A SURVEY ON DIGITAL IMAGE COMPRESSION
Sun Huijie
1, 2
and Deng Tingquan
1
1
School of Computer Science and Technology, Harbin Engineering University, 150001, Harbin, China
2
College of Computer Science & Information Engineering, Harbin Normal University, 150080, Harbin, China
Keywords: Image encoding wavelet transform JPEG.
Abstract: Image compression enencoding is one of the key techniques in modern multimedia and communication
field. The methods of image encoding are of a great variety currently every kind of encoding methods all
exist respectively merit and shortcoming. This paper summarizes the traditiona image compression
enencoding methods and the modern image encoding methods and the hybrid image encoding methods. The
paper presents the direction of researching image compression enencoding next step.
1 INTRODUCTION
Image compression technology is in the field of
modern multimedia and key technology of
communication. Since 1948 digital TV signals to the
idea of proposed, science instant began after of
image compression research, has been 60 years of
history. In recent years, with the advent of the
information age and digital multimedia computer
technology development, no matter be traditional air
broadcast, cable TV companies or digital STBS,
television, mobile multimedia, makes people face
the various data quantity increased, signal is
transmitted by bandwidth when the restriction,
especially the wide application of computer network,
the more promoted the image compression
technology and related theory research and
development. In 1988, formed the ITU - T h. 261
draft, the draft to pass in 1990, marks the image
encoding the important step towards practical. This
paper mainly introduces the development history
and image encoding encoding method, introduced
the main technical image encoding.
2 THE RESEARCH STATUS
OF IMAGE ENCODING
Image encoding is present information science
research, one of the most active areas in pixels or
pixel blocks for encode entity image encoding
technology (such as the entropy encoding, transform
encoding, forecast encoding, motion compensation
mature gradually, etc) has been widely used in
JPEG, MPEG - 1, mpeg-2,, 261, h. h. 263 and other
international standards. Along with the development
of multimedia communication, computer,
communication, and consumer electronics, promoted
the crossover fusion image encoding technology
research, the new image encoding method are
continuously emerging.
Image compression technology from the time
development can be divided into two generations.
the first generation is based on statistics, removning
redundant data compression method l; The second
generation is based on the content, remove the
content is redundant, which are based on object
method called middle-level compression encoding
method based on grammar, senior encoding method
as the method.
2.1 The First Generation of Image
Encoding Technology
(1) Predictive Encoding
Predictive encoding predicts the current values with
transmited pixels, to encode the difference between
values and the actual value. Predictive encoding is
the earliest image encoding technology with motion
compensation frames between forecast encoding for
computational complexity is lower, facilitate real-
time realization and often for various image
encoding adopted by the international standard.
(2) Statistical Encoding
Statistical encoding say entropy encoding, classical
statistics encoding have Haffmann encoding,
arithmetic encoding and run-length encoding etc.
527
Huijie S. and Tingquan D..
A SURVEY ON DIGITAL IMAGE COMPRESSION.
DOI: 10.5220/0003591605270534
In Proceedings of the 13th International Conference on Enterprise Information Systems (ITLS-2011), pages 527-534
ISBN: 978-989-8425-56-0
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
Hoffmann encoding method is based on the source
of various symbols appear the probability of
encoding, the encode is simple and effective. The
arithmetic encoding is completely abandoned with
special characters instead of input characters
thoughts, it is to the data with 0 to 1 between the
floating-point number for enencoding, when the
source of the probability of symbols is more
adjacent, arithmetic encoding efficiency than
hoffmann encoding, but the realization of the
arithmetic encoding than hoffmann encoding is more
complex. The run-length encoding is relatively
simple encoding technology, it is a zero called run-
length, convert instead of special character, reducing
the amount of data, mainly used in image quantified
appear under the condition of the continuous zero.
(3) Transform Encoding
Transform encoding is a certain function transform,
from a representation space change to another
representation space, then transform domain, on the
transformation of signal encoded. This
transformation encoding essence is to pass transform
the way of the original image energy mainly
concentrated in a few parts of the coefficient, so can
more easily to do image compression.
2.2 The Second Generation of Image
Encoding Technology
The traditional encoding method has many
shortcomings, such as high compression ratio restore
images appear serious square effect, the human
visual characteristics not easy is introduced to the
compression algorithm. To overcome the
shortcomings of traditional compression method
have been put forward several new coding method
based on wavelet transform, compression method,
fractal compression method and neural network
method, etc..
(1) Wavelet Transform Method
The theory of wavelet transform in recent years is
the emergence of new branch of mathematics, which
is the Fourier transform again after a landmark
development. Now, wavelet analysis method has
been widely used in signal processing, image
processing, pattern recognition, speech recognition,
seismic exploration, CT imaging, computer vision,
aviation and aerospace technology, fault monitoring,
communication and electronic systems and so on
themultitudinous disciplines and related technology
research. Wavelet image compression is by using
wavelet transform and has good spatial resolution
and the frequency resolution character, make the
energy and transform coefficient in frequency and
space, so as to achieve the concentration of
removing pixel redundancy role.
(2) Fractal Compression Method
In various multimedia services and digital
communication and other fields of application,
image compression/coding is crucial technology.
The vast literature published in recent years in
display, image coding has made important progress,
many new ideas are proposed. Fractal coding is
among them one of the most prominent technology,
it opened a new image compression coding ideas.
Since the early 1990s, fractal coding has more than
ten years in short has made remarkable achievement.
Barnsley fractal coding is put forward by the
first iteration function system, from the fractal
geometry theory (the important composition part). In
fractal coding, an image from a make it approximate
constant compression affine transformation said
reconstruction images is compressed transform fixed
point, compression affine transformation of the
parameters of the original image fractal yards.
Therefore, an image fractal coding is looking for a
suitable compression affine transformation, its fixed
point is the original image possible good
approximation. Fractal decoding is a relatively
simple rapid iteration process, decoded image fractal
codes by compressed transform iterative function
said in any initial image to approach.
Fractal image coding is the search for the basic
ideas of image among different regions under
different scales similarities. Therefore, and usually,
as the image coding method of fractal coding system
design of the first step is for image segmentation,
which divided into some taller image for coding
regions (R block), each branch area in the images of
the corresponding to an object or object, the next
part of the main steps of each branch area is its
affine similar for large area (D block). As such, each
for a group of block R affine transform coefficient,
regardless of the segmentation information and if,
then nearly yards coding coefficient fractal codes is
proportional to the file size. The number of pieces of
R Therefore, partition is the key factor than
determines compression.
Segmentation is to determine the decoded image
quality and a key factor, a good segmentation
scheme should reflect the image similarity across the
scale. Image both smooth uniform regions
(brightness constant or slow-moving area), and have
high contrast area (such as edge regions). In uniform
regional part, use large can achieve good collage,
meanwhile, high contrast area are need to use small
size block just might come to hope the image
quality. To achieve this, must adopt more flexible
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
528
segmentation method (variable size block division).
Contains coded quardtree segmentation, the various
variable size block segmentation method has been
widely adopted by the fractal coding literature.
Fractal coding (a compressed affine
transformation description) is a division of
information and quantitative transform parameter of
different segmentation scheme, the size of the
segmentation occupies information is different.
From compressed perspective, segmentation scheme
possession of information has been jumped over
lesser, but if an irregular segmentation brings good
subjective and objective coding quality, spending a
little coding cost is worth it. Therefore, in the
influence of fractal image coding performance of
various factors, part should be one of the most
important factors. Choose which segmentation
scheme need to weigh the compression ratio (or
digital rate) and coding quality and then selecting a
compromise plan.
(3) Neural Network Method
PCNN (pulse coupled neural network) is a kind of
irregular segmented regions based on the image
compression method. PCNN itself to the image
details have better keep characteristics, blurred
image after still can reach good segmentation effect.
The role of local connected domain PCNN and
threshold means that has similar properties, make
attenuation of gray-scale characteristics can
simultaneously in near pixels, constitutes the
activated PCNN segmentation characteristics of the
foundation. Through the various parameters PCNN
model adjust, can make the image segmentation
results can better contains the original image detail
information, and can avoid some meaningless small
segmentation pieces of produce. In keeping the
detail of image PCNN has incomparable advantage
over traditional coding, but its reconstruction effect
is not good, so this method is also a recent research
hot spot of experts and scholars.
Another is worth mentioning, in recent years, the
application of mathematical morphology covering
the image processing almost all areas, including
character recognition, medical image processing,
image compression, visual inspection, materials
science and robot vision, etc. Mathematical
morphology is a new nonlinear image signal
processing and analysis theory, it rejected traditional
numerical modeling and analysis of the Angle of
view and set to depict and analyze image, with a
complete theory, method and algorithm system.
Mathematical morphology is a comprehensive
multidisciplinary intellectual crossover science, its
theoretical basis is quite deep, but the basic principle
is simple. So far, there is no one way to like as
mathematical morphology both solid theoretical
foundation, concise, simple, unified basic idea, but
also has so extensive practical application value.
Therefore, someone will morphology and wavelet
transform combined with DCT transform or
application in image compression, also made a
surprising results, this undoubtedly for image coding
method to develop the field of the ideas and
direction, more enriched image coding method.
3 THE IMAGE COMPRESSION
TECHNOLOGY RELATED
BASIC THEORY
3.1 Information Measure
The basic principle of image compression originated
from the 1940s Shannon (Shannon) information
theory. Shannon's coding theorem tells us, in not any
distortion, through before the coding, for every
reasonable distribution of a source code words differ
long symbols, average code length can be arbitrary
close to source entropy.
In information theory with the "entropy" to
measure the size of the information. For individual
events (such as a character) speaking, its entropy
defined as:
H (i) = - log
2
(Pi) (bit) (1)
Type (1) says the probability of occurrence in the
event for Pi (characters) has the information.
Measure the size of the unit is "information from"
(bit). Its physical meaning for said the event
(characters) need at least digits.
Although shannon, discusses the information
coding should follow the rule, but did not give a
specific.
A message queue average information entropy
defined as:
() ()log(())
2
1
n
Hi px px
ii
i
=−
=
(2)
Type (2) the p (xi) says an event, xi the
probability of occurrence in. The probability of one
incident, the smaller the information entropy, the
higher the amount of information contained.
A SURVEY ON DIGITAL IMAGE COMPRESSION
529
3.2 Data Coding and Compression
Concept
Say simply, so-called compression is trying to
remove all sorts of redundancy, keep really useful
information. To signal compression called encoding.
Restoration of compression information process
called decoding.
Data compression was originally an important
topic in the study of information theory, in
information theory called the source coding,
shannon information theory tells us that source
entropy is source code without distortion of the
limit. That is to say, no matter what compression
algorithm, which compressed digital rate is not less
than the data of entropy, if less than words, this
compression is necessarily distortion, and a
distortion of source code, and to follow the
information rate-distortion function in relationships.
3.3 The Classification of Data
Ompression
Although shannon, discusses the information coding
should follow the rule, but did not give a specific
coding method. Therefore, the coding researchers
continuously put forward various coding method.
Several forms of the space because a signal is
correlated, such as storage space reduce also means
transmission efficiency and occupy bandwidth
saves, that is, as long as adopt some methods to
reduce a signal space, can compress data.
Data compression method of classification, and
yet many unified. Data compression is consisted
with entropy encoding and entropy compression.
Among them, Huffmann encoding and LZW coding
is relatively commonly used, wavelet transform
coding, fractal coding for relatively frontier
compression technology.
3.4 The Basic Principle of Image
Compression
Judging from the perspective of the information
theory, compression is to remove the redundant
information uncertainty, namely reserves, remove
the information to determine the information
(know), and it is more close to information in a
description of nature to replace the original
redundancy description.
Image compression implementation principle of
two aspects:
(1) The original image data (stationary or
exercise) exist great redundancy, such as still images
between adjacent pixels in spatio-temporal
correlation between before and after the moving
pictures and temporal correlation are large, source
have redundant.
(2) The second is the application of the
multimedia system in the field, who is a major
recipient of image information, the eyes are receiver,
so it might be possible to use visual drastic changes
to edge is not sensitive (masking effect) and eye
vision of image information sensitivity, but for
brightness color resolution weak etc physiology
characteristic to realize high compression ratio, and
make by compressed data recovery image signal still
have satisfactory visual quality.
Develop multimedia application system, meet
the greatest obstacle to progress is the multimedia
information huge data quantity of data acquisition,
storage, processing and transmission. Among them,
the largest amount of data is digital image
information. For example: a picture of a resolution
color images 640 x 480 (24 bits per pixel), the
amount of data about 0.92 MB. If again with 30
frames per second, the video signal speed playback
volume of data as high as 27.6 MB. If the cd-rom
650M in again, without considering audio signal,
each piece CDS also can only play and 24 seconds.
Obviously, image compression technology is one of
the key technology of multimedia technology.
4 THE IMAGE COMPRESSION
STANDARD JPEG
4.1 JPEG Background
JPEG (Group) is a Photographic has or by ISO and
IEC two organizations together, an image panel, be
responsible for making static digital image data
compression coding standard, this Group developed
algorithm called JPEG arithmetic, JPEG image has
become common international standards of its
applicable scope, is gray image, image with color,
the compression of still images, video sequence
frame image compression; JPEG can adjust big
range code-rate and quality image.
The core of JPEG is mainly DCT and DPCM.
4.2 JPEG Operating Mode
For an image component, JPEG stipulated the four
operating mode:
(1) The order of dct-based coding mode (baseline
CODEC).
The order of dct-based coding mode is the single
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
530
time scanning to complete a image component
coding, scanning the order from left to right, from
the top down. Its algorithm basic steps as follows:
A. Will with original image brightness,
chromatism says (component image sampling 4:1:1)
B. Into 8 x 8 pieces of data [0 to 255], [- convert
128 ~ 127]
C. Positive discrete cosine transform (FDCT)
D. Quantification (quantization)
E.Z Glyph quantitative results (zigzag arranged
scan)
F. Use of DC coefficient DPCM code (DC)
G. Use to exchange coefficient trip coding code
(AC)
H. Entropy entropy coding (for): hoffmann or
arithmetic coding
(2) Based on DPCM (difference pulse code
modulation) nondestructive coding mode
Based on DPCM nondestructive coding model
mainly adopts three neighborhood two-dimensional
forecast coding and entropy coding, fig.1 shows:
Figure 1: DPCM Predicting coding.
This mode and sequence model coding steps,
basically the same differs each image of increasing
mode to weight after multiple scanning coding to
finish. The first scan only a rough compression, then
according to the data reconstruction a picture first,
the image quality is low after scanning to make fine
scanning, make the reconstructed image quality
enhances unceasingly, until satisfaction,
nondestructive coding compression ratio can reach
2:1.
(3) The gradual dct-based coding mode
The gradual dct-based coding mode is through
multiple scanning an image component coding,
provides a from coarse to fine gradual streaming
structure. Mainly divided into the following two
modes:
A. Press band: a scan, gradual only to the certain
frequency conversion DCT coefficients of the code
Transmitted, and then to other band of
progressive way encoded and transmitted, until the
end of all coefficient relay
B. Bitwise gradual: according to its digital of
DCT coefficients from high to low into segments,
which in turn into to paragraphs
Do compression coding, first on the most
effective bit code of N a transmission, until the
transfer over. All coefficient
(4) Dct-based coding mode of layered
Dct-based coding model algorithm of layered
basic step as follows:
A. Reduce spatial resolution of the original
image.
B. Has been reduced to the resolution of the
image according to order coding mode compression
and stored or transmitted.
C. On low resolution image, then use by
decoding the improvement of image interpolation
method of resolution.
D. Will increase as the image resolution have
predictive value of original image, and put it and
poor value of original image dct-based coding.
E repeat steps c, d until achieve full resolution
images.
5 IMAGE MIXED ENCODING
Currently image compression algorithm sort is
various, classic compression algorithm theory has
more mature, as people in these traditions coding
method of thorough research and application of
these methods, found many shortcomings, such as
high compression ratio restore images appear serious
square effect, the human visual characteristics not
easy is introduced to the compression algorithm, in
order to overcome the shortcomings of traditional
compression method, puts forward a new coding
method except outside, another important solution is
take different coding method combining now, many
of the ideas in mixed domestic and foreign scholars
have the image coding method research attempts
were made, and made a lot of achievements. It
should be noted that many international codec
standards also USES a hybrid coding scheme
introduced here several typical hybrid image coding
algorithm of principle, implementing methods and
research status.
5.1 Dct-based Transform Hybrid
Coding
5.1.1 Dct-based Transform the Fractal
Image Coding
Fractal image coding is mainly the mathematical
basis of fractal geometry iteration function system
theory, the theory of fixed point and collage
theorem. Iterative Function System (IFS) are Iterated
views, and its research group in Barnsley's proposed
Hutchinson 1981 Iterated Function Theory (Iterated
A SURVEY ON DIGITAL IMAGE COMPRESSION
531
guys) are developed on the basis of. Fixed point
theory is an important branch of functional analysis,
it ensures that any of the iteration function system
exists only attractor, and this iteration function
system with initial set the attractor is irrelevant.
Collage theorem is Banach fixed point theorem for
coding the simple inference, problems are times
optimal solution provides an answer. The task of
fractal image coding is to seek for code iteration
function the parameters of the system image, fractal
image decoding is according to the parameters of the
iteration function system to find out the attractor.
Iterative function system is to use a mathematical
system to construct a categories analytical man-
made or natural, has the similar structure since affine
fractal, i.e. use simple iterative function system
(IFS) coding can generate and natural scenery
similar has boundless complex graphics, and for a
given any an image, fractal coding thoughts are
looking for a IFS makes attractor is original image
or the approximation of the original image.
Dct-based transform the flow of fractal image
coding as shown in figure 4 shows.
5.1.2 Based on the Forecast Coding and
DCT Combination of Image Coding
And the forecast method is the most simple and
practical video compression coding method, through
coding not pixels after transmission, but the
sampling value itself the sampling forecast value and
the actual value difference. Because the same image
of adjacent pixels with correlation between, so take
advantage of these strong relevance to forecast
coding.
Frame forecast coding method mainly have the
best forecast coding, considering subjective visual
effect quantizer and linear predictor, predictive
coding. The most commonly used difference pulse
code modulation (DPCM), its coding model shown
in figure 5 shows.
5.1.3 Wavelet Transform And Out of
Combining Neural Network Mixed
Encoding
Because neural network has massively parallel
processing and distributed information storage
advantage, good adaptability, self-organizing and
fault tolerance, a strong learning and associative
memory function, and artificial neural network has
many brain and a similar information processing
capability, with strong data compression ability, can
put the neural network theory applied to image
coding. In 1985, Hinton etc who first Ackley and the
multilayer neural network models for data
compression transform. In 1990, Z.H e and H.L I use
multilayer feedforward neural network nonlinear
predictive coding for image. At present, according to
the neural network in the image compression, the
application of the model and the algorithm can be
applied types concluded to the following four:
(1) Use of data compression characteristics with
neural network, namely direct implementation image
compression using neural network realization vector
quantization algorithm. If Kohonen self-organizing
feature mapping (SOFM) neural network for code
book design vector quantization approach has not
easily affected by the influence of initial code book,
and can keep the topological structure of image data
etc.
(2) Neural network is applied to compression,
indirect based on neural network coding method, the
image transformation in the existing algorithm used
to implement a local stage, some of these steps. If
use neural network to realize the orthogonal
transformation of orthogonal transform code Hofield
neural network operation, including image
transformation coding. In addition, neural network
transform code are not limited to the orthogonal
transformation, can be extended to the orthogonal,
nonlinear, dimension reduction or fractal transform,
can make the image in the transformation process
directly compressed.
(3) Principal component analysis neural network
coding. Principal component analysis (PCA) is a
linear transform feature space dimension reducing
feature selection method, it only retains the data, the
main components of the dropped relatively minor
component, achieved the purpose of data
compression. Principal component analysis (PCA) is
a kind of effective image transformation coding
algorithm, it can draw the main characteristics of
image data in reducing weight, and therefore, can
input data dimension image of compressed image
but also to minimize distortion. In addition to the
principal component analysis and Kohonen self-
organizing feature mapping (SOFM) algorithm
combining used in image compression.
(4) Neural network and the existing some
combination of image compression algorithm coding
method, i.e. put some advanced algorithm developed
into learning algorithm and establishing neural
network model. Such as wavelet neural network by
network and nature.these, cent predicted neural
networks.
The principle mixed encoding scheme is shown
as shown in figure 9.
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
532
5.2 Other Types of Mixed Image
Coding Method
Based on the genetic algorithm is this image fractal
compression encoding.
Based on genetic algorithm is fractal image
compression with matching block the coordinates (x,
y) and the regional blocks of rotation transformation
(consists of eight kinds of rotating), chromosome in
fractal coding algorithm, each need to match with
genetic algorithm to search, according to defined
blocks of chromosome and fitness function (such as
regional blocks and matching block matching mean-
square error) to search for, when chromosome
maternal convergence optimal individuals when as
determined from the matching block position and
transform parameter is matching result i.e. coding.
In addition, there are a lot of genetic algorithm based
on image fractal compression of improvement, such
as searching for optimal use of domain block
domain block two parameters is relative to the range
block horizontal and vertical displacement of the
weight, will the improved genetic algorithm is
applied to fractal coding, the improved genetic
algorithm search speed, overcome fractal
compression classification matching algorithm, the
local optimal and random search problem.
GA used in fractal image compression, enhance
the compression ratio and compression precision,
because in high compression ratio descend signal-to-
noise ratio has been greatly improved, so it can be
used in low bit rate image compression. Moreover,
GA has can parallel computation of fractal features,
and can reduce the computation time, quickly find
optimal solutions. But the control parameters,
experiment, most will depend on experience many,
therefore, how to get adaptive to control these
parameters, further improve the compression ratio
and decoding quality, still remain in research and
exploration. Because the good properties, it GA: s
combined with fractal image coding method broad
prospect of application.
The image based on neural network 4.3.2 fractal
compression encoding.
Image coding technology, fractal image
compression has higher compression ratio and low
attrition, but the biggest shortage is iterative function
system (IFS) automatic fractal image coding method
to calculate the amount is large, thus limiting its
practical application, and artificial neural network is
learning, memory, identification and reasoning
features, using neural network method to parallel
waycompletes fractal image compression coding,
plenty of calculation for fractal image compression
application provides a new solution.
For the first time, the application of neural
networks is presented Stark IFS Hopfield neural
network based on the fractal image coding method,
can effectively solve the problem of linear
accumulative total, but meanwhile neural network
method are only used for IFS unzip process. The
application of neural networks with fractal image
compression encoding principle is compressed using
neural network model, make the fractal codes
automatic acquisition, specifically is one neuron in
the images of the a pixel, representing the weight
and threshold as fractal codes, appropriate weight
and threshold value can be obtained in the
compression process, the initial image compression
process in solution can be constructed out. This
method is put forward two different neural network
model with fractal image compression solution
compress, these two models of the architecture is
same, mainly adopts define different linear model
and the nonlinear model of functions. The results
show that, with the basic automatic fractal image
coding method compared, in the basic guarantee for
the quality of reconstructed image premise, the
operation time and bit rate decreased, visible neural
network technology used in fractal image
compression and unzip the feasibility and efficiency.
At present, the neural network in the research of
fractal image coding are not much, but neural
network of parallel
Computation ability to solve the fractal image
compression of calculation problem is very
important and deserves further study.
6 SUMMARY
The existing various image compression method,
any kind of coding algorithm is not perfect. Wavelet
transform existing wavelet base selection and its
calculating complex and difficult high compression
ratio Gibbs effects exist the problems to be solved;
Fractal coding also exists matching operation
problem; DCT coding exist block effect, etc. Only
make full use of the advantages of different coding
method to use it to its mix of win-win results. This
paper introduces the traditional encoding method
and modern coding method, and the emphasis on the
hybrid coding various technical, future will further
study of image coding, truly realize image
compression ratio and reconstruction quality win-
win.
A SURVEY ON DIGITAL IMAGE COMPRESSION
533
REFERENCES
Abdallah, E. E, Hamza, A. B., & Bhattacharya, P. An
improved image watermarking schemeusing fast
Hadamard and discrete wavelet transforms. Journalof
Electronic Imaging, 2007.16(3), 033020-033020
Iqbal, M. A., Javed, M. Y., & Qayyum, U. Curvelet-based
image compression with SPIHT. In International
conference on convergence information technology.
IEEE ComputerSociety.2007.
Seo, K. K. Anapplication of one-classsupport vector
machines in content based imageretrieval. Expert
Systems with Applications, 2007.33(2), 491-498.
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
534