21-05-2012, 10:40 AM
ARTIFICIAL NEURAL NETWORKS FOR COMPRESSION OF
DIGITAL IMAGES
[attachment=22519]
ABSTRACT
Digital images require large amounts of memory for storage. Thus, the transmission of an image from one
computer to another can be very time consuming. By using data compression techniques, it is possible to
remove some of the redundant information contained in images, requiring less storage space and less time
to transmit. Artificial Neural networks can be used for the purpose of image compression. Successful
applications of neural networks to vector quantization have now become well established, and other aspects
of neural networks for image compression are stepping up to play significant roles in assisting the
traditional compression techniques. This paper discusses various neural network architectures for image
compression. Among the architectures presented are the Kohonen self-organized maps (KSOM),
Hierarchical Self-organized maps (HSOM), the Back-Propagation networks (BPN), Modular Neural
networks (MNN), Wavelet neural Networks, Fractal Neural Networks, Predictive coding Neural Networks
and Cellular Neural Networks (CNN). The architecture of these networks, and their performance issues are
compared with those of the conventional compression techniques.
Keywords: Neural Networks, Image compression,Vector quantization
1. INTRODUCTION
Digital image compression is a key technology in
the field of communications and multimedia
applications. A large number of techniques have
been developed [13] to make the storage and
transmission of images economical. These
methods can be lossy or lossless. Over the last
decade, numerous attempts have been made to
apply artificial neural networks (ANNs) for image
compression [9, 34, and 35]. The Kohonen selforganizing
maps (SOM) [16],[1], Hierarchical
SOMs (HSOMs)[3], Back-propagation networks
(BPNs) [15], Modular Neural networks (MNNs)
[32], and Cellular Neural Networks[30 ], among
others have been proposed in the literature.
Research on neural networks for image
compression is still making steady advances. In
this paper, we discuss the various neural network
architectures for compression of still images. A
comparison on the performance of these networks
is made with those of the conventional algorithms
for image compression. This paper consists of five
sections. Section 2 discusses the SOMs and
HSOMs for image compression using VQ. Section
3 describes the various back propagation neural
networks for image compression. In Section 4
modular neural networks are presented. Wavelet
and Fractal neural networks are discussed in
section 5. Cellular neural networks are described
in section 6. Finally, Section 7 gives the
conclusions and the scope for future research work
in this area.
2. SELF-ORGANIZING MAPS (SOMS)
2.1 Kohonen’s Self-organizing maps (KSOM)
One of the common methods to compress images
is to code them through vector quantization (VQ)
techniques [16]. VQ is a lossy compression
technique. First, the image is split in to square
blocks of size τxτ (4x4 or 8x8) pixels; each block
is considered as a vector in a 16 or 64 dimensional
spaces. Second, a limited number (l) of vectors
International Journal of Reviews in Computing
© 2009-2010 IJRIC& LLS. All rights reserved. IJRIC
ISSN: 2076-3328 www.ijric.org E-ISSN: 2076-3336
76
(code words) in this space is selected in order to
approximate as much as possible, the distribution
of initial vectors extracted form the image. Third,
each vector from the original image is replaced by
its nearest codeword. Finally, during transmission,
the index of the codeword is transmitted.
Compression is achieved if the number of bits used
to transmit the index (log2 l) is less than the
number of initial bits of the block (τ x τ x m);
where m is the number of bits per pixel.
Kohonen’s self-organized feature map (KSOM) is
a reliable and efficient method to achieve VQ for
image compression. The basic structure of KSOM
is shown in Figure 1. Compression ratios of 10:1
to 100:1 have been reported using the SOM [25].
The search complexity is of order O (N).
Figure 1 SOM Neural Network for Vector Quantization
22 Hierarchical Self-organizing maps (HSOM)
HSOM is an extension of the conventional SOM.
A tree-structure is defined, where each node is a
SOM, trained with one determined data set. The
map in level-1 is trained with the full set of data,
and in accordance with the quantization of each
neuron, the map children are trained with
subgroups of this. Fig.2 illustrates the
configuration of the HSOM. The trained HSOM is
executed sequentially, i.e., from the highest to the
lowest level of the tree. The complexity of
search is reduced to O(log N) in the HSOM.
Figure 2 Configuration of Hierarchical SOM
3. BACK-PROPAGATION NETWORKS
3.1. Basic back-propagation neural network
Back-propagation neural networks are directly
applied to image compression [9, 24]. The neural
network structure is shown in Figure 3. Three
layers, one input layer, one output layer and one
hidden layer are designed. The input layer and
output layer are fully connected to the hidden
layer. Compression is achieved by designing the
value of K, the number of neurons at the hidden
layer less than that of neurons at both input and the
output layers. The input image is split up into
blocks or vectors of 8x8, 4x4 or 16x16 pixels.
When the input vector is referred to as Ndimensional
which is equal to the number of pixels
included in each block, all the coupling weights
connected to each neuron at the hidden layer can
be represented by {wji , j=1, 2,…, K and i=1, 2,…,
N}, which can also be described by a matrix of
order KxN. From the hidden layer to the output
layer, the connections can be represented by {w’ij :
1≤ i ≤ N, 1 ≤ j≤ K} which is another weight matrix
of order NxK. Image compression is achieved by
training the network in such a way that the
coupling weights, {wji }, scale the input vector of
N-dimension into a narrow channel of Kdimension
(K<N) at the hidden layer and produce
optimum output value which makes the quadratic
error between input and output minimum.
Figure 3 back propagation Neural Network
3.2. Hierarchical back-propagation neural network
The basic back-propagation network is further
extended to construct a hierarchical neural network
International Journal of Reviews in Computing
© 2009-2010 IJRIC& LLS. All rights reserved. IJRIC
ISSN: 2076-3328 www.ijric.org E-ISSN: 2076-3336
77
by adding two more hidden layers into the existing
network as proposed in [24]. The Hierarchical
neural network structure is illustrated in Fig. 4 in
which the three hidden layers are termed as the
combiner layer, the compressor layer and the decombiner
layer. The idea is to exploit correlation
between pixels by inner hidden layer and to exploit
correlation between blocks of pixels by outer
hidden layers
Figure 4. Hierarchical Neural Network Structure
From the input layer to the combiner layer and
from the de-combiner layer to the output layer,
local connections are designed which have the
same effect as M fully connected neural subnetworks.
As seen in Fig. 4, all three hidden layers
are fully connected. The basic idea is to divide an
input image into M disjoint sub-scenes and each
sub-scene is further partitioned into T pixel blocks
of size pxp. For a standard image of 512x512, as
proposed [24], it can be divided into 8 sub-scenes
and each sub-scene has 512 pixel blocks of size
8x8. Accordingly, the proposed neural network
structure is designed to have the following
parameters:
The total number of neurons at the input layer is
Mxp2 = 8x64 =512. Total number of neurons at the
combiner layer is MxNh =8x8=64. Total number of
neurons at the compressor layer is Q = 8. The
total number of neurons for the de-combiner layer
and the output layer is the same as that of the
combiner layer and the input layer, respectively. A
so-called nested training algorithm (NTA) is
proposed to reduce the overall neural network
training time, which comprises the following
steps:
Step1: Outer loop neural network (OLNN)
training.
Step2: Inner loop neural network (ILNN) training.
Step3: Reconstruction of the overall neural
networks.
After training is completed, the neural network is
ready for image compression in which half of the
network acts as an encoder and the other half as a
decoder. The neuron weights are maintained the
same throughout the compression process.
3.3. Adaptive back-propagation neural network
Further to the basic narrow channel backpropagation
image compression neural network, a
number of adaptive schemes are proposed [4, 9]
based on the principle that different neural
networks are used to compress image blocks with
different complexity. The general structure for the
adaptive schemes are shown in Figure-4 in which a
group of neural networks with increasing number
of hidden neurons, (hmin, hmax), is designed. The
basic idea is to classify the input image blocks into
a few sub-sets with different features according to
their complexity measurement. A fine tuned neural
network then compresses each sub-set. Four
schemes are proposed [9] to train the neural
networks which are classified as parallel training,
serial training, activity-based training and activity
and direction based training schemes.
3.4. Performance Considerations
Considering the different settings for the
experiments reported in various sources
[9],[23],[24], it is difficult to make a comparison
among all the algorithms presented in this section.
Fig. 5 Adaptive Neural Network Structure
International Journal of Reviews in Computing
© 2009-2010 IJRIC& LLS. All rights reserved. IJRIC
ISSN: 2076-3328 www.ijric.org E-ISSN: 2076-3336
78
To make the best use of all the experimental
results available, we take Lena as the standard
image sample and summarize the related
experiments for all the algorithms as illustrated in
Table1 which are grouped into basic back
propagation, hierarchical back propagation and
adaptive back propagation.
Table 1 Basic Back Propagation on Lena (256x256)
Table 2 Hierarchical Back Propagation on Lena
(512x512)
Table 3 Adaptive Back Propagation on Lena (256x256)
4. MODULAR NEURAL NETWORKS
Though a single neural network can compress the
average characteristic for the image data, it is
difficult to compress both the edge and flat regions
with the same precision. For this, Rahim et al.
have proposed the compression method using two
neural networks [26]. One of the neural networks
is used for the compression of the original image
and the other is used for the compression of the
residual image. However, it is effective to
compress for each region, which is divided in to
the edge and flat regions. A modular structured
neural network consisting of multiple neural
networks with different block sizes (the number of
input units) for region segmentation has been
proposed by Watanbe and Mori [32]. By the
region segmentation, each neural network is
assigned to each region such as the edge or the flat
region.
From simulation results, it is shown that the
proposed method yields a better compression
compared with the conventional compression
technique using a single neural network.
5. NEURAL NETWORK DEVELOPMENT OF EXISTING TECHNOLOGY
In this section, we show that the existing
conventional image compression technology can
be developed right into various learning algorithms
to build up neural networks for image
compression. This will be a significant
development in the sense that various existing
image compression algorithms can actually be
implemented by one neural network architecture
empowered with different learning algorithms.
Hence, the powerful parallel computing and
learning capability with neural networks can be
fully exploited to build up a universal test bed
where various compression algorithms can be
evaluated and assessed. Three conventional
techniques are covered in this section, which
include wavelet transforms, fractals and predictive
coding.
DIGITAL IMAGES
[attachment=22519]
ABSTRACT
Digital images require large amounts of memory for storage. Thus, the transmission of an image from one
computer to another can be very time consuming. By using data compression techniques, it is possible to
remove some of the redundant information contained in images, requiring less storage space and less time
to transmit. Artificial Neural networks can be used for the purpose of image compression. Successful
applications of neural networks to vector quantization have now become well established, and other aspects
of neural networks for image compression are stepping up to play significant roles in assisting the
traditional compression techniques. This paper discusses various neural network architectures for image
compression. Among the architectures presented are the Kohonen self-organized maps (KSOM),
Hierarchical Self-organized maps (HSOM), the Back-Propagation networks (BPN), Modular Neural
networks (MNN), Wavelet neural Networks, Fractal Neural Networks, Predictive coding Neural Networks
and Cellular Neural Networks (CNN). The architecture of these networks, and their performance issues are
compared with those of the conventional compression techniques.
Keywords: Neural Networks, Image compression,Vector quantization
1. INTRODUCTION
Digital image compression is a key technology in
the field of communications and multimedia
applications. A large number of techniques have
been developed [13] to make the storage and
transmission of images economical. These
methods can be lossy or lossless. Over the last
decade, numerous attempts have been made to
apply artificial neural networks (ANNs) for image
compression [9, 34, and 35]. The Kohonen selforganizing
maps (SOM) [16],[1], Hierarchical
SOMs (HSOMs)[3], Back-propagation networks
(BPNs) [15], Modular Neural networks (MNNs)
[32], and Cellular Neural Networks[30 ], among
others have been proposed in the literature.
Research on neural networks for image
compression is still making steady advances. In
this paper, we discuss the various neural network
architectures for compression of still images. A
comparison on the performance of these networks
is made with those of the conventional algorithms
for image compression. This paper consists of five
sections. Section 2 discusses the SOMs and
HSOMs for image compression using VQ. Section
3 describes the various back propagation neural
networks for image compression. In Section 4
modular neural networks are presented. Wavelet
and Fractal neural networks are discussed in
section 5. Cellular neural networks are described
in section 6. Finally, Section 7 gives the
conclusions and the scope for future research work
in this area.
2. SELF-ORGANIZING MAPS (SOMS)
2.1 Kohonen’s Self-organizing maps (KSOM)
One of the common methods to compress images
is to code them through vector quantization (VQ)
techniques [16]. VQ is a lossy compression
technique. First, the image is split in to square
blocks of size τxτ (4x4 or 8x8) pixels; each block
is considered as a vector in a 16 or 64 dimensional
spaces. Second, a limited number (l) of vectors
International Journal of Reviews in Computing
© 2009-2010 IJRIC& LLS. All rights reserved. IJRIC
ISSN: 2076-3328 www.ijric.org E-ISSN: 2076-3336
76
(code words) in this space is selected in order to
approximate as much as possible, the distribution
of initial vectors extracted form the image. Third,
each vector from the original image is replaced by
its nearest codeword. Finally, during transmission,
the index of the codeword is transmitted.
Compression is achieved if the number of bits used
to transmit the index (log2 l) is less than the
number of initial bits of the block (τ x τ x m);
where m is the number of bits per pixel.
Kohonen’s self-organized feature map (KSOM) is
a reliable and efficient method to achieve VQ for
image compression. The basic structure of KSOM
is shown in Figure 1. Compression ratios of 10:1
to 100:1 have been reported using the SOM [25].
The search complexity is of order O (N).
Figure 1 SOM Neural Network for Vector Quantization
22 Hierarchical Self-organizing maps (HSOM)
HSOM is an extension of the conventional SOM.
A tree-structure is defined, where each node is a
SOM, trained with one determined data set. The
map in level-1 is trained with the full set of data,
and in accordance with the quantization of each
neuron, the map children are trained with
subgroups of this. Fig.2 illustrates the
configuration of the HSOM. The trained HSOM is
executed sequentially, i.e., from the highest to the
lowest level of the tree. The complexity of
search is reduced to O(log N) in the HSOM.
Figure 2 Configuration of Hierarchical SOM
3. BACK-PROPAGATION NETWORKS
3.1. Basic back-propagation neural network
Back-propagation neural networks are directly
applied to image compression [9, 24]. The neural
network structure is shown in Figure 3. Three
layers, one input layer, one output layer and one
hidden layer are designed. The input layer and
output layer are fully connected to the hidden
layer. Compression is achieved by designing the
value of K, the number of neurons at the hidden
layer less than that of neurons at both input and the
output layers. The input image is split up into
blocks or vectors of 8x8, 4x4 or 16x16 pixels.
When the input vector is referred to as Ndimensional
which is equal to the number of pixels
included in each block, all the coupling weights
connected to each neuron at the hidden layer can
be represented by {wji , j=1, 2,…, K and i=1, 2,…,
N}, which can also be described by a matrix of
order KxN. From the hidden layer to the output
layer, the connections can be represented by {w’ij :
1≤ i ≤ N, 1 ≤ j≤ K} which is another weight matrix
of order NxK. Image compression is achieved by
training the network in such a way that the
coupling weights, {wji }, scale the input vector of
N-dimension into a narrow channel of Kdimension
(K<N) at the hidden layer and produce
optimum output value which makes the quadratic
error between input and output minimum.
Figure 3 back propagation Neural Network
3.2. Hierarchical back-propagation neural network
The basic back-propagation network is further
extended to construct a hierarchical neural network
International Journal of Reviews in Computing
© 2009-2010 IJRIC& LLS. All rights reserved. IJRIC
ISSN: 2076-3328 www.ijric.org E-ISSN: 2076-3336
77
by adding two more hidden layers into the existing
network as proposed in [24]. The Hierarchical
neural network structure is illustrated in Fig. 4 in
which the three hidden layers are termed as the
combiner layer, the compressor layer and the decombiner
layer. The idea is to exploit correlation
between pixels by inner hidden layer and to exploit
correlation between blocks of pixels by outer
hidden layers
Figure 4. Hierarchical Neural Network Structure
From the input layer to the combiner layer and
from the de-combiner layer to the output layer,
local connections are designed which have the
same effect as M fully connected neural subnetworks.
As seen in Fig. 4, all three hidden layers
are fully connected. The basic idea is to divide an
input image into M disjoint sub-scenes and each
sub-scene is further partitioned into T pixel blocks
of size pxp. For a standard image of 512x512, as
proposed [24], it can be divided into 8 sub-scenes
and each sub-scene has 512 pixel blocks of size
8x8. Accordingly, the proposed neural network
structure is designed to have the following
parameters:
The total number of neurons at the input layer is
Mxp2 = 8x64 =512. Total number of neurons at the
combiner layer is MxNh =8x8=64. Total number of
neurons at the compressor layer is Q = 8. The
total number of neurons for the de-combiner layer
and the output layer is the same as that of the
combiner layer and the input layer, respectively. A
so-called nested training algorithm (NTA) is
proposed to reduce the overall neural network
training time, which comprises the following
steps:
Step1: Outer loop neural network (OLNN)
training.
Step2: Inner loop neural network (ILNN) training.
Step3: Reconstruction of the overall neural
networks.
After training is completed, the neural network is
ready for image compression in which half of the
network acts as an encoder and the other half as a
decoder. The neuron weights are maintained the
same throughout the compression process.
3.3. Adaptive back-propagation neural network
Further to the basic narrow channel backpropagation
image compression neural network, a
number of adaptive schemes are proposed [4, 9]
based on the principle that different neural
networks are used to compress image blocks with
different complexity. The general structure for the
adaptive schemes are shown in Figure-4 in which a
group of neural networks with increasing number
of hidden neurons, (hmin, hmax), is designed. The
basic idea is to classify the input image blocks into
a few sub-sets with different features according to
their complexity measurement. A fine tuned neural
network then compresses each sub-set. Four
schemes are proposed [9] to train the neural
networks which are classified as parallel training,
serial training, activity-based training and activity
and direction based training schemes.
3.4. Performance Considerations
Considering the different settings for the
experiments reported in various sources
[9],[23],[24], it is difficult to make a comparison
among all the algorithms presented in this section.
Fig. 5 Adaptive Neural Network Structure
International Journal of Reviews in Computing
© 2009-2010 IJRIC& LLS. All rights reserved. IJRIC
ISSN: 2076-3328 www.ijric.org E-ISSN: 2076-3336
78
To make the best use of all the experimental
results available, we take Lena as the standard
image sample and summarize the related
experiments for all the algorithms as illustrated in
Table1 which are grouped into basic back
propagation, hierarchical back propagation and
adaptive back propagation.
Table 1 Basic Back Propagation on Lena (256x256)
Table 2 Hierarchical Back Propagation on Lena
(512x512)
Table 3 Adaptive Back Propagation on Lena (256x256)
4. MODULAR NEURAL NETWORKS
Though a single neural network can compress the
average characteristic for the image data, it is
difficult to compress both the edge and flat regions
with the same precision. For this, Rahim et al.
have proposed the compression method using two
neural networks [26]. One of the neural networks
is used for the compression of the original image
and the other is used for the compression of the
residual image. However, it is effective to
compress for each region, which is divided in to
the edge and flat regions. A modular structured
neural network consisting of multiple neural
networks with different block sizes (the number of
input units) for region segmentation has been
proposed by Watanbe and Mori [32]. By the
region segmentation, each neural network is
assigned to each region such as the edge or the flat
region.
From simulation results, it is shown that the
proposed method yields a better compression
compared with the conventional compression
technique using a single neural network.
5. NEURAL NETWORK DEVELOPMENT OF EXISTING TECHNOLOGY
In this section, we show that the existing
conventional image compression technology can
be developed right into various learning algorithms
to build up neural networks for image
compression. This will be a significant
development in the sense that various existing
image compression algorithms can actually be
implemented by one neural network architecture
empowered with different learning algorithms.
Hence, the powerful parallel computing and
learning capability with neural networks can be
fully exploited to build up a universal test bed
where various compression algorithms can be
evaluated and assessed. Three conventional
techniques are covered in this section, which
include wavelet transforms, fractals and predictive
coding.