92 97

ISSN: 2278 – 1323
International Journal of Advanced Research in Computer Engineering & Technology
Volume 1, Issue 4, June 2012

VLSI Architecture and implementation for 3D Neural Network based image
compression

Deepa.S1, Dr.P.Cyril Prasanna Raj 2, Dr.M.Z.Kurian3, Manjula.Y4
1-Final year M.Tech VLSI & embedded systems deepa2292959@gmail.com,
2-Professor, member of IEEE cyrilyahoo@gmail.com,
3- Dean,HOD,E&C,SSIT,Tumkur mzkurian@gmail.com,
4-Lecturer,Dept of E&C,SSIT,Tumkur manjulayarava@gmail.com

Abstract: Image compression is one of the key option for reducing the number of bits in
image processing techniques in signal processing transmission. This in turn helps increase the volume
and communication systems. Compression of of data transferred in a space of time, along with
images leads to reduction of storage space and reducing the cost required. It has become
reduces transmission bandwidth and hence also increasingly important to most computer networks, as
the cost. Advances in VLSI technology are rapidly the volume of data traffic has begun to exceed their
changing the technological needs of common man. capacity for transmission. Traditional techniques that
One of the major technological domains that are have already been identified for data compression
directly related to mankind is image compression. include: Predictive Coding, Transform coding and
Neural networks can be used for image Vector Quantization. In brief, predictive coding
compression. Neural network architectures have refers to the decorrelation of similar neighboring
proven to be more reliable, robust, and pixels within an image to remove redundancy.
programmable and offer better performance Following the removal of redundant data, a more
when compared with classical techniques. compressed image or signal may be transmitted.
In this work the main focus is on development of Transform-based compression techniques have also
new architectures for hardware implementation of been commonly employed. These techniques execute
3-D neural network based image compression transformations on images to produce a set of
optimizing area, power and speed as specific to coefficients. A subset of coefficients is chosen that
ASIC implementation, and comparison with allows good data representation (minimum distortion)
FPGA. while maintaining an adequate amount of
compression for transmission. The results achieved
Key words: Image compression, 3-D neural network, with a transform based technique is highly dependent
FPGA, ASIC on the choice of transformation used (cosine,
wavelet, etc.). Finally vector quantization techniques
require the development of an appropriate codebook
I. INTRODUCTION to compress data. Usages of codebooks do not
guarantee convergence and hence do not necessarily
Neural network for image compression and deliver infallible decoding accuracy. Also the process
decompression have been adopted as they achieve may be very slow for large codebooks as the process
better compression and also work in noisy requires extensive searches through the entire
environment. Many approaches have been reported in codebook.
realizing the Neural Network architectures on
software and hardware for real-time applications. Artificial Neural Networks (ANNs) have been
Today’s technological growth, has led to scaling of applied to many problems and have demonstrated
transistors and hence complex and massively parallel their superiority over traditional methods when
architecture are possible to realize on dedicated dealing with noisy or incomplete data. One such
hardware consuming low power and less area. This application is for image compression. Neural
work reviews the neural network approaches for Networks seem to be well suited to this particular
image compression and proposes hardware function, as they have the ability to preprocess input
implementation schemes for Neural Network. The patterns to produce simpler patterns with fewer
transport of images across communication paths is an components. This compressed information (stored in
expensive process. Image compression provides an

92
All Rights Reserved © 2012 IJARCET

ISSN: 2278 – 1323

a hidden layer) preserves the full information
obtained from the external environment. Not only
can ANN based techniques provide sufficient
compression rates of the data in question, but security
is easily maintained. This occurs because the
compressed data that is sent along a communication
line is encoded and does not resemble its original
form.

II. SYSTEM DESIGN

Image
compre 3D
Image ssion enco
Image prepro
Image using ding
acquis cessing 3D archi
ition
Neural tectu
Networ re
k

Fig 1: Block diagram overview
Fig 2: Proposed 3DNN architecture
The fig 1 gives the overall view of the work. Here
image may be any picture or a photograph or a visual
data. This is captured or acquired by an optical lence
such as a camera. The image is “discretized” i.e.,
defined on a discrete grid and stored as two-
dimensional collection of light values (or gray
values).The image preprocessing does the function of
adjustment of pixels and removal of noise from the
captured image. The image is compressed using 3D
Neural Network architecture and encoded for image
transmission.

The 3D Neural Network architecture for image
compression is main topic of interest of this project
work. An attempt to design and develop 3-D adaptive
neural network architecture for image compression
and decompression that can work in real time, noisy
environment is made here. Fig 3: Basic architecture for image compression using NN

The proposed 3-DNN architecture is as shown in M-dimensional coefficient vector y is calculated as
figure 2. Typical way of compression with neural Y=Zx X
network is using hidden layer with lower dimension then the inverse transformation is given by the
then input or output layer. In this case, network input transpose of the forward transformation matrix
and output is an image. The input and hidden layers resulting in the reconstructed vector
perform the compression, it transforms the input X =ZT x Y
vector of dimension d(IN) to hidden layer space of Here 16 input neurons, 4 neurons each in 4 hidden
dimension d(HI). The dimension of hidden layer is layers and 16 output neurons are considered. purlin
lower than of input layer d(HI)<d(IN). Output layer function and purlin functions are used to train the
with the help of hidden layers performs input and output neurons of hidden layers. Back
decompression. It transforms the vector from hidden propagation technique is also used.
layer with lower dimension to output vector space
with higher dimension d(HI)<d(OUT). The
dimension of input and output layer is the same III. SYSTEM IMPLEMENTATION/
d(IN)=d(OUT). SIMULATION
Matlab program is written in the following steps:
The basic architecture for image compression using The input image is read by the instruction imread.
neural network is shown in figure 3 This will read image saved in the specified path. Next
a part of image is read by specifying particular row

93

ISSN: 2278 – 1323

and column values. This part of the image is
displayed. Rearrangement of image is the selected
application of the multilayer perceptron. Next is the
training procedure where specified weight and bias is
obtained for image compression & decompression.
The reshaping of a block matrix in to a matrix of
vector occurs. The design consists of matrix
multiplication of two matrices, one is the input image
samples(each column), and the second is the weight
matrix obtained after training. This multiplied output
is added with bias to obtain the compressed output
that gets transmitted or stored in compressed format.
On the decompression side, the compressed data in Fig 5
matrix form is multiplied with the output weight
matrix and added with output bias to get back the Here the image is read in figure. Next a part of
original image. The image quality of the image(64*64) is read for compression and the
decompressed image depends on the weight matrix decompressed output is shown.
and the compression ratios. The image data of size
Fig 6: Training the neuron in matlab tool
16x16 is taken in each column wise i.e column 1 of
size 16x1 then multiplied by the weight matrix of
4x16 and added with bias matrix of size 4x1 to get a
compressed output of 4x1. On the decompression
side 4x1 input matrix (compressed image) is
multiplied with the weight matrix of size 16x4 and
added with output bias matrix of size16x1to get
output decompressed matrix of size 16x1. This
procedure is repeated for 15 more times to
reproduces the original image. In order to achieve
better compression nonlinear functions are used both
at transmitter and receiver section. The Neural
Network has 4 hidden and 16 output layers. Purlin The above snapshot shows training the neuron in
and purlin functions are used for input and output
matlab tool.
hidden layers. Back propagation technique is used.
Finally rearrangement of the vectors in to block
matrix to display the image. Simulation Results in matlab

Image input The different compression algorithms can be
compared based on certain performance measures.
Read image Compression Ratio (CR) is the ratio of the number of
bits required to represent the data before compression
to the number of bits required after compression. Bit
Read a part of the image
rate is the average number of bits per sample or pixel
(bpp), in the case of image. The image quality can be
Display the image evaluated objectively and subjectively. A standard
objective measure of image quality is reconstruction
error given by equation 1.
Rearrangement of image
Error E = Original Image – Reconstructed image (1)
Training procedure (fixing the
compression ratios & training Two of the error metrics used to compare the various
functions) image compression techniques are the mean square
error (MSE) and the Peak Signal to Noise Ratio
(PSNR). MSE refers to the average value of the
Rearrange matrix to display
square of the error between the original signal and the
reconstruction as given by equation 2. The important
Output image parameter that indicates the quality of the
Fig 4: Flow chart of matlab program reconstruction is the peak signal-to-noise ratio

94

ISSN: 2278 – 1323

(PSNR). PSNR is defined as the ratio of square of the image. Combination of different compression ratios
peak value of the signal to the mean square error, (CR)with different transfer functions gives different
expressed in decibels. PSNR values for the same input image.

MSE = E / (SIZE OF IMAGE) (2) Fig 7
PSNR 1:16 CR
The MSE is the cumulative squared error between the
compressed and the original image, whereas PSNR is 20
a Σmeasure of the peak error. The mathematical 15
formulae for the computation of MSE & PSNR is : 10
M N 5
MSE = 1/ MN [ Σ Σ (Ixy – I 'xy)2] (3) 0
psnr
i=1j=1
PSNR = 20 * log (255 / sqrt(MSE)) (4) funct

where I(x,y) is the original image, I'(x,y) is the
approximated version (which is actually the
decompressed image) and M, N are the dimensions of PSNR 2:16 CR
the images, 255 is the peak signal value. A lower
value for MSE means lesser error, and as seen from 20
the inverse relation between the MSE and PSNR. 15
Higher values of PSNR produce better image 10
compression because it means that the ratio of Signal 5
to Noise is higher. Here, the 'signal' is the original 0 psnr
image, and the 'noise' is the error in reconstruction.
So, a compression scheme having a lower MSE (and funct
a high PSNR), can be recognized as a better one.

Compression results by adaptive BPNN Structure
for size=64*64,r=4,CR=4:16(75%)
4:16 CR
Sl Test Image Image Image PSNR 21
no. max MSE 20
Error 19
18
1 cameraman 108 262.01 23.94 17
2 board 166 958.40 17.7765 16
15
3 cell 38 26.462 33.9045 psnr
4 circuit 15 14.113 36.6344
5 Lena 61 100.00 28.1305
6 sun 44 115.84 27.2919
7 girl 58 42.806 31.8157
8 Blue hills 8 22.182 44.7420
9 Sunset 55 39.371 32.1790
10 Water lilies 72 31.379 33.164
11 Winter 36 47.626 31.352
12 Drawing 255 170.77 15.194
13 College 57 97.780 24.037
14 Garden 78 87.990 25.000
15 My photo 101 197.22 22.991
16 Holi 97 175.34 25.7
17 Bhagavadgeetha 67 65.63 29.96
18 Devinecouple 74 80.20 29.1
19 krishna 39 7.1 39.6
20 Goddess 65 45.048 31.5940
Table 1: COMPRESSION RESULTS
Combination of different transfer functions gives
different decompressed images for the same input

95

ISSN: 2278 – 1323

Fig9

For a better image, the max error must be as low as
possible, the mean square error(MSE) must also be
small and the peak signal to noise ratio must be high.
From the above observations the 'tansig','purelin' and Inputs(inputs with scaling) are storeed as serial in parallel out
registers
'purelin','purelin' functions give consistently high
PSNRs' for the respective compression ratios(CR). Fig10
Trainrp is a network training function used in this
work that updates weight and bias values according
to the resilient backpropagation algorithm
(Rprop).Trainlm may also be used which is also a
network training function that updates weight and
bias values according to Levenberg-Marquardt
optimization but consumes huge amount of memory.
In this work back propagation technique is used. The
above results are for 64*64 image. The NN with BP
algorithm is also trained with 128*128, 256*256 and
521*512 images. The network trained using
backpropagation algorithm is realized using
multipliers and adders for FPGA realization .The output of compression and decompression in HDL coding.
work is being carried out with the design of fast
multiplier which uses Wallace technique. A 9 bit Fig11
Wallace multiplier and a 18 bits carry save adder is
designed for this project work.

The inputs for 9 bit Wallace tree multiplier and 18
bits carry save adder is obtained by the MATLab
program.HDL coding is being carried out for FPGA
realization. A 3D Neural network algorithm is
planned for the design.

HDL/VHDL and FPGA simulation results
Fig8

Output obtained by the vertex5 chip scope design

IV.HARDWARE AND SOFTWARE
REQUIREMENTS
The Neural Network archicture will be synthesized
on FPGA to estimate the hardware complexity for
output of 9 bits Wallace multiplier and 18 bits carry save efficient ASIC implementation. The design is
adder mapped on VERTEX5 device from Xilinx 13.1

96

ISSN: 2278 – 1323

vertion. As the design is mapped on FPGA, it image. Further this is coded in HDL for FPGA
supports Reconfigurability. Reconfigurability can be realization and the HDL code is modified for ASIC
achieved by changing the input layer values for better Designing and implementation. Low power
compression and decompression. techniques are used to reduce power dissipation of
the complex architecture.
The HDL code for FPGA implementation is modified
for ASIC implementation. The general coding styles VII. REFERENCE
is adopted for building optimized RTL code for ASIC
implementation. [1]. K.VenkataRamanaiah and Cyril Prasanna Raj VLSI
To support higher order multiplication it is modeled Architecuture for Neural Network Based Image Compression, ©
using HDL, and there is a need to develop an RTL 2010 IEEE
model using efficient coding styles for ASIC
[2]. Hamdy Soliman Computer Science Department New Mexico
Synthesis using Synopsys Design Compiler Version
2007.03. The timing analysis will be carried out Tech Neural Net Simulation: SFSN Model For Image
using Synopsys Prime Time. Compression, Proceedings of the IEEE, 2001
[3]. Daniel Matolin, Jörg Schreiter, Stefan Getzlaff and René
In brief the hardware used are the FPGA platform, Schüffny,”An Analog VLSI Pulsed Neural Network
and ASIC libraries.
Implementation for Image”, Proceedings of the International
The software requirements are Xilinx for HDL
coding , Design compiler and Matlab. Conference on Parallel Computing in Electrical Engineering
(PARELEC’04) IEEE
[4]. Arbib, Michael A. (Ed.) (1995). The Handbook of Brain
V. ADVANTAGES AND LIMITATIONS Theory and Neural Networks.
[5]. Alspector, U.S. Patent 4,874,963 "Neuromorphic learning
Advantages:
Among the all present method, neural network is of networks". October 17, 1989.
special interest due to the success it has in many [6]. Ivan Vilovic,” An Experience in Image Compression
applications. UsingNeural Networks”,48th International Symposium ELMAR-
1. Learning ability, 2006, 07-09 June 2006, Zadar, Croatia
2. System identification,
[7]. Hadi Veisi, Mansour Jamzad,” A Complexity-Based Approach
3. Robustness against noise,
4. Capability of optimum estimation and being used in Image Compression using Neural Networks”, International
in parallel structures Journal of Signal Processing 5-2-2009
5.Various kinds of neural networks like multilayer [8]. Rafid Ahmed Khalil,” Hardware Implementation of
perceptron (MLP), Hopfield, learning vector Backpropagation Neural Networks on Field programmable Gate
quantization (LVQ), self-organizing map (SOM) and
Array (FPGA)”, Al-Rafidain Engineering Vol.16 No.3 Aug. 2008
principal component neural networks have been used
to compress images. The VLSI Architecture for 3D [9]. Rehna. V. J, Jeya Kumar. M. K,”Hybrid Approaches to Image
Neural Network based image compression consume Coding: A Review” (IJACSA) International Journal of Advanced
less power and space, hence suitable for low cost and Computer Science and Applications, Vol. 2, No. 7, 2011
reliable Hardware implementation, reconfigurable on [10]. Jan Pohl, Petr Polách, Václav Jirsík,“ALTERNATIVE WAY
FPGA and good time to market. The process of OF IMAGE COMPRESSION WITH NEURAL NETWORK”,
image compression is very fast as it uses neural
[11]. Hossein SAHOOLIZADEH and Amir Abolfazl SURATGAR
network rather than codebook.
Adaptive Image Compression Using Neural Networks 5th
Limitations: International Conference: Sciences Of Electronics, Technologies
This technique may have small amount of distortion of Information and Telecommunications March 22-26, 2009
in the decompressed image TUNISIA
[12]. Matlab Image Processing Toolbox User’s Guide version-2
VI. CONCLUSION [13]. Image processing study guide by Rafael C. Gonzalez,Richard
E. Woods
The neural network architecture for image [14]. Digital Image Processing By Gonzalez 2Nd Edition 2002,
compressions has been analyzed on FPGA and ASIC (Ebook) Prentice Hall.
platforms and need to implement on 3D neural
[15]. Neural Network Toolbox™ 6 User’s Guide by Howard
network. A Matlab program is written to train the
neurons for compression and decompression of Demuth , Mark Beale , Martin Hagan.

97

92 97

More Related Content

What's hot (15)

Viewers also liked (9)

Similar to 92 97 (20)

More from Editor IJARCET (20)

Recently uploaded (20)

92 97