EDGE PRESERVING IMAGE COMPRESSION FOR

MAGNETIC RESONANCE IMAGES

1
 CONTENTS

1. INTRODUCTION

1.1 OVERVIEW 1
1.2 STEPS INVOLVED IN COMPRESSING AN IMAGE 3
1.3 CHARACTERISTICS OF AN APPROPRIATE 4
IMAGE COMPRESSION
1.4 MAGNETIC RESONANCE IMAGING (MRI) 5
1.5 NEED OF IMAGE COMPRESSION : MRI 6
1.6 GOALS FOR DEVELOPING AN IMAGE 7
COMPRESSION ARCHITECTURE : MRI

2. BASIC CONCEPT S

2.1 SOME BASIC FACTORS OF IMAGE COMPRESSION 8

2.2 DATA COMPRESSION USING NEURAL NETWORKS 10

3. EDGE PRESERVING IMAGE COMPRESSOR

ARCHITECTURE

3.1 EDGE PRESERVING IMAGE COMPRESSOR

12
3.2 GOALS OF COMPRESSION ARCHITECTURE
13
3.3 USE OF BANK OF NEURAL NETWORKS
14
3.4 COMPRESSOR AND DECOMPRESSOR OVERVIEW
15

4. EXTENSION OF DANN ARCHITECTURE & TRAINING

ALGORITHMS FOR EPIC 17

5. EPIC COMPRESSION ALGORITHM 21

6. ADVANTAGES OF THE EPIC COMPRESSION 24

7. FUTURE ENHANCEMENTS AND CONCLUSION 26

8. BIBLIOGRAPHY
27
2
 Chapter-One
INTRODUCTION

1.1 OVERVIEW

Image compression is a field that is growing exponentially, due to its

widespread applications in digital imaging and transmission. With the
evolution of digital storage of visual information from grayscale still images to
full-color real-time video images, ever increasing bandwidths will be needed if
very high compression ratio real-time image compression techniques were not
available. Image compression serves to alleviate this problem by reducing the
required bandwidths and allowing new applications, such as
videoconferencing, to coexist with existing data and information transmission
on computer networks.
Nonetheless, existing image compression techniques are able to
achieve compression ratios from 2:1 to 20:1 for most applications. The higher
compression ratios have been achieved using “lossy” techniques that remove
visual information that is not perceived by the human eye. Higher compression
ratios can only be achieved at the expense of higher distortion that degrades
the quality of the image. To meet the challenges of achieving even high
compression ratios, adaptive compression schemes have been developed. They
are capable of adapting to the characteristics of the inputs as they vary. In this
way, the image quality is preserved even though compression ratios are much
higher diagnoses.
Edge information and image textures are important components
that must be preserved by the image compression system. However, Shannon’s

3
entropy relation indicates that the achievable compression ratio for general
data is about 2:1 using lossless techniques.
Our visual and auditory senses perform some “masking” of the
received inputs, rendering a portion of the received input redundant and
unperceivable. This enables us to devise lossy data compression techniques
which remove this redundant information, thus achieving much higher
compression ratios.
JPEG still image compression scheme, which provides
compression ratios of 5:1 to 20:1 with reasonable distortions.
For JPEG and other transform-based compression techniques, the
cost-function is the energy or frequency content of the image , which is
concentrated in the lower values. The use of transform-based coding
techniques effectively results in the blurring of edges and other high-frequency
contents of the image . Furthermore, at high compression ratios (which is
controlled by the Q-Factor in JPEG), the distortion rate increases significantly
as well, resulting in “patchy” or “blocky” images. Real-time implementations
of the Discrete Cosine Transform (DCT) used in JPEG are also costly, since
DCT is computationally
Medical image compression, such as the compression of magnetic
resonance images, requires the detection and recognition of boundaries in the
tissue structures, for detection of lesions and tumors. This requires the
preservation of edges in the image, which defines the various boundaries
between the various anatomical structures and tissues in the image.

4
Neural networks are adaptive in nature, due to their use of a training phase for
learning the relationships between the input data and required outputs, and
their generalization capabilities provide a means of coping with noisy or
incomplete data.

1.2 STEPS INVOLVED IN COMPRESSING AN IMAGE:

1) Specifying the rate (bits available) and Distortion (tolerable error)

parameters for the target image.

2) Dividing the image data into various classes, based on their importance.

3) Dividing the available bit budget among these classes, such that the
distortion is minimum.

4) Quantize each class separately using the bit allocation information derived
in step 3.

5) Encode each class separately using an entropy coder and write to the file.

5
1.3 THE CHARACTERISTICS OF AN APPROPRIATE
IMAGE COMPRESSION SCHEME CAN BE DEFINED
AS FOLLOWS:

1) The ability to specify which components of the image are vital for visual
integrity and contain the most information, and therefore, need to be
preserved with very little or no loss.

2) The ability to achieve as high compression ratios as possible for the other
portions of the image, leading to significant savings in transmission and
storage requirements.

3) The need to control the distortion incurred by the high compression

approach within user specified levels.

4) The ability to adapt to changes in the input data stream.

5) The ability to perform the compression and decompression as fast as

possible, for use in real-time applications.

6
1.4 MAGNETIC RESONANCE IMAGING (MRI)

Magnetic resonance imaging (MRI) is a test that uses a magnetic field and
pulses of radio wave energy to make pictures of organs and structures inside
the body.
In many cases MRI gives information that cannot be seen on an X-
ray, ultrasound, or computed tomography (CT) scan. For an MRI test, the
area of the body being studied is placed inside a special machine that is a
strong magnet. MRI scanner detects changes in the normal structure and
characteristics of your head. The MRI can also detect tissue damage or
disease, such as infection, inflammation, or a tumor.
Information from an MRI can be saved and stored on a computer for
more study. Photographs or films of certain views can also be made.
In some cases, a dye (contrast material) may be used during the MRI to
show pictures of organs or structures more clearly.
MRI can be used to find problems such as bleeding, tumors, infection,
blockage, or injury in the brain, organs and glands, blood vessels, and joints.
In some cases, a contrast material may be used during the MRI scan to
enhance the images of certain structures. The contrast material may help
evaluate blood flow, detect some types of tumors, and locate areas of
inflammation.

While MRI is a safe and valuable test for looking at structures and organs
inside the body and it is more expensive than other image techniques

7
1.5 NEED OF IMAGE COMPRESSION WITH RESPECT TO
MAGNETIC RESONANCE IMAGING (MRI)

Preservation of edges in the image after compression is atmost necessary as

Boundaries and edges in the tissue structures are vital for detection of
lesions and tumors.
Medical image compression, such as the compression of magnetic
resonance images, requires the detection and recognition of boundaries in
the tissue structures, for detection of lesions and tumors. This requires the
preservation of edges in the image, which defines the various boundaries
between the various anatomical structures and tissues in the image. In
addition, the permissible distortion rate is zero to very low values for
medical image compression due to its nature, resulting in a conflict between
requirements
(i) The ability to achieve as high compression ratios as possible for
the other portions of the image, leading to significant savings in
transmission and storage requirements.
(ii) The need to control the distortion incurred by the high compression
approach in (i) to within user specified levels.
Since distortion rate and achievable compression ratios are directly related
(a high compression ratio implies a high distortion rate).
The edge information extraction is useful for other reasons as well.
Correlation of anatomical features between images obtained using different
modalities are vital for various diagnostic functions involving tissue
identification,
classification, and metabolic modeling. For instance, CT and MRI image
features are used to localize anatomical features obtained from PET and
other modalities with less distinct image features and boundaries.
8
1.6 HOW TO ACHIEVE THE GOALS FOR DEVELOPING
AN IMAGE COMPRESSION ARCHITECTURE FOR
MRI:

1) The preservation of edge information in the recovered image with no

distortion.

2) The use of an image classifier to separate image blocks into different

classes and to compress the different classes of images using DANN-
based networks that are optimally trained for each class.

3) The improvement in the training set generation criteria, to eliminate

redundant patterns and speed training.

4) The extension of the DANN-based neural network training approach to

improve the speed of convergence during training.

5) The combination of the different areas into a compression system.

9
 Chapter-Two

SURVEY OF PREVIOUS WORK

AND BASIC CONCEPTS

2.1 SOME BASIC FACTORS OF IMAGE COMPRESSION

2.1.1 BASIC CATEGORIES OF IMAGE COMPRESSION

Image data compression and decompression can be classified into two

basic categories, lossless compression and lossy compression.Data
compression is alternatively termed source coding for the case of lossless
compression, since it entails finding the most compact representation of the
data source. In general, lossless techniques are first generation techniques
which utilize algorithmic approaches to compressing data. Lossy techniques
usually adapt to the characteristics of the human visual system for determining
what information is visually important. The predominant lossy compression
technique in use today is based on the Discrete Cosine Transform (DCT),
which is used in the JPEG compression technique.

2.1.2 DCT TECHNIQUE

Fourier Quantization is a transform technique which has been investigated

for digitizing hand images. In contrast, DCT is a fast but sub optimal
transform which operates on a small section of the image at a time. This
invariably results in “blocking” effects where sub images show sharp

10
transition artifacts or block-like appearance at the edges of each sub image. In
addition, transform techniques involve a lot of computationally intensive steps
during compression and decompression, making software implementations
cumbersome for real-time applications. However, dedicated processors have
started to appear for performing the transform manipulations quickly.
Variations of the DCT technique attempt to improve its
performance and reduce the blocking artifacts. One approach is to perform the
DCT on the whole image as a single frame . This alleviates the blocking effect
at the expense of compressor complexity and memory requirements.
Acceptable compression ratios of 4:1 for MR images to 20:1 for CR chest
images were recorded. Adaptive DCT (ADCT) has been investigated, which
uses a variable encoding scheme for the transformed image, allocating more
bits to coefficients with higher values . This improves the compression ratio by
25- 30%.

2.1.3 VECTOR QUANTIZATION

Another approach utilizes Vector Quantization techniques for the lossy

compression of MRI images . Techniques based on Vector Quantization have
asymmetrical processing requirements. Decoding is greatly simplified due to
the use of a Vector Codebook which is used to reconstruct the image.
However, the computational overhead involved in selecting optimal vectors
for encoding the images is quite high, as well as the choice of the codebook
itself. Variations of the technique can be used to provide for image error
correction and distortion control capabilities using variable-sized image
blocks as well. Sub band coding with DPCM quantization and post
quantization compression [17] has been utilized for image compression of
non-medical images. It utilizes the characteristics of the human visual

11
perception system to filter out portions of the image spectrum to which the
eye is insensitive. The technique has achieved very good compression ratios,
with the JPEG standard test image achieving up to 0.23 bits/pixel.

2.2. DATA COMPRESSION USING NEURAL NETWORKS

2.2.1. NEURAL NETWORK IMPLEMENTATION OF VECTOR

QUANTIZATION

A recent work using neural network approaches to image compression is based

on Vector Quantization. Since the selection of the appropriate vectors for
representing a given image is basically a classification problem, neural
networks which have been trained to do so optimally will greatly increase the
compression speed. Nevertheless, reported results still indicate a rather high
distortion rate, with a signal to noise ratio (SNR) of 26.9 for a 20:1
compression ratio. Further improvements in the error rates are necessary
before this technique can be applied to solve medical image compression
problems.

2.2.2. Cottrell/Munro/Zipser Technique

Most of the neural network based techniques are derived from the work by
Cottrell, Munro and Zipser . They developed a multi layered perceptron neural
network with back propagation as the error learning function. Compression is
achieved by the means of multi-layered neural networks that have small
12
hidden layers (in comparison to the output layer). Multi-Level Neural Nets are
trained using back-propagation with the block sampled original image at the
input and output in order to obtain a reduced vector representation of each
block in the hidden layer. The internal (hidden) representation of the block is
the compressed form of the image to be transmitted since there are less hidden
layer units than at the input or output. By applying the internal vector to the
decompressing network, the original
image can be recalled.
This approach uses a form of self-organization for obtaining the
reduced dimension feature space. Neural networks are trained using a set of
sample images, and then the weights obtained through the training phase are
used to compress actual images.
New neural network based adaptive compression methods which
overcome the traditional drawbacks associated with back-propagation based
neural networks, such as the static nature of network architectures, unbalanced
errors of individual training patterns. This new neural network architecture,
termed Dynamic Auto associative Neural Networks (DANN) has been shown
to provide excellent control over the error rates while maintaining user-
selectable compression ratios.

13
 Chapter-Three

EDGE PRESERVING IMAGE COMPRESSOR

(EPIC) ARCHITECTURE

3.1 EDGE PRESERVING IMAGE COMPRESSOR(EPIC)

The proposed Edge Preserving Image Compressor (EPIC) architecture is a new

hybrid image compression technique combining conventional techniques with
DANN-based neural networks. EPIC utilizes techniques that are similar to the
approach of the edge/non-edge compression algorithm, which in turn is
derived from the synthetic-highs compression technique.
Neural network based compression is designed to be modular in
nature, and improvements in each module can be incorporated easily into the
architecture, resulting in better accuracy and performance of the system. One
of the major criteria of this architecture is the preservation of edge information
of the original image. Since edge detection is a vital step in any image
processing and manipulation task, the inclusion of this step into the image
compression architecture results in two distinct advantages: the elimination of
the need for edge detection in subsequent downstream processing of the
images, and the guarantee that edge information is not lost
during compression, unlike many transform-based methods.
Since edge information has been preserved, we can compress the
rest of the image at high compression ratios with a user-controlled distortion
rate without significantly affecting the overall picture quality of the
reconstructed image. Nonetheless, for certain applications such as medical

14
imaging, very minimal or no loss compression is required where image
integrity is of utmost importance.

3.2 GOALS OF COMPRESSION ARCHITECTURE:

Neural network based techniques, which have shown significant advantages

over other techniques in their adaptation and generalization capabilities, have
been selected for this purpose. The neural networks are trained to compress
image features present in specific types of images. By targeting a specific data
compression domain for the neural network compressor, the architecture aims
to achieve both objectives through the effective use of adaptation and
generalization capabilities inherent in neural networks.
The second goal of the compression architecture is to create a
system that is easily converted to parallel implementations. Neural network
systems are well suited to such implementations due to their uniform structure
and the distributed nature of the neural computation, output, and weight
adaptation processes. Hardware implementation of neural network topologies
will lead to the development of compression systems that can perform their
tasks in real-time. The edge detection step enables us to select the most
appropriate edge detection technique for the given images, while
improvements in edge detection algorithms and changes in image integrity
criteria can be “plugged into” the compression system without affecting the
rest of the compression process.

15
3.3 USE OF BANK OF NEURAL NETWORKS

Another means of parallelizing the image compression process is through the

use of a bank of neural networks for the compression process. The image is
subdivided into uniform sized blocks that are fed to a bank of neural networks.
Each network has been previously trained for a specific class of image data
based on the variance value of the image block. The appropriate network is
selected for the given image block, and compression of each block proceeds
independently of the other blocks. It is therefore possible to pipeline and
parallelize this process for the compression of blocks belonging to different
classes, achieving significant improvements in processing speed.
The advantage of using a bank of neural networks tailored to
specific image classes is the capability for parallelism during the training
process. The entire training set is divided over the bank of networks, and the
number of patterns that each network is required to learn is therefore reduced,
or alternatively, the system is capable of learning more images, improving its
adaptation capabilities.

3.4 COMPRESSOR AND DECOMPRESSOR OVERVIEW

Among the various edge detectors that have been developed, the Canny filter
have proven to be very accurate in terms of edge localization due to its
adjustable window size and edge thresholds. The Canny filter produces very
clean edges without the need for noise removal and is used by EPIC to perform
the edge extraction step during the compression phase. A classifier network
examines each block of the image and determines which of bank of neural nets
previously trained is best suited for the compression task using a simple
variance classifier. The ability of the neural network to generalize and be
trained to adapt to specific types of patterns in the input image is expected to

16
yield better compression ratios and image fidelity as compared to compression
using conventional DCT techniques, and is inherently more parallelizable in
hardware implementations due to its uniform structure and simple processing
requirements.

The neural network architecture utilized in EPIC is a multi-

layered Perceptron network with two hidden layers, which has been used
successfully for image data compression . The input vector is compressed into
a neural network pattern representation using a compression ratio dictated by
the ratio between the number of input neurons (equal to the number of pixels
in the image block) and the number of neurons in the second hidden layer. The
compressed pattern outputs from the neural network compressor is then coded
using either a variable bit-rate linear predictive coder, or quantized using a
fixed bit-rate with marginally higher distortion. Optional Lempel-Ziv-Welch
(LZW) encoding is applied to the outputs of both the edge coding and neural
network compressors to remove any residual redundancies. These patterns are
then stored or transmitted for use in reconstructing the images at the
decompressor.
The decompressor performs the corresponding optional LZW
decoding on the received edges and compressed patterns, as well as
corresponding linear predictive decoding on the compressed patterns (if
required), and passes it through a bank of neural networks which converts the
compressed patterns into recovered output blocks that are used to form the
non-edge image. This non-edge image is then overlaid with the decoded edge
pixels to reconstruct the reconstructed image.

The compression system can be shown in the following diagrams

(Figures 3.1 and 3.2):

17
Figure 3.1:

Figure 3.2:

18
 Chapter-Four

EXTENSION OF DANN ARCHITECTURE &

TRAINING ALGORITHMS FOR EPIC

4.1 USE OF DANN-NETWORK

Dynamic Associative Neural Networks (DANN) have been used effectively

for image compression . Consequently EPIC utilizes DANN-based neural
networks for compressing the non-edge image. However, DANN training
suffers from the problem of slow convergence, since the descending epsilon
training technique is a very slow process. There are often specific input
patterns that are considered very difficult to learn. The single network used for
DANN compression to compress all possible input patterns therefore has to
account for those difficult patterns in its learning process as well. This further
increases the required network capacity and complicates the learning process
since the network often has to adapt to input patterns with very different
characteristics. All these factors limit the ability to extend DANN-based
compression techniques to provide very high data compression ratios while
providing very low error rates for use with application where data fidelity is
extremely important.

19
4.2 EPIC IMPROVES UPON DANN-BASED COMPRESSION
THROUGH THE FOLLOWING FEATURES:

1) The use of a bank of DANN networks in place of a single network for

processing the incoming data stream, under the control of a variance
classifier.

2) Improvements in the training set generation procedure through the

elimination of duplicate training vectors which are close together in
Euclidean space, controlled by the use of a similarity threshold.

3) Improvements in the training process, by specifying termination criteria for

pathological training conditions, where the network error is increasing
continually or is stuck for a given number of training epochs. The changes
to DANN architectures and training algorithms are detailed in the
following sections.

4.3 VARIANCE CLASSIFIER SELECTOR FOR

CONTROLLING A BANK OF DANN-NETWORKS

A bank of four-layered feed forward perceptron networks (similar to those

used in DANN compression) is used to process the incoming data stream. The
data is classified using a simple variance classifier into 1 of P classes, where P
is determined heuristically from variance characteristics of the data, and
compressed by the corresponding network. In this way, the cost of
compressing complex patterns can be isolated from the cost of compressing
the other patterns. The network used to compress a pattern with low variance
will require a small first hidden layer and provide very high compression
ratios, in contrast to other networks which have larger first hidden layers, and

20
possibly lower compression ratios. The outline of the training process is given
in Figure 4.1.
Figure 4.1

4.4. THRESHOLDED TRAINING SET GENERATION

PROCEDURE

The generation of the training set for training the bank of P neural networks is
to take a set of training images, and partition each image into N x N sized
blocks, overlapping by N/2 pixels, for use in generating the training set. The
variance of the block is calculated for use in the classification step. The blocks
are classified into one of P classes, based on predefined variance threshold
determined empirically through statistical analysis of the characteristics of
typical images, and converted into a training vector. The candidate vector
belonging to the p-th class is then compared with all existing training vectors
in the training set for that class to determine if it exceeds a “similarity
threshold” defined to eliminate redundant training vectors lying close together
in Euclidean space. The user controllable “similarity threshold” serves to
reduce the size of the training set and avoids over training through the use of
highly similar training vectors, which reduces the generalization properties of
the networks.

21
4.5. Modified DANN Training Procedure

The neural network training algorithm used in EPIC is based on the DANN
training procedure . Each neural network in the bank of P networks is
subjected to DANN training, with some modifications designed to detect
pathological conditions and reduce time spent in unproductive training. Among
the improvements are convergence detection primitives, which measure the
slope of the error curve to determine if the training process is stuck (wobbling)
or if the error function is increasing as training proceeds (increasing error).
Thresholds are set for the number of iterations the network can exist in these
conditions before that particular phase of training is aborted and control
returned to the calling routine. The EPIC training process for each of the P
networks can be done in parallel, decreasing the time required for learning a
given set of images compared to the original DANN training process.

22
 Chapter-Five

EPIC COMPRESSION ALGORITHM

5.1 ALGORITHMS FOR EPIC COMPRESSION:

The EPIC compression technique is essentially an asymmetric compression

scheme. The processing overhead is incurred at the compression phase, due to
the edge detection and block classification steps. The decompression phase
only requires the reconstruction of the non-edge image and the overlaying of
the edge pixels onto that image to result in the reconstructed image.

An incoming image of size n x n is passed through an edge

detector for extracting the edge pixels for lossless chain coding while
simultaneously being subdivided into N x N non-overlapping blocks for
processing by the neural network based lossy compressor. The block is
classified into 1 of P classes. This process of decomposing the image into
individual blocks for use in classification can be performed in parallel. The
classified blocks are then fed to the bank of neural networks for compression.
Each network has been trained previously using image blocks that have the
same variance characteristics and is thus well adapted for those blocks. The
second hidden layer compressed patterns are then extracted, and either coded
using a zeroth order Linear Predictive Coder at a variable bit rate, or
alternatively, quantized at a fixed bit-rate with marginally higher distortion.
Associated with each compressed pattern is the network class for use in
decompression.

23
The chain coded edge image, as well as the compressed patterns, are optionally
further compressed using the LZW-based UNIX compress command to
remove any further redundancies in the data stream. This usually results in no
more than 10% improvement in the compression ratio, however, and is
obtained only from the chain codes. It is conceivable that this step could be
eliminated with the use of a Huffman coder for the chain codes.
Decompression is relatively straightforward. The compressed non edge pattern
and edge image are first decompressed using the UNIX uncompress command
as necessary, and the compressed patterns extracted and fed to the
decompressors based on their network class for decompression into the
respective image blocks. The block decompression step can be performed in
parallel. The various blocks are used to form the non-edge image. In addition,
the chain coded edge image is decoded and overlaid onto the non-edge image
to give the reconstructed image.
The algorithms for compression and decompression
are given in Figures 5.1 and 5.2
For given set of images
Perform Edge Detection on the image
Store Edges and their Gray values using chain coding
Partition image into N x N non-overlapping blocks
For each block (parallelizable)
classify the block into p P classes using predefined variance thresholds
compress the block using the pth neural network
Quantize output of second hidden layer
endfor
Store compressed patterns using differential linear predictive coding,
or fixed-bit rate quantization for higher compression ratio
Optional edge compression stage using LZW encoding for storage/transmission
endfor
Fig. 5.1: EPIC Image Compression Pseudocode

24
For given set of images
LZW decoding (correspond to optional encoding) of edge compressed images
Decode chain of Edges and their Gray values
Perform differential linear predictive decoding of compressed patterns (as needed)
For each block (parallelizable)
decompress block using the p-th neural network specified in pattern’s class info.
Form N x N non-edge image block
endfor
Overlay Edge pixels on non-edge image to obtain reconstructed image
endfor
Fig 5.2: EPIC Image Decompression Pseudocode

25
 Chapter-Six

ADVANTAGES OF THE EPIC COMPRESSION

1) Edge preservation: Since the edges in the image are coded using a
lossless scheme, it does not suffer from the drawback of conventional
JPEG, which tends to lose high frequency components (edges) during
compression.

2) Feature determination: A major advantage of the edge extraction step is

that it provides the first step towards the extraction of features from the
image for use in further image processing. This utilization of an
essential step in image processing in the compression process helps
eliminate redundant processing later on.

3) High compression ratio for non-edge images: the separation of edge

information from the rest of the image affords the use of lossy
compression schemes that achieve high compression ratios. In addition,
the use of neural networks for performing this step enables us to tailor
the compression architecture for specific types of images.

4) Progressive Transmission: Since neural network compressors are able to

generalize and compensate for noisy inputs, it is therefore possible to
provide progressive transmission of images by transmitting the most
significant bits of the compressed vectors first. This is effectively
similar to performing quantization on the transmitted vectors, where the
26
 received compressed vector is improved successively using more
significant bits, increasing the number of quantization levels and
converging to the final image.

5) Temporal adaptability: Neural network based compressors are by

definition adaptive. In addition, they can be configured to adapt to
changes over time in the input data stream by the use of shadow neural
network compressors which continue to learn from patterns extracted
from the input stream. If the error rates of the on-line network exceed a
given threshold, the shadow compressors can replace the existing
network easily, thus achieving temporal adaptability.

27
 Chapter-Seven

FUTURE ENHANCEMENTS AND CONCLUSION

Further improvements to the DANN training algorithms are still needed, to

improve its speed of convergence and convergence ability beyond that
achieved through the techniques proposed in PIC. This is necessary due to the
non-linear nature of the neural network training process, as well as the high
degrees of freedom present in parameter selection. Edge removal prior to
neural network compression would also help reduce the variability in the
image and help convergence.
In order for the compression architecture to perform optimally,
continuous monitoring of the compression process should be maintained. This
is accomplished by using a similar bank of neural compressors in parallel with
the actual compressor (called the shadow compressor), and if the error exceeds
a predefined threshold, the shadow compressor network will be retrained in
order to return the system to its operational goals. Once the training is
completed, the updated weights of the neural networks will be transferred to
the operational compressor network. Variable sized image blocks can also be
used to reduce “blocking” effects, while large image blocks can be used to
provide even higher compression ratios for the uniform dark regions
surrounding the MR images.

28
 BIBLIOGRAPHY

1. Lic. Lanzarini Laura1, ‘Images Compression for Medical

Diagnosis using Neural Networks’, 2004

2. Tat Chee Wan, ‘Edge Preserving Image Compression

for Magnetic Resonance Images Using DANN-Based Neural
Networks’, 2003

3. Constantio Carlos, ‘Edge Preserving Image Compression

for Magnetic Resonance Images Using DANN-Based Neural
Networks’, 2004

4. Netravali, A., Limb, J., “Picture Coding: A Review,”

Proceedings of the IEEE, March 1990, Vol. 68, No. 3

5. Jain, A., “Image Data Compression: A Review,”

Proceedings of the IEEE, March 1991, Vol. 69, No. 3

6. Riskin, E. Lookabaugh, T., et. al., “Variable Rate

Vector Quantization for Medical Image
Compression,” IEEE Transactions on Medical
Imaging, Sept. 2001, Vol. 9, Issue No. 3

7. Garrido, D.Nunes, R., “An Edge/Non-Edge Image

Compression Algorithm,” Proceedings of the IEEE
International Conference On Image Processing, Singapore,
Sept. 2005.

8. Hong yan, “ Image Analysis for digital media application ”

IEEE computer Graphics Application, Jan 2005,
Vol.-21 Issue no – 1

29