Lossless compression in
lossy compression systems
Almost every lossy compression system contains a lossless
compression system
Lossy compression system
Transform
Quantizer
Lossless
Encoder
Lossless
Decoder
Dequantizer
Inverse
Transform
Lossless compression system
We discuss the basics of lossless compression first,
then move on to lossy compression
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 1
Topics in lossless compression
Binary decision trees and variable length coding
Entropy and bit-rate
Prefix codes, Huffman codes, Golomb codes
Joint entropy, conditional entropy, sources with memory
Fax compression standards
Arithmetic coding
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 2
Example: 20 Questions
Alice thinks of an outcome (from a finite set), but does not
disclose her selection.
Bob asks a series of yes/no questions to uniquely
determine the outcome chosen. The goal of the game is to
ask as few questions as possible on average.
Our goal: Design the best strategy for Bob.
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 3
Example: 20 Questions (cont.)
Which strategy is better?
0 (=no)
1 (=yes)
B C
0
1
1 0
1 C D 0
Observation: The collection of questions and answers yield
a binary code for each outcome.
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 4
Fixed length codes
B C
DE
F G
Average description length for K outcomes lav log2 K
Optimum for equally likely outcomes
Verify by modifying tree
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 5
Variable length codes
If outcomes are NOT equally probable:
Use shorter descriptions for likely outcomes
Use longer descriptions for less likely outcomes
Intuition:
Optimum balanced code trees, i.e., with equally likely outcomes, can
be pruned to yield unbalanced trees with unequal probabilities.
The unbalanced code trees such obtained are also optimum.
Hence, an outcome of probability p should require about
1
log 2 bits
p
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 6
Entropy of a random variable
Consider a discrete, finite-alphabet random variable X
Alphabet X { 0 ,1 , 2 ,..., K 1}
PMF f X x P X x
for each x X
Information associated with the event X=x
hX x log 2 f X x
Entropy of X is the expected value of that information
H X E hX X f X x log 2 f X x
x
Unit: bits
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 7
Information and entropy: properties
Information hX x 0
Information hX(x) strictly increases with decreasing
probability fX(x)
Boundedness of entropy
0 H ( X ) log 2
Equality if only one
outcome can occur
Equality if all outcomes
are equally likely
Very likely and very unlikely events do not substantially
change entropy
p log2 p 0 for p 0 or p 1
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 8
Example: Binary random variable
H X p log2 p (1 p)log 2 (1 p)
1
H X
0.9
0.8
Equally
likely
0.7
0.6
0.5
0.4
0.3
deterministic
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
Bernd Girod: EE398A Image and Video Compression
0.6
0.7
0.8
0.9
Entropy and Lossless Coding no. 9
Entropy and bit-rate
Consider IID random process X n (or source) where each
sample X n (or symbol) possesses identical entropy H(X)
H(X) is called entropy rate of the random process.
Noiseless Source Coding Theorem [Shannon, 1948]
The entropy H(X) is a lower bound for the average word length R of
a decodable variable-length code for the symbols.
Conversely, the average word length R can approach H(X), if
sufficiently large blocks of symbols are encoded jointly.
Redundancy of a code:
RH X 0
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 10
Variable length codes
X with alphabet
Given IID random process
PMF f X x
Task: assign a distinct code word, cx, to each element,
x X , where cx is a string of cx bits, such that each
symbol xn can be determined, even if the codewords cx
n
are directly concatenated in a bitstream
Codes with the above property are said to be
uniquely decodable.
Prefix codes
and
No code word is a prefix of any other codeword
Uniquely decodable, symbol by symbol,
in natural order 0, 1, 2, . . . , n, . . .
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 11
Example of non-decodable code
Encode sequence of source symbols 0 , 2 , 3 , 0 , 1
Resulting bit-stream
0 10 11 0 01
Encode sequence of source symbols 1 , 0 , 3 , 0 , 1
Resulting bit-stream
01 0 11 0 01
Same bit-stream for different sequences of source symbols:
ambiguous, not uniquely decodable
BTW: Not a prefix code.
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 12
Unique decodability: McMillan and Kraft conditions
Necessary condition for unique decodability [McMillan]
cx
Given a set of code word lengths ||cx|| satisfying McMillan
condition, a corresponding prefix code always exists [Kraft]
Hence, McMillan inequality is both necessary and sufficient.
Also known as Kraft inequality or Kraft-McMillan inequality.
No loss by only considering prefix codes.
Prefix code is not unique.
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 13
Prefix Decoder
Shift register
to hold longest
code word
...
Input buffer
Advance
||cx|| bits
Bernd Girod: EE398A Image and Video Compression
Code
word LUT
...
Code
word length
LUT
Entropy and Lossless Coding no. 14
Binary trees and prefix codes
Any binary tree can be
converted into a prefix code
by traversing the tree from
root to leaves.
Any prefix code corresponding
to a binary tree meets McMillan
condition with equality
cx
0
0
00
1
1
01 10
1
1
1100
1101
111
3 22 2 24 23 1
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 15
Binary trees and prefix codes (cont.)
Augmenting binary tree by two
new nodes does not change
McMillan sum.
Pruning binary tree does not
change McMillan sum.
1
1
Bernd Girod: EE398A Image and Video Compression
1
1
2 l
2
McMillan sum for simplest
binary tree
1
0
l 1
l 1
2 l
21 21 1
Entropy and Lossless Coding no. 16
Instantaneous variable length encoding
without redundancy
A code without redundancy, i.e.
Example
R H(X )
requires all individual code
word lengths
lk log 2 f X k
All probabilities would have to
be binary fractions:
f X ( k ) 2
lk
Bernd Girod: EE398A Image and Video Compression
H X 1.75 bits
R 1.75 bits
0
Entropy and Lossless Coding no. 17
Huffman Code
Design algorithm for variable length codes proposed by
Huffman (1952) always finds a code with minimum
redundancy.
Obtain code tree as follows:
1 Pick the two symbols with lowest probabilities and
merge them into a new auxiliary symbol.
2 Calculate the probability of the auxiliary symbol.
3 If more than one symbol remains, repeat steps
1 and 2 for the new auxiliary alphabet.
4 Convert the code tree into a prefix code.
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 18
Huffman Code - Example
R fixed 4 bits/symbol
Fixed length coding:
RHuffman 2.77 bits/symbol
Huffman code:
Entropy
H( X ) 2.69 bits/symbol
Redundancy of the Huffman code:
0.08 bits/symbol
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 19
Redundancy of prefix code for general distribution
Huffman code redundancy 0 1 bit/symbol
Theorem: For any distribution fX, a prefix code can be found, whose
rate R satisfies
Proof
H X R H X 1
Left hand inequality: Shannons noiseless coding theorem
Right hand inequality:
Choose code word lengths cx log 2 f X x
Resulting rate
R f X x log 2 f X x
x
f X x 1 log 2 f X x
H X 1
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 20
Vector Huffman coding
Huffman coding very inefficient for H(X) << 1 bit/symbol
Remedy:
Combine m successive symbols to a new block-symbol
Huffman code for block-symbols
Redundancy
1
H X R H X
m
Can also be used to exploit statistical dependencies between
successive symbols
m
Disadvantage: exponentially growing alphabet size X
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 21
Truncated Huffman Coding
Idea: reduce size of Huffman code table and maximum
Huffman code word length by Huffman-coding only the most
probable symbols.
Combine J least probable symbols of an alphabet of size K into an
auxillary symbol ESC
Use Huffman code for alphabet consisting of remaining K-J most
probable symbols and the symbol ESC
If ESC symbol is encoded, append
log 2 J bits to specify exact
symbol from the full alphabet
Results in increased average code word length trade off
complexity and efficiency by choosing J
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 22
Adaptive Huffman Coding
Use, if source statistics are not known ahead of time
Forward adaptation
Measure source statistics at encoder by analyzing entire data
Transmit Huffman code table ahead of compressed bit-stream
JPEG uses this concept (even though often default tables are
transmitted)
Backward adaptation
Measure source statistics both at encoder and decoder, using the
same previously decoded data
Regularly generate identical Huffman code tables at transmitter and
receiver
Saves overhead of forward adaptation, but usually poorer code
tables, since based on past observations
Generally avoided due to computational burden at decoder
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 23
Unary coding
Geometric source
Alphabet
0,1,...
PMF f X x 2 x 1 , x 0
Optimal prefix code with redundancy 0 is unary code
(comma code)
c0 "1" c1 "01" c2 "001" c3 "0001"
Consider geometric source with faster decay
1
PMF f X x 1 , with 0 ; x 0
2
x
Unary code is still optimum prefix code (i.e., Huffman code), but
not redundancy-free
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 24
Golomb coding
For geometric source with slower decay
1
PMF f X x 1 , with 1; x 0
2
x
Idea: Express each x as
x
xq
m
Distribution of new random variables
x mxq xr
with
m 1
f X q xq f X mxq i
xr x mod m
and
mxq
i 0
m 1
f i
i 0
1 xr
f X r xr
for 0 xr m
1 m
X q and X r statistically independent.
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 25
Golomb coding (cont.)
Golomb coding
1
2
m
Choose integer divisor
Encode xq optimally by unary code
Encode xr by a modified binary code, using code word lengths
ka log2 m
kb log2 m
Concatenate bits for xq and xr
In practice, m=2k is often used, so xr can be encoded by constant
code word length log 2 m
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 26
Golomb code examples
Unary
Code
1
01
001
0001
00001
000001
0000001
00000001
000000001
.
.
.
Unary
Code
10
11
Constant
010
011
length code
0010
0011
00010
00011
000010
m=2
000011
Unary
0000010
Code
0000011
00000010
00000011
.
.
.
Bernd Girod: EE398A Image and Video Compression
100
101
110
111
0100
0101
0110
0111
00100
00101
00110
00111
000100
000101
000110
000111
Constant
length code
m=4
.
.
.
Entropy and Lossless Coding no. 27
Golomb parameter estimation
Expected value for geometric distribution
E X 1 x
x
x 0
Approximation for E X 1
E X
1 E X
m
1
EX 2
1
EX
2
1
k max 0, log 2 E X
2
EX
1 E X
Rule for optimum
performance of
Golomb code
m 2k
Bernd Girod: EE398A Image and Video Compression
Reasonable setting,
even if E X 1
does not hold
Entropy and Lossless Coding no. 28
Adaptive Golomb coder (JPEG-LS)
Pick the
best
Golomb
code
Initial estimate
of mean
Initialize A x and N 1
For each n 0,1, 2,
A
Set k max 0, log 2
2
N
Code symbol xn using Golomb code with parameter k
If N N max
Set A
A / 2 and N
N / 2
Update A
A xn and N
N 1
Avoid overflow
and slowly
forget the past
Bernd Girod: EE398A Image and Video Compression
Entropy and Lossless Coding no. 29
