings based on these probabilities. How the probabilities are generated is part of the model component of the algorithm, which is discussed in Section 4. In practice we typically use probabilities for parts of a larger message rather than for the complete message, e.g., each character or word in a text. To be consistent with the terminology in the previous section, we will consider each of these components a message on its own, and we will use the term message sequence for the larger message made up of these components. In general each little message can be of a different type and come from its own probability distribution. For example, when sending an image we might send a message specifying a color followed by messages specifying a frequency component of that color. Even the messages specifying the color might come from different probability distributions since the probability of particular colors might depend on the context. We distinguish between algorithms that assign a unique code (bitstring) for each message, and ones that “blend” the codes together from more than one message in a row. In the ?rst class we will consider Huffman codes, which are a type of pre?x code. In the later category we consider arithmetic codes. The arithmetic codes can achieve better compression, but can require the encoder to delay sending messages since the messages need to be combined before they can be sent.
3.1 Pre?x Codes
A code C for a message set S is a mapping from each message to a bit string. Each bit string is called a codeword, and we will denote codes using the syntax C = {(s1 , w1 ), (s2 , w2), · · · , (sm , wm )}. Typically in computer science we deal with ?xedlength codes, such as the ASCII code which maps every printable character and some control characters into 7 bits. For compression, however, we would like codewords that can vary in length based on the probability of the message. Such variable length codes have the potential problem that if we are sending one codeword after the other it can be hard or impossible to tell where one codeword ?nishes and the next starts. For example, given the code {(a, 1), (b, 01), (c, 101), (d, 011)}, the bitsequence 1011 could either be decoded as aba, ca, or ad. To avoid this ambiguity we could add a special stop symbol to the end of each codeword (e.g., a 2 in a 3valued alphabet), or send a length before each symbol. These solutions, however, require sending extra data. A more ef?cient solution is to design codes in which we can always uniquely decipher a bit sequence into its code words. We will call such codes uniquely decodable codes. A pre?x code is a special kind of uniquely decodable code in which no bitstring is a pre?x of another one, for example {(a, 1), (b, 01), (c, 000), (d, 001)}. All pre?x codes are uniquely decodable since once we get a match, there is no longer code that can also match. Exercise 3.1.1. Come up with an example of a uniquely decodable code that is not a pre?x code. Pre?x codes actually have an advantage over other uniquely decodable codes in that we can decipher each message without having to see the start of the next message. This is important when sending messages of different types (e.g., from different probability distributions). In fact in certain 10
applications one message can specify the type of the next message, so it might be necessary to fully decode the current message before the next one can be interpreted. A pre?x code can be viewed as a binary tree as follows ? Each message is a leaf in the tree ? The code for each message is given by following a path from the root to the leaf, and appending a 0 each time a left branch is taken, and a 1 each time a right branch is taken. We will call this tree a pre?xcode tree. Such a tree can also be useful in decoding pre?x codes. As the bits come in, the decoder can follow a path down to the tree until it reaches a leaf, at which point it outputs the message and returns to the root for the next bit (or possibly the root of a different tree for a different message type). In general pre?x codes do not have to be restricted to binary alphabets. We could have a pre?x code in which the bits have 3 possible values, in which case the corresponding tree would be ternary. In this chapter we only consider binary codes. Given a probability distribution on a set of messages and associated variable length code C , we de?ne the average length of the code as la (C ) =
(s,w )∈C
p (s )l (w )
where l(w ) is the length of the codeword w . We say that a pre?x code C is an optimal pre?x code if la (C ) is minimized (i.e., there is no other pre?x code for the given probability distribution that has a lower average length). 3.1.1 Relationship to Entropy It turns out that we can relate the average length of pre?x codes to the entropy of a set of messages, as we will now show. We will make use of the KraftMcMillan inequality Lemma 3.1.1. KraftMcMillan Inequality. For any uniquely decodable code C , 2?l(w) ≤ 1 ,
(s,w )∈C
where l(w ) it the length of the codeword w . Also, for any set of lengths L such that 2 ?l ≤ 1 ,
l∈L
there is a pre?x code C of the same size such that l(wi ) = li (i = 1, . . . L). The proof of this is left as a homework assignment. Using this we show the following Lemma 3.1.2. For any message set S with a probability distribution and associated uniquely decodable code C , H (S ) ≤ la (C ) 11
Proof. In the following equations for a message s ∈ S , l(s) refers to the length of the associated code in C . 1 ? p (s )
H (S ) ? la (C ) =
s∈S
p(s) log2
p (s )l (s )
s∈S
1 = p(s) log2 ? l (s ) p ( s ) s∈S =
s∈S
p(s) log2 p(s) log2
s∈S
1 ? log2 2l(s) p (s )
=
2?l(s) p (s )
≤ log2 (
s∈S
2?l(s) )
≤ 0
The second to last line is based on Jensen’s inequality which states that if a function f (x) is concave then i pi f (xi ) ≤ f ( i pi xi ), where the pi are positive probabilities. The logarithm function is concave. The last line uses the KraftMcMillan inequality. This theorem says that entropy is a lower bound on the average code length. We now also show an upper bound based on entropy for optimal pre?x codes. Lemma 3.1.3. For any message set S with a probability distribution and associated optimal pre?x code C , la (C ) ≤ H (S ) + 1 . Proof. Take each message s ∈ S and assign it a length l(s) = ?log p(1s) ?. We have 2?l(s) =
s∈S s∈S
2??log p(s) ? 2? log p(s)
s∈S
1
1
≤ =
s∈S
p (s )
= 1
Therefore by the KraftMcMillan inequality there is a pre?x code C ′ with codewords of length
12
l(s). Now la (C ′ ) =
(s,w )∈C ′
p (s )l (w ) p(s)?log
(s,w )∈C ′
= ≤
(s,w )∈C ′
1 ? p (s ) 1 ) p (s ) 1 p (s )
p(s)(1 + log p(s) log
(s,w )∈C ′
= 1+ = 1 + H (S )
By the de?nition of optimal pre?x codes, la (C ) ≤ la (C ′ ). Another property of optimal pre?x codes is that larger probabilities can never lead to longer codes, as shown by the following theorem. This theorem will be useful later. Theorem 3.1.1. If C is an optimal pre?x code for the probabilities {p1 , p2 , . . . , pn } then pi > pj implies that l(ci ) ≤ l(cj ). Proof. Assume l(ci ) > l(cj ). Now consider the code gotten by switching ci and cj . If la is the average length of our original code, this new code will have length
′ = la + pj (l(ci ) ? l(cj )) + pi (l(cj ) ? l(ci )) la = la + (pj ? pi )(l(ci ) ? l(cj ))
(2) (3)
Given our assumptions the (pj ? pi )(l(ci ) ? l(cj )) is negative which contradicts the assumption that la is an optimal pre?x code.
3.2 Huffman Codes
Huffman codes are optimal pre?x codes generated from a set of probabilities by a particular algorithm, the Huffman Coding Algorithm. David Huffman developed the algorithm as a student in a class on information theory at MIT in 1950. The algorithm is now probably the most prevalently used component of compression algorithms, used as the back end of GZIP, JPEG and many other utilities. The Huffman algorithm is very simple and is most easily described in terms of how it generates the pre?xcode tree. ? Start with a forest of trees, one for each message. Each tree contains a single vertex with weight wi = pi ? Repeat until only a single tree remains 13
– Select two trees with the lowest weight roots (w1 and w2 ). – Combine them into a single tree by adding a new root with weight w1 + w2 , and making the two trees its children. It does not matter which is the left or right child, but our convention will be to put the lower weight root on the left if w1 = w2 . For a code of size n this algorithm will require n ? 1 steps since every complete binary tree with n leaves has n ? 1 internal nodes, and each step creates one internal node. If we use a priority queue with O (log n) time insertions and ?ndmins (e.g., a heap) the algorithm will run in O (n log n) time. The key property of Huffman codes is that they generate optimal pre?x codes. We show this in the following theorem, originally given by Huffman. Lemma 3.2.1. The Huffman algorithm generates an optimal pre?x code. Proof. The proof will be on induction of the number of messages in the code. In particular we will show that if the Huffman code generates an optimal pre?x code for all probability distributions of n messages, then it generates an optimal pre?x code for all distributions of n + 1 messages. The base case is trivial since the pre?x code for 1 message is unique (i.e., the null message) and therefore optimal. We ?rst argue that for any set of messages S there is an optimal code for which the two minimum probability messages are siblings (have the same parent in their pre?x tree). By lemma 3.1.1 we know that the two minimum probabilities are on the lowest level of the tree (any complete binary tree has at least two leaves on its lowest level). Also, we can switch any leaves on the lowest level without affecting the average length of the code since all these codes have the same length. We therefore can just switch the two lowest probabilities so they are siblings. Now for induction we consider a set of message probabilities S of size n + 1 and the corresponding tree T built by the Huffman algorithm. Call the two lowest probability nodes in the tree x and y , which must be siblings in T because of the design of the algorithm. Consider the tree T ′ gotten by replacing x and y with their parent, call it z , with probability pz = px + py (this is effectively what the Huffman algorithm does). Lets say the depth of z is d, then la (T ) = la (T ′ ) + px (d + 1) + py (d + 1) ? pz d = la (T ′ ) + px + py . (4) (5)
To see that T is optimal, note that there is an optimal tree in which x and y are siblings, and that wherever we place these siblings they are going to add a constant px + py to the average length of any pre?x tree on S with the pair x and y replaced with their parent z . By the induction hypothesis la (T ′ ) is minimized, since T ′ is of size n and built by the Huffman algorithm, and therefore la (T ) is minimized and T is optimal. Since Huffman coding is optimal we know that for any probability distribution S and associated Huffman code C H (S ) ≤ la (C ) ≤ H (S ) + 1 .
14
3.2.1 Combining Messages Even though Huffman codes are optimal relative to other pre?x codes, pre?x codes can be quite inef?cient relative to the entropy. In particular H (S ) could be much less than 1 and so the extra 1 in H (S ) + 1 could be very signi?cant. One way to reduce the permessage overhead is to group messages. This is particularly easy if a sequence of messages are all from the same probability distribution. Consider a distribution of six possible messages. We could generate probabilities for all 36 pairs by multiplying the probabilities of each message (there will be at most 21 unique probabilities). A Huffman code can now be generated for this new probability distribution and used to code two messages at a time. Note that this technique is not taking advantage of conditional probabilities since it directly multiplies the probabilities. In general by grouping k messages the overhead of Huffman coding can be reduced from 1 bit per message to 1/k bits per message. The problem with this technique is that in practice messages are often not from the same distribution and merging messages from different distributions can be expensive because of all the possible probability combinations that might have to be generated. 3.2.2 Minimum Variance Huffman Codes The Huffman coding algorithm has some ?exibility when two equal frequencies are found. The choice made in such situations will change the ?nal code including possibly the code length of each message. Since all Huffman codes are optimal, however, it cannot change the average length. For example, consider the following message probabilities, and codes. symbol probability code 1 a 0.2 01 b 0.4 1 c 0.2 000 d 0.1 0010 e 0.1 0011 code 2 10 00 11 010 011
Both codings produce an average of 2.2 bits per symbol, even though the lengths are quite different in the two codes. Given this choice, is there any reason to pick one code over the other? For some applications it can be helpful to reduce the variance in the code length. The variance is de?ned as p(c)(l(c) ? la (C ))2
c∈C
With lower variance it can be easier to maintain a constant character transmission rate, or reduce the size of buffers. In the above example, code 1 clearly has a much higher variance than code 2. It turns out that a simple modi?cation to the Huffman algorithm can be used to generate a code that has minimum variance. In particular when choosing the two nodes to merge and there is a choice based on weight, always pick the node that was created earliest in the algorithm. Leaf nodes are assumed to be created before all internal nodes. In the example above, after d and e are joined, the pair will have the same probability as c and a (.2), but it was created afterwards, so we join c and 15
0 0
b
1 1 0
a
1
c
0
d
1
e
Figure 2: Binary tree for Huffman code 2 a. Similarly we select b instead of ac to join with de since it was created earlier. This will give code 2 above, and the corresponding Huffman tree in Figure 2.
3.3 Arithmetic Coding
Arithmetic coding is a technique for coding that allows the information from the messages in a message sequence to be combined to share the same bits. The technique allows the total number of bits sent to asymptotically approach the sum of the self information of the individual messages 1 ). (recall that the self information of a message is de?ned as log2 p i To see the signi?cance of this, consider sending a thousand messages each having probability .999. Using a Huffman code, each message has to take at least 1 bit, requiring 1000 bits to be sent. 1 On the other hand the self information of each message is log2 p = .00144 bits, so the sum of this i selfinformation over 1000 messages is only 1.4 bits. It turns out that arithmetic coding will send all the messages using only 3 bits, a factor of hundreds fewer than a Huffman coder. Of course this is an extreme case, and when all the probabilities are small, the gain will be less signi?cant. Arithmetic coders are therefore most useful when there are large probabilities in the probability distribution. The main idea of arithmetic coding is to represent each possible sequence of n messages by a separate interval on the number line between 0 and 1, e.g. the interval from .2 to .5. For a sequence of messages with probabilities p1 , . . . , pn , the algorithm will assign the sequence to an interval of size n i=1 pi , by starting with an interval of size 1 (from 0 to 1) and narrowing the interval by a factor of pi on each message i. We can bound the number of bits required to uniquely identify an interval of size s, and use this to relate the length of the representation to the self information of the messages. In the following discussion we assume the decoder knows when a message sequence is complete either by knowing the length of the message sequence or by including a special endof?le message. This was also implicitly assumed when sending a sequence of messages with Huffman codes since the decoder still needs to know when a message sequence is over. We will denote the probability distributions of a message set as {p(1), . . . , p(m)}, and we de?ne the accumulated probability for the probability distribution as
j ?1
f (j ) =
i=1
p(i)
(j = 1, . . . , m). 16
(6)
1.0 c 0.7 b 0.2 0.0 a
0.7 c 0.55 b 0.3 0.2 a
0.3 c 0.27 b 0.22 0.2 a
0.27 c 0.255 b 0.23 0.22 a
Figure 3: An example of generating an arithmetic code assuming all messages are from the same probability distribution a = .2, b = .5 and c = .3. The interval given by the message sequence babc is [.255, .27). So, for example, the probabilities {.2, .5, .3} correspond to the accumulated probabilities {0, .2, .7}. Since we will often be talking about sequences of messages, each possibly from a different probability distribution, we will denote the probability distribution of the ith message as {pi (1), . . . , pi(mi )}, and the accumulated probabilities as {fi (1), . . . , fi (mi )}. For a particular sequence of message values, we denote the index of the ith message value as vi . We will use the shorthand pi for pi (vi ) and fi for fi (vi ). Arithmetic coding assigns an interval to a sequence of messages using the following recurrences li = si = fi li?1 + fi ? si?1 pi si?1 ? pi i=1 1<i≤n (7)
i=1 1<i≤n
where ln is the lower bound of the interval and sn is the size of the interval, i.e. the interval is given by [ln , ln + sn ). We assume the interval is inclusive of the lower bound, but exclusive of the upper bound. The recurrence narrows the interval on each step to some part of the previous interval. Since the interval starts in the range [0,1), it always stays within this range. An example of generating an interval for a short message sequences is illustrated in Figure 3. An important property of the intervals generated by Equation 7 is that all unique message sequences of length n will have non overlapping intervals. Specifying an interval therefore uniquely determines the message sequence. In fact, any number within an interval uniquely determines the message sequence. The job of decoding is basically the same as encoding but instead of using the message value to narrow the interval, we use the interval to select the message value, and then narrow it. We can therefore “send” a message sequence by specifying a number within the corresponding interval. The question remains of how to ef?ciently send a sequence of bits that represents the interval, or a number within the interval. Real numbers between 0 and 1 can be represented in binary fractional notation as .b1 b2 b3 . . .. For example .75 = .11, 9/16 = .1001 and 1/3 = .0101, where w means that the sequence w is repeated in?nitely. We might therefore think that it is adequate to represent each interval by selecting the number within the interval which has the fewest bits in binary fractional notation, and use that as the code. For example, if we had the intervals [0, .33), 17
[.33, 67), and [.67, 1) we would represent these with .01(1/4), .1(1/2), and .11(3/4). It is not hard to show that for an interval of size s we need at most ??log2 s? bits to represent such a number. The problem is that these codes are not a set of pre?x codes. If you sent me 1 in the above example, I would not know whether to wait for another 1 or interpret it immediately as the interval [.33, 67). To avoid this problem we interpret every binary fractional codeword as an interval itself. In particular as the interval of all possible completions. For example, the codeword .010 would represent the interval [1/4, 3/8) since the smallest possible completion is .0100 = 1/4 and the largest possible completion is .0101 = 3/8 ? ?. Since we now have several kinds of intervals running around, we will use the following terms to distinguish them. We will call the current interval of the message sequence (i.e [li , li + si )) the sequence interval, the interval corresponding to the probability of the ith message (i.e., [fi , fi + pi )) the message interval, and the interval of a codeword the code interval. An important property of code intervals is that there is a direct correspondence between whether intervals overlap and whether they form pre?x codes, as the following Lemma shows. Lemma 3.3.1. For a code C , if no two intervals represented by its binary codewords w ∈ C overlap then the code is a pre?x code. Proof. Assume codeword a is a pre?x of codeword b, then b is a possible completion of a and therefore its interval must be fully included in the interval of a. This is a contradiction. To ?nd a pre?x code, therefore, instead of using any number in the interval to be coded, we select a codeword who’s interval is fully included within the interval. Returning to the previous example of the intervals [0, .33), [.33, 67), and [.67, 1), the codewords .00[0, .25), .100[.5, .625), and .11[.75, 1) are adequate. In general for an interval of size s we can always ?nd a codeword of length ??log2 s? + 1, as shown by the following lemma. Lemma 3.3.2. For any l and an s such that l, s ≥ 0 and l + s < 1, the interval represented by taking the binary fractional representation of l + s/2 and truncating it to ?? log2 s? + 1 bits is contained in the interval [l, l + s). Proof. A binary fractional representation with l digits represents an interval of size less than 2?l since the difference between the minimum and maximum completions are all 1s starting at the l + 1th location. This has a value 2?l ? ?. The interval size of a ??log2 s? + 1 bit representation is therefore less than s/2. Since we truncate l + s/2 downwards the upper bound of the interval represented by the bits is less than l + s. Truncating the representation of a number to ??log2 s? +1 bits can have the effect of reducing it by at most s/2. Therefore the lower bound of truncating l + s/2 is at least l. The interval is therefore contained in [l, l + s). We will call the algorithm made up of generating an interval by Equation 7 and then using the truncation method of Lemma 3.3.2, the RealArithCode algorithm. Theorem 3.3.1. For a sequence of n messages, with self informations s1 , . . . , sn the length of the arithmetic code generated by RealArithCode is bounded by 2 + n i=1 si , and the code will not be a pre?x of any other sequence of n messages. 18
Proof. Equation 7 will generate a sequence interval of size s = n i=1 pi . Now by Lemma 3.3.2 we know an interval of size s can be represented by 1 + ?? log s? bits, so we have
n
1 + ?? log s? = 1 + ?? log2 (
i=1 n
p i )?
= 1+?
i=1 n
? log2 pi ? si ?
i=1 n
= 1+? < 2+
i=1
si
The claim that the code is not a pre?x of other messages is taken directly from Lemma 3.3.1. The decoder for RealArithCode needs to read the input bits on demand so that it can determine when the input string is complete. In particular it loops for n iterations, where n is the number of messages in the sequence. On each iteration it reads enough input bits to narrow the code interval to within one of the possible message intervals, narrows the sequence interval based on that message, and outputs that message. When complete, the decoder will have read exactly all the characters generated by the coder. We give a more detailed description of decoding along with the integer implementation described below. From a practical point of view there are a few problems with the arithmetic coding algorithm we described so far. First, the algorithm needs arbitrary precision arithmetic to manipulate l and s. Manipulating these numbers can become expensive as the intervals get very small and the number of signi?cant bits get large. Another problem is that as described the encoder cannot output any bits until it has coded the full message. It is actually possible to interleave the generation of the interval with the generation of its bit representation by opportunistically outputting a 0 or 1 whenever the interval falls within the lower or upper half. This technique, however, still does not guarantee that bits are output regularly. In particular if the interval keeps reducing in size but still straddles .5, then the algorithm cannot output anything. In the worst case the algorithm might still have to wait until the whole sequence is received before outputting any bits. To avoid this problem many implementations of arithmetic coding break message sequences into ?xed size blocks and use arithmetic coding on each block separately. This approach also has the advantage that since the group size is ?xed, the encoder need not send the number of messages, except perhaps for the last group which could be smaller than the block size. 3.3.1 Integer Implementation It turns out that if we are willing to give up a little bit in the ef?ciency of the coding, we can used ?xed precision integers for arithmetic coding. This implementation does not give precise arithmetic codes, because of roundoff errors, but if we make sure that both the coder and decoder 19
are always rounding in the same way the decoder will always be able to precisely interpret the message. For this algorithm we assume the probabilities are given as counts c(1), c(2), . . . , c(m) , and the cumulative count are de?ned as before (f (i) = as m T =
j =1 i?1 j =1 c(j )).
The total count will be denoted
c(j ) .
Using counts avoids the need for fractional or real representations of the probabilities. Instead of using intervals between 0 and 1, we will use intervals between [0..(R ? 1)] where R = 2k (i.e., is a power of 2). There is the additional restriction that R > 4T . This will guarantee that no region will become too small to represent. The larger R is, the closer the algorithm will come to real arithmetic coding. As in the noninteger arithmetic coding, each message can come from its own probability distribution (have its own counts and accumulative counts), and we denote the ith message using subscripts as before. The coding algorithm is given in Figure 4. The current sequence interval is speci?ed by the integers l (lower) and u (upper), and the corresponding interval is [l, u + 1). The size of the interval s is therefore u ? l + 1. The main idea of this algorithm is to always keep the size greater than R/4 by expanding the interval whenever it gets too small. This is what the inner while loop does. In this loop whenever the sequence interval falls completely within the top half of the region (from R/2 to R) we know that the next bit is going to be a 1 since intervals can only shrink. We can therefore output a 1 and expand the top half to ?ll the region. Similarly if the sequence interval falls completely within the bottom half we can output a 0 and expand the bottom half of the region to ?ll the full region. The third case is when the interval falls within the middle half of the region(from R/4 to 3R/4). In this case the algorithm cannot output a bit since it does not know whether the bit will be a 0 or 1. It, however, can expand the middle region and keep track that is has expanded by incrementing a count m. Now when the algorithm does expand around the top (bottom), it outputs a 1 (0) followed by m 0s (1s). To see why this is the right thing to do, consider expanding around the middle m times and then around the top. The ?rst expansion around the middle locates the interval between 1/4 and 3/4 of the initial region, and the second between 3/8 and 5/8. After m expansions the interval is narrowed to the region (1/2 ? 1/2m+1 , 1/2 + 1/2m+1 ). Now when we expand around the top we narrow the interval to (1/2, 1/2 + 1/2m+1 ). All intervals contained in this range will start with a 1 followed by m 0s. Another interesting aspect of the algorithm is how it ?nishes. As in the case of realnumber arithmetic coding, to make it possible to decode, we want to make sure that the code (bit pattern) for any one message sequence is not a pre?x of the code for another message sequence. As before, the way we do this is to make sure the code interval is fully contained in the sequence interval. When the integer arithmetic coding algorithm (Figure 4) exits the for loop, we know the sequence interval [l, u] completely covers either the second quarter (from R/4 to R/2) or the third quarter (from R/2 to 3R/4) since otherwise one of the expansion rules would have been applied. The 20
function IntArithCode(?le, k, n) R = 2k l=0 u=R?1 m=0 for i = 1 to n s=u?l+1 u = l + ?s · fi (vi + 1)/T ? ? 1 l = l + ?s · fi (vi )/T ? while true if (l ≥ R/2) // interval in top half WriteBit(1) u = 2u ? R + 1 l = 2l ? R for j = 1 to m WriteBit(0) m=0 else if (u < R/2) // interval in bottom half WriteBit(0) u = 2u + 1 l = 2l for j = 1 to m WriteBit(1) m=0 else if (l ≥ R/4 and u < 3R/4) // interval in middle half u = 2u ? R/2 + 1 l = 2l ? R/2 m=m+1 else continue // exit while loop end while end for if (l ≥ R/4) // output ?nal bits WriteBit(1) for j = 1 to m WriteBit(0) WriteBit(0) else WriteBit(0) for j = 1 to m WriteBit(1) WriteBit(1) Figure 4: Integer Arithmetic Coding.
21
algorithm therefore simply determines which of these two regions the sequence interval covers and outputs code bits that narrow the code interval to one of these two quarters—a 01 for the second quarter, since all completions of 01 are in the second quarter, and a 10 for the third quarter. After outputting the ?rst of these two bits the algorithm must also output m bits corresponding to previous expansions around the middle. The reason that R needs to be at least 4T is that the sequence interval can become as small as R/4 + 1 without falling completely within any of the three halves. To be able to resolve the counts C (i), T has to be at least as large as this interval. An example: Here we consider an example of encoding a sequence of messages each from the same probability distribution, given by the following counts. c(1) = 1, c(2) = 10, c(3) = 20 The cumulative counts are f (1) = 0, f (2) = 1, f (3) = 11 and T = 31. We will chose k = 8, so that R = 256. This satis?es the requirement that R > 4T . Now consider coding the message sequence 3, 1, 2, 3. Figure 5 illustrates the steps taken in coding this message sequence. The full code that is output is 01011111101 which is of length 11. The sum of the selfinformation of the messages is ?(log2 (20/31) + log2 (10/31) + log2 (1/31) + log2 (10/31)) = 8.85 . Note that this is not within the bound given by Theorem 3.3.1. This is because we are not generating an exact arithmetic code and we are loosing some coding ef?ciency. We now consider how to decode a message sent using the integer arithmetic coding algorithm. The code is given in Figure 6. The idea is to keep separate lower and upper bounds for the code interval (lb and ub ) and the sequence interval (l and u). The algorithm reads one bit at a time and reduces the code interval by half for each bit that is read (the bottom half when the bit is a 0 and the top half when it is a 1). Whenever the code interval falls within an interval for the next message, the message is output and the sequence interval is reduced by the message interval. This reduction is followed by the same set of expansions around the top, bottom and middle halves as followed by the encoder. The sequence intervals therefore follow the exact same set of lower and upper bounds as when they were coded. This property guarantees that all rounding happens in the same way for both the coder and decoder, and is critical for the correctness of the algorithm. It should be noted that reduction and expansion of the code interval is always exact since these are always changed by powers of 2.
4 Applications of Probability Coding
To use a coding algorithm we need a model from which to generate probabilities. Some simple models are to count characters for text or pixel values for images and use these counts as probabilities. Such counts, however, would only give a compression ratio of about 4.7/8 = .59 for English 22
i start 1 2 + 3 + + + + + + 4 + end
vi 3 2 1
f (vi ) f (vi + 1) 11 1 0 31 11 1
2
1
11
l 0 90 95 62 62 124 120 112 96 64 0 6 12
u 255 255 147 167 64 129 131 135 143 159 191 67 135
s m 256 166 0 53 0 106 1 3 1 6 0 12 1 24 2 48 4 96 5 192 6 62 6 124 0 0
expand rule
output
(middle half) (bottom half) (middle half) (middle half) (middle half) (middle half) (middle half) (bottom half) (?nal out) 01
0111111 01
Figure 5: Example of integer arithmetic coding. The rows represent the steps of the algorithm. Each row starting with a number represents the application of a contraction based on the next message, and each row with a + represents the application of one of the expansion rules. text as compared to the best compression algorithms that give ratios of close to .2. In this section we give some examples of more sophisticated models that are used in realworld applications. All these techniques take advantage of the “context” in some way. This can either be done by transforming the data before coding (e.g., runlength coding, movetofront coding, and residual coding), or directly using conditional probabilities based on a context (JBIG and PPM). An issue to consider about a model is whether it is static or dynamic. A model can be static over all message sequences. For example one could predetermine the frequency of characters and text and “hardcode” those probabilities into the encoder and decoder. Alternatively, the model can be static over a single message sequence. The encoder executes one pass over the sequence to determine the probabilities, and then a second pass to use those probabilities in the code. In this case the encoder needs to send the probabilities to the decoder. This is the approach taken by most vector quantizers. Finally, the model can be dynamic over the message sequence. In this case the encoder updates its probabilities as it encodes messages. To make it possible for the decoder to determine the probability based on previous messages, it is important that for each message, the encoder codes it using the old probability and then updates the probability based on the message. The advantages of this approach are that the coder need not send additional probabilities, and that it can adapt to the sequence as it changes. This approach is taken by PPM. Figure 7 illustrates several aspects of our general framework. It shows, for example, the interaction of the model and the coder. In particular, the model generates the probabilities for each possible message, and the coder uses these probabilities along with the particular message to generate the codeword. It is important to note that the model has to be identical on both sides. Furthermore the model can only use previous messages to determine the probabilities. It cannot use the current 23
function IntArithDecode(?le,k, n) R = 2k l =0 u=R?1 // sequence interval lb = 0 ub = R ? 1 // code interval j=1 // message number while j ≤ n do s=u?l+1 i=0 do // ?nd if the code interval is within one of the message intervals i=i+1 u′ = l + ?s · fj (i + 1)/Tj ? ? 1 l′ = l + ?s · fj (i)/Tj ? while i ≤ mj and not((lb ≥ l′ ) and (ub ≤ u′ )) if i > mj then // halve the size of the code interval by reading a bit b = ReadBit(?le) sb = ub ? lb + 1 lb = lb + b(sb /2) ub = lb + sb /2 ? 1 else Output(i) // output the message in which the code interval ?ts ′ ′ u=u l=l // adjust the sequence interval j =j+1 while true if (l ≥ R/2) // sequence interval in top half u = 2u ? R + 1 l = 2l ? R ub = 2ub ? R + 1 lb = 2lb ? R else if (u < R/2) // sequence interval in bottom half u = 2u + 1 l = 2l ub = 2ub + 1 lb = 2lb else if (l ≥ R/4 and u < 3R/4) // sequence interval in middle half u = 2u ? R/2 + 1 l = 2l ? R/2 ub = 2ub ? R/2 + 1 lb = 2lb ? R/2 else continue // exit inner while loop end if end while Figure 6: Integer Arithmetic Decoding
24
Compress
Model Static Part Dynamic Part Message 6 s∈S
Uncompress
Model Static Part Dynamic Part 6 s∈S ?
{p(s) : s ∈ S}
Coder
codeword
w ≈ i(s) 1 = log p( s)
{p(s) : s ∈ S } Decoder
6
Transform 6 In
Inverse Transform ? Out
Figure 7: The general framework of a model and coder. message since the decoder does not have this message and therefore could not generate the same probability distribution. The transform has to be invertible.
4.1 Runlength Coding
Probably the simplest coding scheme that takes advantage of the context is runlength coding. Although there are many variants, the basic idea is to identify strings of adjacent messages of equal value and replace them with a single occurrence along with a count. For example, the message sequence acccbbaaabb could be transformed to (a,1), (c,3), (b,2), (a,3), (b,2). Once transformed, a probability coder (e.g., Huffman coder) can be used to code both the message values and the counts. It is typically important to probability code the runlengths since short lengths (e.g., 1 and 2) are likely to be much more common than long lengths (e.g., 1356). An example of a realworld use of runlength coding is for the ITUT T4 (Group 3) standard for Facsimile (fax) machines2 . At the time of writing (1999), this was the standard for all home and business fax machines used over regular phone lines. Fax machines transmit blackandwhite images. Each pixel is called a pel and the horizontal resolution is ?xed at 8.05 pels/mm. The vertical resolution varies depending on the mode. The T4 standard uses runlength encoding to code each sequence of black and white pixels. Since there are only two message values black and white, only the runlengths need to be transmitted. The T4 standard speci?es the start color by placing a dummy white pixel at the front of each row so that the ?rst run is always assumed to be a white run. For example, the sequence bbbbwwbbbbb would be transmitted as 1,4,2,5. The
2
ITUT is part of the International Telecommunications Union (ITU, http://www.itu.ch/.
25
runlength 0 1 2 3 4 .. 20 .. 64+ 128+
white codeword 00110101 000111 0111 1000 1011 0001000 11011 10010
black codeword 0000110111 010 11 10 011 00001101000 0000001111 000011001000
Table 3: ITUT T4 Group 3 Runlength Huffman codes. T4 standard uses static Huffman codes to encode the runlengths, and uses a separate codes for the black and white pixels. To account for runs of more than 64, it has separate codes to specify multiples of 64. For example, a length of 150, would consist of the code for 128 followed by the code for 22. A small subset of the codes are given in Table 4.1. These Huffman codes are based on the probability of each runlength measured over a large number of documents. The full T4 standard also allows for coding based on the previous line.
4.2 MoveToFront Coding
Another simple coding schemes that takes advantage of the context is movetofront coding. This is used as a substep in several other algorithms including the BurrowsWheeler algorithm discussed later. The idea of movetofront coding is to preprocess the message sequence by converting it into a sequence of integers, which hopefully is biases toward integers with low values. The algorithm then uses some form of probability coding to code these values. In practice the conversion and coding are interleaved, but we will describe them as separate passes. The algorithm assumes that each message comes from the same alphabet, and starts with a total order on the alphabet (e.g., [a, b, c, d, . . .]). For each message, the ?rst pass of the algorithm outputs the position of the character in the current order of the alphabet, and then updates the order so that the character is at the head. For example, coding the character c with an order [a, b, c, d, . . .] would output a 3 and change the order to [c, a, b, d, . . .]. This is repeated for the full message sequence. The second pass converts the sequence of integers into a bit sequence using Huffman or Arithmetic coding. The hope is that equal characters often appear close to each other in the message sequence so that the integers will be biased to have low values. This will give a skewed probability distribution and good compression.
26
4.3 Residual Coding: JPEGLS
Residual compression is another general compression technique used as a substep in several algorithms. As with movetofront coding, it preprocesses the data so that the message values have a better skew in their probability distribution, and then codes this distribution using a standard probability coder. The approach can be applied to message values that have some meaningful total order (i.e., in which being close in the order implies similarity), and is most commonly used for integers values. The idea of residual coding is that the encoder tries to guess the next message value based on the previous context and then outputs the difference between the actual and guessed value. This is called the residual. The hope is that this residual is biased toward low values so that it can be effectively compressed. Assuming the decoder has already decoded the previous context, it can make the same guess as the coder and then use the residual it receives to correct the guess. By not specifying the residual to its full accuracy, residual coding can also be used for lossy compression Residual coding is used in JPEG lossless (JPEG LS), which is used to compress both greyscale and color images.3 Here we discuss how it is used on gray scale images. Color images can simply be compressed by compressing each of the three color planes separately. The algorithm compresses images in raster order—the pixels are processed starting at the topmost row of an image from left to right and then the next row, continuing down to the bottom. When guessing a pixel the encoder and decoder therefore have as their disposal the pixels to the left in the current row and all the pixels above it in the previous rows. The JPEG LS algorithm just uses 4 other pixels as a context for the guess—the pixel to the left (W), above and to the left (NW), above (N), and above and to the right (NE). The guess works in two stages. The ?rst stage makes the following guess for each pixel value. ? max(N, W ) ≤ NW ? min(W, N ) max(W, N ) min(N, W ) < NW (8) G = ? N + W ? NW otherwise
This might look like a magical equation, but it is based on the idea of taking an average of nearby pixels while taking account of edges. The ?rst and second clauses capture horizontal and vertical edges. For example if N > W and N ≤ NW this indicates a horizontal edge and W is used as the guess. The last clause captures diagonal edges. Given an initial guess G a second pass adjusts that guess based on local gradients. It uses the three gradients between the pairs of pixels (NW, W ), (NW, N ), and (N, NE ). Based on the value of the gradients (the difference between the two adjacent pixels) each is classi?ed into one of 9 groups. This gives a total of 729 contexts, of which only 365 are needed because of symmetry. Each context stores its own adjustment value which is used to adjust the guess. Each context also stores information about the quality of previous guesses in that context. This can be used to predict variance and can help the probability coder. Once the algorithm has the ?nal guess for the pixel, it determines the residual and codes it.
3
This algorithm is based on the LOCOI (LOw COmplexity LOssless COmpression for Images) algorithm and the of?cial standard number is ISO144951/ITUT.87.
27
O O O O O O O A O O ? (a) O O O O O A O O O O ? (b)
Figure 8: JBIG contexts: (a) threeline template, and (b) twoline template. ? is the current pixel and A is the “roaming pixel”.
4.4 Context Coding: JBIG
The next two techniques we discuss both use conditional probabilities directly for compression. In this section we discuss using contextbased conditional probabilities for Bilevel (blackandwhite) images, and in particular the JBIG1 standard. In the next section we discuss using a context in text compression. JBIG stands for the Joint Bilevel Image Processing Group. It is part of the same standardization effort that is responsible for the JPEG standard. The algorithm we describe here is JBIG1, which is a lossless compressor for bilevel images. JBIG1 typically compresses 2080% better than ITU Groups III and IV fax encoding outlined in Section 4.1. JBIG is similar to JPEG LS in that it uses a local context of pixels to code the current pixel. Unlike JPEG LS, however, JBIG uses conditional probabilities directly. JBIG also allows for progressive compression—an image can be sent as a set of layers of increasing resolution. Each layer can use the previous layer to aid compression. We ?rst outline how the initial layer is compressed, and then how each following layer is compressed. The ?rst layer is transmitted in raster order, and the compression uses a context of 10 pixels above and to the right of the current pixel. The standard allows for two different templates for the context as shown in Figure 8. Furthermore, the pixel marked A is a roaming pixel and can be chosen to be any ?xed distance to the right of where it is marked in the ?gure. This roaming pixel is useful for getting good compression on images with repeated vertical lines. The encoder decides on which of the two templates to use and on where to place A based on how well they compress. This information is speci?ed at the head of the compressed message sequence. Since each pixel can only have two values, there are 210 possible contexts. The algorithm dynamically generates the conditional probabilities for a black or white pixel for each of the contexts, and uses these probabilities in a modi?ed arithmetic coder—the coder is optimized to avoid multiplications and divisions. The decoder can decode the pixels since it can build the probability table in the same way as the encoder. The higherresolution layers are also transmitted in raster order, but now in addition to using a context of previous pixels in the current layer, the compression algorithm can use pixels from the previous layer. Figure 9 shows the context templates. The context consists of 6 pixels from the current layer, and 4 pixels from the lower resolution layer. Furthermore 2 additional bits are needed to specify which of the four con?gurations the coded pixel is in relative to the previous layer. This gives a total of 12 bits and 4096 contexts. The algorithm generates probabilities in the same way as for the ?rst layer, but now with some more contexts. The JBIG standard also speci?es how to generate lower resolution layers from higher resolution layers, but this won’t be discussed here. 28
O A O O O O ? O O ?
O A O O
O A O O O O ? O O ?
O A O O
Figure 9: JBIG contexts for progressive transmission. The dark circles are the low resolution pixels, the 0s are the highresolution pixels, the A is a roaming pixel, and the ? is the pixel we want to code/decode. The four context con?gurations are for the four possible con?gurations of the highresolution pixel relative to the low resolution pixel. The approach used by JBIG is not well suited for coding greyscale images directly since the number of possible contexts go up as mp , where m is the number of greyscale pixel values, and p is the number of pixels. For 8bit greyscale images and a context of size 10, the number of possible contexts is 230 , which is far too many. The algorithm can, however, be applied to greyscale images indirectly by compressing each bitposition in the grey scale separately. This still does not work well for greyscale levels with more than 2 or 3 bits.
4.5 Context Coding: PPM
Over the past decade, variants of this algorithm have consistently given either the best or close to the best compression ratios (PPMC, PPM? , BOA and RK from Table 2 all use ideas from PPM). They are, however, are not very fast. The main idea of PPM (Prediction by Partial Matching) is to take advantage of the previous K characters to generate a conditional probability of the current character. The simplest way to do this would be to keep a dictionary for every possible string s of k characters, and for each string have counts for every character x that follows s. The conditional probability of x in the context s is then C (xs)/C (s), where C (xs) is the number of times x follows s and C (s) is the number of times s appears. The probability distributions can then be used by a Huffman or Arithmetic coder to generate a bit sequence. For example, we might have a dictionary with qu appearing 100 times and e appearing 45 times after qu. The conditional probability of the e is then .45 and the coder should use about 1 bit to encode it. Note that the probability distribution will change from character to character since each context has its own distribution. In terms of decoding, as long as the context precedes the character being coded, the decoder will know the context and therefore know which probability distribution to use. Because the probabilities tend to be high, arithmetic codes work much better than Huffman codes for this approach. There are two problems with the basic dictionary method described in the previous paragraph. 29
Order 0 Order 1 Order 2 Context Counts Context Counts Context Counts empty a=4 a c=3 ac b=1 b=2 c=2 c=5 b a=2 ba c=1 c a=1 b=2 ca a=1 c=2 cb a=2 cc a=1 b=1
Figure 10: An example of the PPM table for k = 2 on the string accbaccacba. First, the dictionaries can become very large. There is no solution to this problem other than to keep k small, typically 3 or 4. A second problem is what happens if the count is zero. We cannot use zero probabilities in any of the coding methods (they would imply in?nitely long strings). One way to get around this is to assume a probability of not having seen a sequence before and evenly distribute this probability among the possible following characters that have not been seen. Unfortunately this gives a completely even distribution, when in reality we might know that a is more likely than b, even without knowing its context. The PPM algorithm has a clever way to deal with the case when a context has not been seen before, and is based on the idea of partial matching. The algorithm builds the dictionary on the ?y starting with an empty dictionary, and every time the algorithm comes across a string it has not seen before it tries to match a string of one shorter length. This is repeated for shorter and shorter lengths until a match is found. For each length 0, 1, . . . , k the algorithm keeps statistics of patterns it has seen before and counts of the following characters. In practice this can all be implemented in a single trie. In the case of the length0 contexts the counts are just counts of each character seen assuming no context. An example table is given in Figure 10 for a string accbaccacba. Now consider following this string with a c. Since the algorithm has the context ba followed by c in its dictionary, it can output the c based on its probability in this context. Although we might think the probability should be 1, since c is the only character that has ever followed ba, we need to give some probability of no match, which we will call the “escape” probability. We will get back to how this probability is set shortly. If instead of c the next character to code is an a, then the algorithm does not ?nd a match for a length 2 context so it looks for a match of length 1, in this case the context is the previous a. Since a has never followed by another a, the algorithm still does not ?nd a match, and looks for a match with a zero length context. In this case it ?nds the a and uses the appropriate probability for a (4/11). What if the algorithm needs to code a d? In this case the algorithm does not even ?nd the character in the zerolength context, so it assigns the character a probability assuming all 30
Order 0 Order 1 Order 2 Context Counts Context Counts Context Counts empty a=4 a c=3 ac b=1 b=2 $=1 c=2 c=5 $=2 $=3 b a=2 $=1 ba c=1 $=1 c a=1 b=2 ca a=1 c=2 $=1 $=3 cb a=2 $=1 cc a=1 b=1 $=2
Figure 11: An example of the PPMC table for k = 2 on the string accbaccacba. This assumes the “virtual” count of each escape symbol ($) is the number of different characters that have appeared in the context. unseen characters have even likelihood. Although it is easy for the encoder to know when to go to a shorter context, how is the decoder supposed to know in which sized context to interpret the bits it is receiving. To make this possible, the encoder must notify the decoder of the size of the context. The PPM algorithm does this by assuming the context is of size k and then sending an “escape” character whenever moving down a size. In the example of coding an a given above, the encoder would send two escapes followed by the a since the context was reduced from 2 to 0. The decoder then knows to use the probability distribution for zero length contexts to decode the following bits. The escape can just be viewed as a special character and given a probability within each context as if it was any other kind of character. The question is how to assign this probability. Different variants of PPM have different rules. PPMC uses the following scheme. It sets the count for the escape character to be the number of different characters seen following the given context. Figure 11 shows an example of the counts using this scheme. In this example, the probability of no match for a context of ac is 2/(1 + 2 + 2) = .4 while the probability for a b in that context is .2. There seems to be no theoretical justi?cation for this choice, but empirically it works well. There is one more trick that PPM uses. This is that when switching down a context, the algorithm can use the fact that it switched down to exclude the possibility of certain characters from the shorter context. This effectively increases the probability of the other characters and decreases the code length. For example, if the algorithm were to code an a, it would send two escapes, but 31
then could exclude the c from the counts in the zero length context. This is because there is no way that two escapes would be followed by a c since the c would have been coded in a length 2 context. The algorithm could then give the a a probability of 4/6 instead of 4/11 (.58 bits instead of 1.46 bits!).
5 The LempelZiv Algorithms
The LempelZiv algorithms compress by building a dictionary of previously seen strings. Unlike PPM which uses the dictionary to predict the probability of each character, and codes each character separately based on the context, the LempelZiv algorithms code groups of characters of varying lengths. The original algorithms also did not use probabilities—strings were either in the dictionary or not and all strings in the dictionary were give equal probability. Some of the newer variants, such as gzip, do take some advantage of probabilities. At the highest level the algorithms can be described as follows. Given a position in a ?le, look through the preceeding part of the ?le to ?nd the longest match to the string starting at the current position, and output some code that refers to that match. Now move the ?nger past the match. The two main variants of the algorithm were described by Ziv and Lempel in two separate papers in 1977 and 1978, and are often refered to as LZ77 and LZ78. The algorithms differ in how far back they search and how they ?nd matches. The LZ77 algorithm is based on the idea of a sliding window. The algorithm only looks for matches in a window a ?xed distance back from the current position. Gzip, ZIP, and V.42bis (a standard modem protocal) are all based on LZ77. The LZ78 algorithm is based on a more conservative approach to adding strings to the dictionary. Unix compress, and the Gif format are both based on LZ78. In the following discussion of the algorithms we will use the term cursor to mean the position an algorithm is currently trying to encode from.
5.1 LempelZiv 77 (Sliding Windows)
The LZ77 algorithm and its variants use a sliding window that moves along with the cursor. The window can be divided into two parts, the part before the cursor, called the dictionary, and the part starting at the cursor, called the lookahead buffer. The size of these two parts are parameters of the program and are ?xed during execution of the algorithm. The basic algorithm is very simple, and loops executing the following steps 1. Find the longest match of a string starting at the cursor and completely contained in the lookahead buffer to a string starting in the dictionary. 2. Output a triple (p, n, c) containing the position p of the occurence in the window, the length n of the match and the next character c past the match. 3. Move the cursor n + 1 characters forward.
32
Step 1 2 3 4 5
a a a a a
a a a a a
c c c c c
a a a a a
a a a a a
c c c c c
Input String a b c a b c a b c a b c a b c
a a a a a
b b b b b
a a a a a
a a a a a
a a a a a
c c c c c
Output Code (0, 0, a) (1, 1, c) (3, 4, b) (3, 3, a) (1, 2, c)
Figure 12: An example of LZ77 with a dictionary of size 6 and a lookahead buffer of size 4. The cursor position is boxed, the dictionary is bold faced, and the lookahed buffer is underlined. The last step does not ?nd the longer match (10,3,1) since it is outside of the window. The position p can be given relative to the cursor with 0 meaning no match, 1 meaning a match starting at the previous character, etc.. Figure 12 shows an example of the algorithm on the string aacaacabcababac. To decode the message we consider a single step. Inductively we assume that the decoder has correctly constructed the string up to the current cursor, and we want to show that given the triple (p, n, c) it can reconstruct the string up to the next cursor position. To do this the decoder can look the string up by going back p positions and taking the next n characters, and then following this with the character c. The one tricky case is when n > p, as in step 3 of the example in Figure 12. The problem is that the string to copy overlaps the lookahead buffer, which the decoder has not ?lled yet. In this case the decoder can reconstruct the message by taking p characters before the cursor and repeating them enough times after the cursor to ?ll in n positions. If, for example, the code was (2,7,d) and the two characters before the cursor were ab, the algorithm would place abababa and then the d after the cursor. There have been many improvements on the basic algorithm. Here we will describe several improvements that are used by gzip. Two formats: This improvement, often called the LZSS Variant, does not include the next character in the triple. Instead it uses two formats, either a pair with a position and length, or just a character. An extra bit is typically used to distinguish the formats. The algorithm tries to ?nd a match and if it ?nds a match that is at least of length 3, it uses the offset, length format, otherwise it uses the single character format. It turns out that this improvement makes a huge difference for ?les that do not compress well since we no longer have to waste the position and length ?elds. Huffman coding the components: Gzip uses separate huffman codes for the offset, the length and the character. Each uses addaptive Huffman codes. Non greedy: The LZ77 algorithm is greedy in the sense that it always tries to ?nd a match starting at the ?rst character in the lookahead buffer without caring how this will affect later matches. For some strings it can save space to send out a single character at the current cursor position and then match on the next position, even if there is a match at the current position. For example, consider coding the string 33
d
b
c
a
b
c
d
a
b
c
a
b
In this case LZCC would code it as (1,3,3), (0,a), (0,b). The last two letters are coded as singletons since the match is not at least three characters long. This same buffer could instead be coded as (0,a), (1,6,4) if the coder was not greedy. In theory one could imagine trying to optimize coding by trying all possible combinations of matches in the lookahead buffer, but this could be costly. As a tradeoff that seems to work well in practice, Gzip only looks ahead 1 character, and only chooses to code starting at the next character if the match is longer than the match at the current character. Hash Table: To quickly access the dictionary Gzip uses a hash table with every string of length 3 used as the hash keys. These keys index into the position(s) in which they occur in the ?le. When trying to ?nd a match the algorithm goes through all of the hash entries which match on the ?rst three characters and looks for the longest total match. To avoid long searches when the dictionary window has many strings with the same three characters, the algorithm only searches a bucket to a ?xed length. Within each bucket, the positions are stored in an order based on the position. This makes it easy to select the more recent match when the two longest matches are equal length. Using the more recent match better skews the probability distribution for the offsets and therefore decreases the average length of the Huffman codes.
5.2 LempelZivWelch
In this section we will describe the LZW (LempelZivWelch) variant of LZ78 since it is the one that is most commonly used in practice. In the following discussion we will assume the algorithm is used to encode byte streams (i.e., each message is a byte). The algorithm maintains a dictionary of strings (sequences of bytes). The dictionary is initialized with one entry for each of the 256 possible byte values—these are strings of length one. As the algorithm progresses it will add new strings to the dictionary such that each string is only added if a pre?x one byte shorter is already in the dictionary. For example, John is only added if Joh had previously appeared in the message sequence. We will use the following interface to the dictionary. We assume that each entry of the dictionary is given an index, where these indices are typically given out incrementally starting at 256 (the ?rst 256 are reserved for the byte values). Creates a new dictionary entry by extending an existing dictionary entry given by index C with the byte x. Returns the new index. ′ C = GetIndex(C, x) Return the index of the string gotten by extending the string corresponding to index C with the byte x. If the entry does not exist, return 1. W = GetString(C ) Returns the string W corresponding to index C . Flag = IndexInDict?(C ) Returns true if the index C is in the dictionary and false otherwise. C ′ = AddDict(C, x)
34
function LZW Encode(File) C = ReadByte(File) while C = EOF do x = ReadByte(File) C ′ = GetIndex(C, x) while C ′ = ?1 do C = C′ x = ReadByte(File) C ′ = GetIndex(C, x) Output(C ) AddDict(C, x) C=x
function LZW Decode(File) C = ReadIndex(File) W = GetString(C ) Output(W ) while C = EOF do C ′ = ReadIndex(File) if IndexInDict?(C ′ ) then W = GetString(C ′ ) AddDict(C, W [0]) else C’ = AddDict(C, W [0]) W = GetString(C ′ ) Output(W ) C = C’
Figure 13: Code for LZW encoding and decoding. The encoder is described in Figure 13, and Tables 4 and 5 give two examples of encoding and decoding. Each iteration of the outer loop works by ?rst ?nding the longest match W in the dictionary for a string starting at the current position—the inner loop ?nds this match. The iteration then outputs the index for W and adds the string W x to the dictionary, where x is the next character after the match. The use of a “dictionary” is similar to LZ77 except that the dictionary is stored explicitly rather than as indices into a window. Since the dictionary is explicit, i.e., each index corresponds to a precise string, LZW need not specify the length. The decoder works since it builds the dictionary in the same way as the encoder and in general can just look up the indices it receives in its copy of the dictionary. One problem, however, is that the dictionary at the decoder is always one step behind the encoder. This is because the encoder can add W x to its dictionary at a given iteration, but the decoder will not know x until the next message it receives. The only case in which this might be a problem is if the encoder sends an index of an entry added to the dictionary in the previous step. This happens when the encoder sends an index for a string W and the string is followed by W W [0], where W [0] refers to the ?rst character of W (i.e., the input is of the form W W W [0]). On the iteration the encoder sends the index for W it adds W W [0] to its dictionary. On the next iteration it sends the index for W W [0]. If this happens, the decoder will receive the index for W W [0], which it does not have in its dictionary yet. Since the it is able to decode the previous W , however, it can easily reconstruct W W [0]. This case is handled by the else clause in LZW decode, and shown by the second example. A problem with the algorithm is that the dictionary can get too large. There are several choices of what to do. Here are some of them. 1. Throw dictionary away when reaching a certain size (GIF) 2. Throw dictionary away when not effective (Unix Compress) 35
C x GetIndex(C,x) init a a b 1 b c 1 c a 1 + a b 256 256 c 1 + c a 258 258 EOF 1
AddDict(C,x) 256 (a,b) 257 (b,c) 258 (c,a) 259 (256,c) 
Output(C) a b c 256 258
(a) Encoding C C’ W Init a a a b b b c c c 256 ab 256 258 ca IndexInDict?(C’) true true true true (b) Decoding Table 4: LZW Encoding and Decoding abcabca. The rows with a + for encoding are iterations of the inner while loop. C x GetIndex(C,x) init a a a 1 + a a 256 256 a 1 + a a 256 + 256 a 257 257 EOF 1 AddDict(C,x) 256 (a,a) 257 (256,a) Output(C) a 256 AddDict(C,W[0]) 256 (a,b) 257 (b,c) 258 (c,a) 259 (256,c) Output(W) a b c ab ca

257
(a) Encoding C C’ W Init a a a 256 aa 256 257 aaa IndexInDict?(C’) false false (b) Decoding Table 5: LZW Encoding and Decoding aaaaaa. This is an example in which the decoder does not have the index in its dictionary. 36 AddDict(C,W[0]) 256 (a,a) 257 (256,a) Output(W) a aa aaa
3. Throw Least Recently Used entry away when reaches a certain size (BTLZ  British Telecom Standard) Implementing the Dictionary: One of the biggest advantages of the LZ78 algorithms and reason for its success is that the dictionary operations can run very quickly. Our goal is to implement the 3 dictionary operations. The basic idea is to store the dictionary as a partially ?lled kary tree such that the root is the empty string, and any path down the tree to a node from the root speci?es the match. The path need not go to a leaf since because of the pre?x property of the LZ78 dictionary, all paths to internal nodes must belong to strings in the dictionary. We can use the indices as pointers to nodes of the tree (possibly indirectly through an array). To implement the GetString(C ) function we start at the node pointed to by C and follow a path from that node to the root. This requires that every child has a pointer to its parent. To implement the GetIndex(C, x) operation we go from the node pointed to by C and search to see if there is a child bytevalue x and return the corresponding index. For the AddDict(C, x) operation we add a child with bytevalue x to the node pointed to by C . If we assume k is constant, the GetIndex and AddDict operations will take constant time since they only require going down one level of the tree. The GetString operation requires W  time to follow the tree up to the root, but this operation is only used by the decoder, and always outputs W after decoding it. The whole algorithm for both coding and decoding therefore require time that is linear in the message size. To discuss one more level of detail, lets consider how to store the pointers. The parent pointers are trivial to keep since each node only needs a single pointer. The children pointers are a bit more dif?cult to do ef?ciently. One choice is to store an array of length k for each node. Each entry is initialized to empty and then searches can be done with a single arrary reference, but we need k pointers per node (k is often 256 in practice) and the memory is prohibitive. Another choice is to use a linked list (or possibly balanced tree) to store the children. This has much better space but requires more time to ?nd a child (although technically still constant time since k is “constant”). A compromise that can be made in practice is to use a linked list until the number of children in a node rises above some threshold k’ and then switch to an array. This would require copying the linked list into the array when switching. Yet another technique is to use a hash table instead of child pointers. The string being searched for can be hashed directly to the appropriate index.
6 Other Lossless Compression
6.1 Burrows Wheeler
The Burrows Wheeler algorithm is a relatively recent algorithm. An implementation of the algorithm called bzip, is currently one of the best overall compression algorithms for text. It gets compression ratios that are within 10% of the best algorithms such as PPM, but runs signi?cantly faster. Rather than describing the algorithm immediately, lets try to go through a thought process that leads to the algorithm. Recall that the basic idea of PPM was to try to ?nd as long a context as 37
a ac acc accb accba accbac accbacc accbacca accbaccac accbaccacb
a c c b a c c a c b a (a)
ccbaccacba cbaccacba baccacba accacba ccacba cacba acba cba ba a
ccbaccacba4 cbaccacbaa1 baccacbaac1 accacbaacc2 ccacbaaccb1 cacbaaccba2 acbaaccbac3 cbaaccbacc4 baaccbacca3 aaccbaccac5 accbaccacb2 (b)
a1 c1 c2 b1 a2 c3 c4 a3 c5 b2 a4
cbaccacbaa1 ccbaccacba4 cacbaaccba2 baaccbacca3 accbaccacb2 ccacbaaccb1 baccacbaac1 acbaaccbac3 aaccbaccac5 accacbaacc2 cbaaccbacc4 (c)
c1 a1 c3 c5 a4 a2 c2 c4 b2 b1 a3
Figure 14: Sorting the characters a1 c1 c2 b1 a2 c3 c4 a3 c5 b2 a4 based on context: (a) each character in its context, (b) end context moved to front, and (c) characters sorted by their context using reverse lexicographic ordering. We use subscripts to distinguish different occurences of the same character. possible that matched the current context and use that to effectively predict the next character. A problem with PPM is in selecting k . If we set k too large we will usually not ?nd matches and end up sending too many escape characters. On the other hand if we set it too low, we would not be taking advantage of enough context. We could have the system automatically select k based on which does the best encoding, but this is expensive. Also within a single text there might be some very long contexts that could help predict, while most helpful contexts are short. Using a ?xed k we would probably end up ignoring the long contexts. Lets see if we can come up with a way to take advantage of the context that somehow automatically adapts. Ideally we would like the method also to be a bit faster. Consider taking the string we want to compress and looking at the full context for each character—i.e., all previous characters from the start of the string up to the character. In fact, to make the contexts the same length, which will be convenient later, we add to the head of each context the part of the string following the character making each context n ? 1 characters. Examples of the context for each character of the string accbaccacba are given in Figure 6.1. Now lets sort these contexts based on reverse lexical order, such that the last character of the context is the most signi?cant (see Figure 6.1c). Note that now characters with the similar contexts (preceeding characters) are near each other. In fact, the longer the match (the more preceeding characters that match identically) the closer they will be to each other. This is similar to PPM in that it prefers longer matches when “grouping”, but will group things with shorter matches when the longer match does not exist. The difference is that there is no ?xed limit k on the length of a match—a match of length 100 has priority over a match of 99. In practice the sorting based on the context is executed in blocks, rather than for the full message sequence. This is because the full message sequence and additional data structures required for sorting it, might not ?t in memory. The process of sorting the characters by their context
38
is often refered to as a blocksorting transform. In the dicussion below we will refer to the sequence of characters generated by a blocksorting transform as the contextsorted sequence (e.g., c1 a1 c3 c5 a4 a2 c2 c4 b2 b1 a3 in Figure 6.1). Given the correlation between nearyby characters in a contextsorted sequence, we should be able to code them quite ef?ciently by using, for example, a movetofront coder (Section 4.2). For long strings with somewhat larger character sets this technique should compress the string signi?cantly since the same character is likely to appear in similar contexts. Experimentally, in fact, the technique compresses about as well as PPM even though it has no magic number k or magic way to select the escape probabilities. The problem remains, however, of how to reconstruct the original sequence from contextsorted sequence. The way to do this is the ingenious contribution made by Burrows and Wheeler. You might try to recreate it before reading on. The order of the mostsigni?cant characters in the sorted contexts plays an important role in decoding. In the example of Figure 6.1, these are a1 a4 a2 a3 b2 b1 c1 c3 c5 c2 c4 . The characters are sorted, but equal valued characters do not necessarily appear in the same order as in the input sequence. The following lemma is critical in the algorithm for ef?ciently reconstruct the sequence. Lemma 6.1.1. For the BlockSorting transform, as long as there are at least two distinct characters in the input, equal valued characters appear in the same order in the mostsigni?cant characters of the sorted contexts as in the output (the context sorted sequence). Proof. Since the contexts are sorted in reverse lexicographic order, sets of contexts whose mostsigni?cant character are equal will be ordered by the remaining context—i.e., the string of all previous characters. Now consider the contexts of the contextsorted sequence. If we drop the leastsigni?cant character of these contexts, then they are exactly the same as the remaining context above, and therefore will be sorted into the same ordering. The only time that dropping the leastsigni?cant character can make a difference is when all other characters are equal. This can only happen when all characters in the input are equal. Based on Lemma 6.1.1, it is not hard to reconstruct the sequence from the contextsorted sequence as long as we are also given the index of the ?rst character to output (the ?rst character in the original input sequence). The algorithm is given by the following code. function BW Decode(In,FirstIndex,n) S = MoveToFrontDecode(In,n) R = Rank(S ) j = FirstIndex for i = 1 to n ? 1 Out[i] = S [j ] j = R[j ] For an ordered sequence S , the Rank(S ) function returns a sequence of integers specifying for each character c ∈ S how many characters are either less than c or equal to c and appear before c in S . Another way of saying this is that it speci?es the position of the character if it where sorted using a stable sort. 39
S s s n a ? s m a i s s s a a i (a)
Sort(S ) a1 a2 a3 a4 i1 i2 m1 n1 s1 s2 s3 s4 s5 s6
S s1 s2 n1 a1 s3 m1 a2 i1 s4 s5 s6 a3 a4 i2 (b)
Rank(S ) 9 10 8 1 ? 11 7 2 5 12 13 14 3 4 6
a4 a1 s1 s4 a3 n1 i1 s3 s6 i2 m1 a2 s2 s5
? ? ? ? ? ? ? ? ? ? ? ? ?
Out a1 ? s1 s4 a3 n1 i1 s3 s6 i2 m1 a2 s2 s5 a4
(c) The decoded message sequence is
Figure 15: BurrowsWheeler Decoding Example. assanissimassa.
To show how this algorithms works, we consider an example in which the MoveToFront decoder returns S = ssnasmaisssaai, and in which FirstIndex = 4 (the ?rst a). The example is shown in Figure 15(a). We can generate the most signi?cant characters of the contexts simply by sorting S . The result of the sort is shown in Figure 15(b) along with the rank R. Because of Lemma 6.1.1, we know that equal valued characters will have the same order in this sorted sequence and in S . This is indicated by the subscripts in the ?gure. Now each row of Figure 15(b) tells us for each character what the next character is. We can therefore simply rebuild the initial sequence by starting at the ?rst character and adding characters one by one, as done by BW Decode and as illustrated in Figure 15(c).
7 Lossy Compression Techniques
Lossy compression is compression in which some of the information from the original message sequence is lost. This means the original sequences cannot be regenerated from the compressed sequence. Just because information is lost doesn’t mean the quality of the output is reduced. For example, random noise has very high information content, but when present in an image or a sound ?le, we would typically be perfectly happy to drop it. Also certain losses in images or sound might be completely imperceptible to a human viewer (e.g. the loss of very high frequencies). For this reason, lossy compression algorithms on images can often get a factor of 2 better compression than lossless algorithms with an imperceptible loss in quality. However, when quality does start degrading in a noticeable way, it is important to make sure it degrades in a way that is least objec40
tionable to the viewer (e.g., dropping random pixels is probably more objectionable than dropping some color information). For these reasons, the way most lossy compression techniques are used are highly dependent on the media that is being compressed. Lossy compression for sound, for example, is very different than lossy compression for images. In this section we go over some general techniques that can be applied in various contexts, and in the next two sections we go over more speci?c examples and techniques.
7.1 Scalar Quantization
A simple way to implement lossy compression is to take the set of possible messages S and reduce it to a smaller set S ′ by mapping each element of S to an element in S ′ . For example we could take 8bit integers and divide by 4 (i.e., drop the lower two bits), or take a character set in which upper and lowercase characters are distinguished and replace all the uppercase ones with lowercase ones. This general technique is called quantization. Since the mapping used in quantization is manytoone, it is irreversible and therefore lossy. In the case that the set S comes from a total order and the total order is broken up into regions that map onto the elements of S ′ , the mapping is called scalar quantization. The example of dropping the lower two bits given in the previous paragraph is an example of scalar quantization. Applications of scalar quantization include reducing the number of color bits or grayscale levels in images (used to save memory on many computer monitors), and classifying the intensity of frequency components in images or sound into groups (used in JPEG compression). In fact we mentioned an example of quantization when talking about JPEGLS. There quantization is used to reduce the number of contexts instead of the number of message values. In particular we categorized each of 3 gradients into one of 9 levels so that the context table needs only 93 entries (actually only (93 + 1)/2 due to symmetry). The term uniform scalar quantization is typically used when the mapping is linear. Again, the example of dividing 8bit integers by 4 is a linear mapping. In practice it is often better to use a nonuniform scalar quantization. For example, it turns out that the eye is more sensitive to low values of red than to high values. Therefore we can get better quality compressed images by making the regions in the low values smaller than the regions in the high values. Another choice is to base the nonlinear mapping on the probability of different input values. In fact, this idea can be formalized—for a given error metric and a given probability distribution over the input values, we want a mapping that will minimize the expected error. For certain errormetrics, ?nding this mapping might be hard. For the rootmeansquared error metric there is an iterative algorithm known as the LloydMax algorithm that will ?nd the optimal mapping. An interesting point is that ?nding this optimal mapping will have the effect of decreasing the effectiveness of any probability coder that is used on the output. This is because the mapping will tend to more evenly spread the probabilities in S ′ .
7.2 Vector Quantization
Scalar quantization allows one to separately map each color of a color image into a smaller set of output values. In practice, however, it can be much more effective to map regions of 3d color space 41
Out 4 3 2 1 40 30 20 10 2 3 4 10 20 30 40 In 40 30 20 10
Out 4 3 2 1 10 20 30 40 2 3 4 In
(a)
(b)
Figure 16: Examples of (a) uniform and (b) nonuniform scalar quantization.
200 180 160 140 Weight 120 100 80 60 40 20 1’ 2’ 3’ 4’ 5’ 6’ 7’ 8’ Height
Figure 17: Example of vectorquantization for a heightweight chart. into output values. By more effective we mean that a better compression ratio can be achieved based on an equivalent loss of quality. The general idea of mapping a multidimensional space into a smaller set of messages S ′ is called vector quantization. Vector quantization is typically implemented by selecting a set of representatives from the input space, and then mapping all other points in the space to the closest representative. The representatives could be ?xed for all time and part of the compression protocol, or they could be determined for each ?le (message sequence) and sent as part of the sequence. The most interesting aspect of vector quantization is how one selects the representatives. Typically it is implemented using a clustering algorithm that ?nds some number of clusters of points in the data. A representative is then chosen for each cluster by either selecting one of the points in the cluster or using some form of centroid for the cluster. Finding good clusters is a whole interesting topic on its own.
42
Vector quantization is most effective when the variables along the dimensions of the space are correlated. Figure 17 gives an example of possible representatives for a heightweight chart. There is clearly a strong correlation between people’s height and weight and therefore the representatives can be concentrated in areas of the space that make physical sense, with higher densities in more common regions. Using such representatives is very much more effective than separately using scalar quantization on the height and weight. We should note that vector quantization, as well as scalar quantization, can be used as part of a lossless compression technique. In particular if in addition to sending the closest representative, the coder sends the distance from the point to the representative, then the original point can be reconstructed. The distance is often referred to as the residual. In general this would not lead to any compression, but if the points are tightly clustered around the representatives, then the technique can be very effective for lossless compression since the residuals will be small and probability coding will work well in reducing the number of bits.
7.3 Transform Coding
The idea of transform coding is to transform the input into a different form which can then either be compressed better, or for which we can more easily drop certain terms without as much qualitative loss in the output. One form of transform is to select a linear set of basis functions (φi ) that span the space to be transformed. Some common sets include sin, cos, polynomials, spherical harmonics, Bessel functions, and wavelets. Figure 18 shows some examples of the ?rst three basis functions for discrete cosine, polynomial, and wavelet transformations. For a set of n values, transforms can be expressed as an n × n matrix T . Multiplying the input by this matrix T gives, the transformed coef?cients. Multiplying the coef?cients by T ?1 will convert the data back to the original form. For example, the coef?cients for the discrete cosine transform (DCT) are
iπ 1/n cos (2j +1) 2n iπ 2/n cos (2j +1) 2n
Tij =
i = 0, 0 ≤ j < n 0 < i < n, 0 ≤ j < n
The DCT is one of the most commonly used transforms in practice for image compression, more so than the discrete Fourier transform (DFT). This is because the DFT assumes periodicity, which is not necessarily true in images. In particular to represent a linear function over a region requires many large amplitude highfrequency components in a DFT. This is because the periodicity assumption will view the function as a sawtooth, which is highly discontinuous at the teeth requiring the highfrequency components. The DCT does not assume periodicity and will only require much lower amplitude highfrequency components. The DCT also does not require a phase, which is typically represented using complex numbers in the DFT. For the purpose of compression, the properties we would like of a transform are (1) to decorrelate the data, (2) have many of the transformed coef?cients be small, and (3) have it so that from the point of view of perception, some of the terms are more important than others.
43
Cosine
f(x)
i
0
1
2
Polynomial
0
1
2
Wavelet
0
1
2
Figure 18: Transforms
8 A Case Study: JPEG and MPEG
The JPEG and the related MPEG format make good realworld examples of compression because (a) they are used very widely in practice, and (b) they use many of the compression techniques we have been talking about, including Huffman codes, arithmetic codes, residual coding, runlength coding, scalar quantization, and transform coding. JPEG is used for still images and is the standard used on the web for photographic images (the GIF format is often used for textual images). MPEG is used for video and after many years of debated MPEG2 has become the standard for the transmission of highde?nition television (HDTV). This means in a few years we will all be receiving MPEG at home. As we will see, MPEG is based on a variant of JPEG (i.e. each frame is coded using a JPEG variant). Both JPEG and MPEG are lossy formats.
8.1 JPEG
JPEG is a lossy compression scheme for color and grayscale images. It works on full 24bit color, and was designed to be used with photographic material and naturalistic artwork. It is not the ideal format for linedrawings, textual images, or other images with large areas of solid color or a very limited number of distinct colors. The lossless techniques, such as JBIG, work better for such images. JPEG is designed so that the loss factor can be tuned by the user to tradeoff image size and image quality, and is designed so that the loss has the least effect on human perception. It however does have some anomalies when the compression ratio gets high, such as odd effects across the boundaries of 8x8 blocks. For high compression ratios, other techniques such as wavelet compression appear to give more satisfactory results. An overview of the JPEG compression process is given in Figure 19. We will cover each of the steps in this process. The input to JPEG are three color planes of 8bits perpixel each representing Red, Blue and Green (RGB). These are the colors used by hardware to generate images. The ?rst step of JPEG compression, which is optional, is to convert these into YIQ color planes. The YIQ color planes are 44
R
G
B (optional)
Y
I Q
For each plane Quantization DCT for each 8x8 block
zigzag order
DC difference from prev. block Huffman or Arithmetic RLE Bits
Figure 19: Steps in JPEG compression. designed to better represent human perception and are what are used on analog TVs in the US (the NTSC standard). The Y plane is designed to represent the brightness (luminance) of the image. It is a weighted average of red, blue and green (0.59 Green + 0.30 Red + 0.11 Blue). The weights are not balanced since the human eye is more responsive to green than to red, and more to red than to blue. The I (interphase) and Q (quadrature) components represent the color hue (chrominance). If you have an old blackandwhite television, it uses only the Y signal and drops the I and Q components, which are carried on a subcarrier signal. The reason for converting to YIQ is that it is more important in terms of perception to get the intensity right than the hue. Therefore JPEG keeps all pixels for the intensity, but typically down samples the two color planes by a factor of 2 in each dimension (a total factor of 4). This is the ?rst lossy component of JPEG and gives a factor of 2 compression: (1 + 2 ? .25)/3 = .5. The next step of the JPEG algorithm is to partition each of the color planes into 8x8 blocks. Each of these blocks is then coded separately. The ?rst step in coding a block is to apply a cosine transform across both dimensions. This returns an 8x8 block of 8bit frequency terms. So far this does not introduce any loss, or compression. The blocksize is motivated by wanting it to be large enough to capture some frequency components but not so large that it causes “frequency spilling”. In particular if we cosinetransformed the whole image, a sharp boundary anywhere in a line would cause high values across all frequency components in that line. After the cosine transform, the next step applied to the blocks is to use uniform scalar quantization on each of the frequency terms. This quantization is controllable based on user parameters and is the main source of information loss in JPEG compression. Since the human eye is more perceptive to certain frequency components than to others, JPEG allows the quantization scaling factor to be different for each frequency component. The scaling factors are speci?ed using an 8x8 table that simply is used to elementwise divide the 8x8 table of frequency components. JPEG 45
16 12 14 14 18 24 49 72
11 12 13 17 22 35 64 92
10 14 16 22 37 55 78 95
16 24 40 19 26 58 24 40 57 29 51 87 56 68 109 64 81 104 87 103 121 98 112 100
51 61 60 55 69 56 80 62 103 77 113 92 120 101 103 99
Table 6: JPEG default quantization table, luminance plane.
Figure 20: Zigzag scanning of JPEG blocks.
de?nes standard quantization tables for both the Y and IQ components. The table for Y is shown in Table 6. In this table the largest components are in the lowerright corner. This is because these are the highest frequency components which humans are less sensitive to than the lowerfrequency components in the upperleft corner. The selection of the particular numbers in the table seems magic, for example the table is not even symmetric, but it is based on studies of human perception. If desired, the coder can use a different quantization table and send the table in the head of the message. To further compress the image, the whole resulting table can be divided by a constant, which is a scalar “quality control” given to the user. The result of the quantization will often drop most of the terms in the lower left to zero. JPEG compression then compresses the DC component (upperleftmost) separately from the other components. In particular it uses a difference coding by subtracting the value given by the DC component of the previous block from the DC component of this block. It then Huffman or arithmetic codes this difference. The motivation for this method is that the DC component is often similar from blocktoblock so that difference coding it will give better compression. The other components (the AC components) are now compressed. They are ?rst converted into a linear order by traversing the frequency table in a zigzag order (see Figure 20). The motivation for this order is that it keeps frequencies of approximately equal length close to each other 46
Playback order: 0 1 2 3 4 5 6 7 Frame type: I B B P B B P B Data stream order: 0 2 3 1 5 6 4 8
8 9 B I 9 7
Figure 21: MPEG Bframes postponed in data stream.
in the linearorder. In particular most of the zeros will appear as one large contiguous block at the end of the order. A form of runlength coding is used to compress the linearorder. It is coded as a sequence of (skip,value) pairs, where skip is the number of zeros before a value, and value is the value. The special pair (0,0) speci?es the end of block. For example, the sequence [4,3,0,0,1,0,0,0,1,0,0,0,...] is represented as [(0,4),(0,3),(2,1),(3,1),(0,0)]. This sequence is then compressed using either arithmetic or Huffman coding. Which of the two coding schemes used is speci?ed on a perimage basis in the header.
8.2 MPEG
Correlation improves compression. This is a recurring theme in all of the approaches we have seen; the more effectively a technique is able to exploit correlations in the data, the more effectively it will be able to compress that data. This principle is most evident in MPEG encoding. MPEG compresses video streams. In theory, a video stream is a sequence of discrete images. In practice, successive images are highly interrelated. Barring cut shots or scene changes, any given video frame is likely to bear a close resemblance to neighboring frames. MPEG exploits this strong correlation to achieve far better compression rates than would be possible with isolated images. Each frame in an MPEG image stream is encoded using one of three schemes: Iframe , or intraframe, are coded as isolated images. Pframe , or predictive coded frame, are based on the previous I or Pframe. Bframe , or bidirectionally predictive coded frame, are based on either or both the previous and next I or Pframe. Figure 21 shows an MPEG stream containing all three types of frames. Iframes and Pframes appear in an MPEG stream in simple, chronological order. However, Bframes are moved so that they appear after their neighboring I and Pframes. This guarantees that each frame appears after any frame upon which it may depend. An MPEG encoder can decode any frame by buffering the two most recent I or Pframes encountered in the data stream. Figure 21 shows how Bframes are postponed in the data stream so as to simplify decoder buffering. MPEG encoders are free to mix the frame types in any order. When the scene is relatively static, P and Bframes could be used, while major scene changes could be encoded using Iframes. In practice, most encoders use some ?xed pattern.
47
Figure 22: Pframe encoding.
Since Iframes are independent images, they can be encoded as if they were still images. The particular technique used by MPEG is a variant of the JPEG technique (the color transformation and quantization steps are slightly different). Iframes are very important for use as anchor points so that the frames in the video can be accessed randomly without requiring one to decode all previous frames. To decode any frame we need only ?nd its closest previous Iframe and go from there. This is important for allowing reverse playback, skipahead, or errorrecovery. The intuition behind encoding Pframes is to ?nd matches, i.e., groups of pixels with similar patterns, in the previous reference frame and then coding the difference between the Pframe and its match. To ?nd these “matches” the MPEG algorithm partitions the Pframe into 16x16 blocks. The process by which each of these blocks is encoded is illustrated in Figure 22. For each target block in the Pframe the encoder ?nds a reference block in the previous P or Iframe that most closely matches it. The reference block need not be aligned on a 16pixel boundary and can potentially be anywhere in the image. In practice, however, the xy offset is typically small. The offset is called the motion vector. Once the match is found, the pixels of the reference block are subtracted from the corresponding pixels in the target block. This gives a residual which ideally is close to zero everywhere. This residual is coded using a scheme similar to JPEG encoding, but will ideally get a much better compression ratio because of the low intensities. In addition to sending the coded residual, the coder also needs to send the motion vector. This vector is Huffman coded. The motivation for searching other locations in the reference image for a match is to allow for the ef?cient encoding of motion. In particular if there is a moving object in the sequence of images (e.g., a car or a ball), or if the whole video is panning, then the best match will not be in the same location in the image. It should be noted that if no good match is found, then the block is coded as if it were from an Iframe. 48
In practice, the search for good matches for each target block is the most computationally expensive part of MPEG encoding. With current technology, realtime MPEG encoding is only possible with the help of custom hardware. Note, however, that while the search for a match is expensive, regenerating the image as part of the decoder is cheap since the decoder is given the motion vector and only needs to look up the block from the previous image. Bframes were not present in MPEG’s predecessor, H.261. They were added in an effort to address the following situation: portions of an intermediate Pframe may be completely absent from all previous frames, but may be present in future frames. For example, consider a car entering a shot from the side. Suppose an Iframe encodes the shot before the car has started to appear, and another Iframe appears when the car is completely visible. We would like to use Pframes for the intermediate scenes. However, since no portion of the car is visible in the ?rst Iframe, the Pframes will not be able to “reuse” that information. The fact that the car is visible in a later Iframe does not help us, as Pframes can only look back in time, not forward. Bframes look for reusable data in both directions. The overall technique is very similar to that used in Pframes, but instead of just searching in the previous I or Pframe for a match, it also searches in the next I or Pframe. Assuming a good match is found in each, the two reference frames are averaged and subtracted from the target frame. If only one good match is found, then it is used as the reference. The coder needs to send some information on which reference(s) is (are) used, and potentially needs to send two motion vectors. How effective is MPEG compression? We can examine typical compression ratios for each frame type, and form an average weighted by the ratios in which the frames are typically interleaved. Starting with a 356 × 260 pixel, 24bit color image, typical compression ratios for MPEGI are: Type I P B Avg Size 18 Kb 6 Kb 2.5 Kb 4.8 Kb Ratio 7:1 20:1 50:1 27:1
If one 356 × 260 frame requires 4.8 Kb, how much bandwidth does MPEG require in order to provide a reasonable video feed at thirty frames per second? 30f rames/sec · 4.8Kb/f rame · 8b/bit = 1.2Mbits/sec Thus far, we have been concentrating on the visual component of MPEG. Adding a stereo audio stream will require roughly another 0.25 Mbits/sec, for a grand total bandwidth of 1.45 Mbits/sec. This ?ts nicely within the 1.5 Mbit/sec capacity of a T1 line. In fact, this speci?c limit was a design goal in the formation of MPEG. Reallife MPEG encoders track bit rate as they encode, and will dynamically adjust compression qualities to keep the bit rate within some userselected bound. This bitrate control can also be important in other contexts. For example, video on a multimedia CDROM must ?t within the relatively poor bandwidth of a typical CDROM drive.
49
MPEG in the Real World MPEG has found a number of applications in the real world, including: 1. Direct Broadcast Satellite. MPEG video streams are received by a dish/decoder, which unpacks the data and synthesizes a standard NTSC television signal. 2. Cable Television. Trial systems are sending MPEGII programming over cable television lines. 3. Media Vaults. Silicon Graphics, Storage Tech, and other vendors are producing ondemand video systems, with twenty ?le thousand MPEGencoded ?lms on a single installation. 4. RealTime Encoding. This is still the exclusive province of professionals. Incorporating specialpurpose parallel hardware, realtime encoders can cost twenty to ?fty thousand dollars.
9 Other Lossy Transform Codes
9.1 Wavelet Compression
JPEG and MPEG decompose images into sets of cosine waveforms. Unfortunately, cosine is a periodic function; this can create problems when an image contains strong aperiodic features. Such local highfrequency spikes would require an in?nite number of cosine waves to encode properly. JPEG and MPEG solve this problem by breaking up images into ?xedsize blocks and transforming each block in isolation. This effectively clips the in?nitelyrepeating cosine function, making it possible to encode local features. An alternative approach would be to choose a set of basis functions that exhibit good locality without arti?cial clipping. Such basis functions, called “wavelets”, could be applied to the entire image, without requiring blocking and without degenerating when presented with highfrequency local features. How do we derive a suitable set of basis functions? We start with a single function, called a “mother function”. Whereas cosine repeats inde?nitely, we want the wavelet mother function, φ, to be contained within some local region, and approach zero as we stray further away:
x→±∞
lim φ(x) = 0
The family of basis functions are scaled and translated versions of this mother function. For some scaling factor s and translation factor l, φsl (x) = φ(2s x ? l) A well know family of wavelets are the Haar wavelets, which are derived from the following mother function: 50
φ00 =φ(x)
1
φ10 =φ(2x)
φ11 =φ(2x?1)
? ?
1
φ20 =φ(4x)
φ21 =φ(4x?1)
φ22 =φ(4x?2)
φ23 =φ(4x?3)
? ?
? ?
? ?
1
Figure 23: A small Haar wavelet family of size seven.
φ(x) =
? ?
Figure 23 shows a family of seven Haar basis functions. Of the many potential wavelets, Haar wavelets are probably the most described but the least used. Their regular form makes the underlying mathematics simple and easy to illustrate, but tends to create bad blocking artifacts if actually used for compression. Many other wavelet mother functions have also been proposed. The Morret wavelet convolves a Gaussian with a cosine, resulting in a periodic but smoothly decaying function. This function is equivalent to a wave packet from quantum physics, and the mathematics of Morret functions have been studied extensively. Figure 24 shows a sampling of other popular wavelets. Figure 25 shows that the Daubechies wavelet is actually a selfsimilar fractal. 51
1 : ?1 : ? 0 :
0 < x ≤ 1/2 1/ 2 < x ≤ 1 x ≤ 0 or x > 1
0.06 0.05 0.04 0.03 0.02 0.01 0 0.01 0.02
0.06
Daubechies_6
0.05 0.04 0.03 0.02 0.01 0 0.01
Coiflet_3
0.07 0.06 0.05 0.04 0.03
3 3 x 10 2.5 2 1.5 1 0.5 0 0.5 1 1200 1250 1300 1350 1400 1450 1500
0
500
1000 1500 2000 2500
0
500
1000 1500 2000 2500
0.02 0.01 0 0.01 0.02 0 500
0.05
Haar_4
0
0.05
0
100 200 300
400 500 600
0.06 0.05 0.04 0.03 0.02 0.01 0 0.01 0.02
Symmlet_6
1000
1500
2000
2500
0
500
1000 1500 2000 2500
Figure 24: A sampling of popular wavelets.
Figure 25: Selfsimilarity in the Daubechies wavelet.
Wavelets in the Real World Summus Ltd. is the premier vendor of wavelet compression technology. Summus claims to achieve better quality than JPEG for the same compression ratios, but has been loathe to divulge details of how their wavelet compression actually works. Summus wavelet technology has been incorporated into such items as: ? Waveletsonachip for missile guidance and communications systems. ? Image viewing plugins for Netscape Navigator and Microsoft Internet Explorer. ? Desktop image and movie compression in Corel Draw and Corel Video. ? Digital cameras under development by Fuji. In a sense, wavelet compression works by characterizing a signal in terms of some underlying generator. Thus, wavelet transformation is also of interest outside of the realm of compression. Wavelet transformation can be used to clean up noisy data or to detect selfsimilarity over widely varying time scales. It has found uses in medical imaging, computer vision, and analysis of cosmic Xray sources.
9.2 Fractal Compression
A function f (x) is said to have a ?xed point xf if xf = f (xf ). For example: f (x) = ax + b b ? xf = 1?a 52
This was a simple case. Many functions may be too complex to solve directly. Or a function may be a black box, whose formal de?nition is not known. In that case, we might try an iterative approach. Keep feeding numbers back through the function in hopes that we will converge on a solution: x0 = guess xi = f (xi?1 ) For example, suppose that we have f (x) as a black box. We might guess zero as x0 and iterate from there: x0 x1 x2 x3 x4 x5 x6 x7 x8 = = = = = = = = = 0 f (x0 ) = 1 f (x1 ) = 1.5 f (x2 ) = 1.75 f (x3 ) = 1.875 f (x4 ) = 1.9375 f (x5 ) = 1.96875 f (x6 ) = 1.984375 f (x7 ) = 1.9921875
1 In this example, f (x) was actually de?ned as 2 x +1. The exact ?xed point is 2, and the iterative solution was converging upon this value. Iteration is by no means guaranteed to ?nd a ?xed point. Not all functions have a single ?xed point. Functions may have no ?xed point, many ?xed points, or an in?nite number of ?xed points. Even if a function has a ?xed point, iteration may not necessarily converge upon it. In the above example, we were able to associate a ?xed point value with a function. If we were given only the function, we would be able to recompute the ?xed point value. Put differently, if we wish to transmit a value, we could instead transmit a function that iteratively converges on that value. This is the idea behind fractal compression. However, we are not interested in transmitting simple numbers, like “2”. Rather, we wish to transmit entire images. Our ?xed points will be images. Our functions, then, will be mappings from images to images. Our encoder will operate roughly as follows:
1. Given an image, i, from the set of all possible images, Image. 2. Compute a function f : Image → Image such that f (i) ≈ i. 3. Transmit the coef?cients that uniquely identify f . 53
Figure 26: Identifying selfsimilarity. Range blocks appear on the left; one domain block appears on the left. The arrow identi?es one of several collage function that would be composited into a complete image.
Our decoder will use the coef?cients to reassemble f and reconstruct its ?xed point, the image: 1. Receive coef?cients that uniquely identify some function f : Image → Image. 2. Iterate f repeatedly until its value converges on a ?xed image, i. 3. Present the decompressed image, i. Clearly we will not be using entirely arbitrary functions here. We want to choose functions from some family that the encoder and decoder have agreed upon in advance. The members of this family should be identi?able simply by specifying the values for a small number of coef?cients. The functions should have ?xed points that may be found via iteration, and must not take unduly long to converge. The function family we choose is a set of “collage functions”, which map regions of an image to similar regions elsewhere in the image, modi?ed by scaling, rotation, translation, and other simple transforms. This is vaguely similar to the search for similar macroblocks in MPEG P and Bframe encoding, but with a much more ?exible de?nition of similarity. Also, whereas MPEG searches for temporal selfsimilarity across multiple images, fractal compression searches for spatial selfsimilarity within a single image. Figure 26 shows a simpli?ed example of decomposing an image info collages of itself. Note that the encoder starts with the subdivided image on the right. For each “range” block, the encoder searchers for a similar “domain” block elsewhere in the image. We generally want domain blocks to be larger than range blocks to ensure good convergence at decoding time.
Fractal Compression in the Real World Fractal compression using iterated function systems was ?rst described by Dr. Michael Barnsley and Dr. Alan Sloan in 1987. Although they claimed extraordinary compression rates, the computational cost of encoding was prohibitive. The major vendor of fractal compression technology is Iterated Systems, cofounded by Barnsley and Sloan. 54
Today, fractal compression appears to achieve compression ratios that are competitive with JPEG at reasonable encoding speeds. Fractal compression describes an image in terms of itself, rather than in terms of a pixel grid. This means that fractal images can be somewhat resolutionindependent. Indeed, one can easily render a fractal image into a ?ner or coarser grid than that of the source image. This resolution independence may have use in presenting quality images across a variety of screen and print media.
9.3 ModelBased Compression
We brie?y present one last transform coding scheme, modelbased compression. The idea here is to characterize the source data in terms of some strong underlying model. The popular example here is faces. We might devise a general model of human faces, describing them in terms of anatomical parameters like nose shape, eye separation, skin color, cheekbone angle, and so on. Instead of transmitting the image of a face, we could transmit the parameters that de?ne that face within our general model. Assuming that we have a suitable model for the data at hand, we may be able to describe the entire system using only a few bytes of parameter data. Both sender and receiver share a large body of a priori knowledge contained in the model itself (e.g., the fact that faces have two eyes and one nose). The more information is shared in the model, the less need be transmitted with any given data set. Like wavelet compression, modelbased compression works by characterizing data in terms of a deeper underlying generator. Modelbased encoding has found applicability in such areas as computerized recognition of fourlegged animals or facial expressions.
55