See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/272863745

Hybrid Image Compression Method using ANN and DWT

Article in International Journal of Computer Applications · June 2014

DOI: 10.5120/16641-6608

CITATIONS READS

5 372

2 authors, including:

Yogita Sawant
DKTE Society's Textile and Engineering Institute Ichalkaranji
1 PUBLICATION 5 CITATIONS

SEE PROFILE

All content following this page was uploaded by Yogita Sawant on 15 April 2017.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Applications (0975 – 8887)
Volume 95– No. 11, June 2014

Hybrid Image Compression Method using

ANN and DWT
Yogita Sawant L. S. Admuthe
D.K.T.E’S Text.& Engg. Institute D.K.T.E’S Text.& Engg .Institute
Ichalkaranji. India 416115 Ichalkaranji, India 416115

ABSTRACT more details of the image. Wavelet Image Processing facilitates

Transmission of image and video requires particularly large computers to accumulate an image in many scales of
amount of bandwidth and storage space. Image compression resolutions, decomposing an image into various levels and types
technology is useful to overcome these problems. Various of details. This paper implements a hybrid approach for
method of compression like transform coding, predictive compression of gray scale images using self organizing map
coding, bit plane coding are available and are used efficiently. (SOM) neural network and discrete wavelet transform. The
Compression systems based on hybrid coding which combines compression achieved by SOM neural network is higher as
the advantages of different methods of image compression have compared to other networks. But it is further increased using
also being developed over the last few years. Hybrid coding of wavelet transform which will compress the code vectors of
images, deals with combining various approaches to enhance SOM. The performance is evaluated using various performance
the individual methods and achieve better quality reconstructed measures.
images with higher compression ratio. In this paper, a hybrid
approach to image compression is discussed. A compression 2. COMPRESSION BY NEURAL
method using neural-network and discrete wavelet transform is NETWORK
presented here. This scheme combines the high compression Artificial neural network models are specified by network
ratio of Self organizing map neural network with the good topology (architecture) and learning algorithms [1],[2]. Network
recreation property of discrete wavelet transform (DWT).The topology describes the way in which the neurons (basic
performance of proposed method is compared with the available processing unit) are interconnected and the way in which they
SOFM NN based compression technique considering standard receive input and output. Learning algorithms specify an initial
images. set of weights and indicate how to adapt them during learning in
order to improve network performance. A neural network has
Keywords: Artificial Neural Networks, Image compression, number of simple computational units, called neurons which are
Self organizing feature map neural network (SOFM-NN), DWT. interconnected to form a network, to perform complex
computational tasks. Because of its parallel architecture,
Artificial Neural Networks (ANNs)have been applied to image
1. INTRODUCTION compression problems, due to their superiority over traditional
The main objective of image compression is to decrease the methods when dealing with noisy or incomplete data[7][11].
redundancy of the image data which helps in increasing the ANNs seem to be well suited to image compression, as they
capacity of storage and efficient transmission. Image have the ability to preprocess input patterns to produce simpler
compression aids in decreasing the size in bytes of a digital patterns with fewer components. This compressed information
image without degrading the quality of the image to an preserves the full information obtained from the external
undesirable level. There are two classifications in image environment. Neural Networks based compression gives
compression: lossless and lossy compression. The reduction in security of the data to be compressed. Many different training
file size allows more images to be stored in a given amount of algorithms and architectures have been used. Different types of
disk or memory space. This supports in decreasing the time Artificial Neural Networks have been trained to perform image
required for the image to send or download from internet. Compression. In image compression technique, the compression
is achieved by training a neural network with the image and
Apart from the existing technology on image compression then using the weights and the coefficients from the hidden
represented by series of JPEG, MPEG and H.26x standards, layer as the data to recreate the image.
technology such as neural networks and genetic algorithms are
being developed to explore the future of image coding.
Successful applications of neural networks with back 2.1.1.Self Organizing Feature Map Neural
propagation algorithm have now become well established and
other aspects of neural network involvement in various
Network
technologies are being studied [1]. Research activities on neural Out of three categories of neural network (NN) architectures,
networks for image compression do exist in many types of namely Feed forward networks, Feed Back networks and
networks such as –Multi Layer Perceptron (MLP) [2][3], Recurrent networks ,the self organizing neural net belongs to
Hopfield [14], Self-Organizing Map (SOM), Learning Vector last one that is recurrent, or competitive or unsupervised
Quantization ( LVQ)[5]. networks. A number of researchers have successfully used the
Self Organizing Feature map (SOFM) algorithm to image
Image compression using Wavelet Transforms is a powerful
compression [9][13]. SOFM’s have two properties. First, it
method of lossy compression that is preferred by scientists to
quantizes the input space like any other vector quantization
store the images at higher compression ratios[6]. Wavelet
method, which constitutes a lossy
facilitates image compression using less storage space with

35
 International Journal of Computer Applications (0975 – 8887)
Volume 95– No. 11, June 2014

compression of the image. Then, the topology preserving After all vectors from input dataset are clustered, the BMUs are
property of SOFM, coupled with the hypothesis that consecutive used to represent the blocks of the image as shown in figure 1.
blocks in the image will often be similar, and to a differential
entropic coder, constitutes a second (non-lossy) compression of 3. COMPRESSION BY WAVELET
the information. This means SOFM neural nets can give better TRANSFORM
compression ratios than other networks [4]. Wavelet transform is the representation of any arbitrary signal
SOFM is realized by a two-layer network, as shown in x(t) as a weighted sum of a set of basis functions called mother
Fig.1.The first layer is the input layer or fan-out layer with input wavelet  ,by dilations (  ) and translations (s) given as ,
neurons and the second layer is the output or competitive layer.
The two layers are completely connected. An input vector x ∈
Rk when applied to the input layer, is distributed to each of the 1 t τ
ψ (t)  ψ .......................(1)
s, τ s  s 
(m×n) output nodes Ok in the competitive layer. Each node in
this layer is connected to all nodes in the input layer; hence, it
has a weight vector prototype wij attached to it. The node in the Low frequencies are examined with low temporal resolution
competitive layer are cluster centers while grouping the data while high frequencies with more temporal resolution. A
points. wavelet transform combines both low pass and high pass
filtering in spectral decomposition of signals. A wavelet
transform combines both low pass and high pass filtering in
spectral decomposition of signals. In wavelet transform based
compression method, the original image or data matrix of size
say M×N is first decomposed in the horizontal direction into
two halves by the low pass filter Lh(z) and high pass filter Hh(z).
The output of each filter is then down sampled with a factor of
2. Then the output is further partitioned in the vertical direction
into two halves by the same filters. The output of each filter is
again down sampled with a factor of 2, which yields four image
bands LL, LH, HL and HH of the first scale as shown in Fig.3.
LL, LH and HL are termed as low-frequency bands and HH
band is called as high frequency bands[12],[15].

Figure 1. SOFM network architecture

In SOFM, each data from data set recognizes themselves by

competing for representation. The weight vectors initialization
is the starting process of SOFM mapping[4]. Then the sample
vector from data set is randomly selected and the map of weight
vectors is searched to find which weight best represents that
sample. Each weight vector has neighboring weights that are
close to it. The weight that is chosen is rewarded by being able
Figure 2: Two level wavelet decomposition
to become more like that randomly selected sample vector. The
neighbors of that weights are also rewarded by being able to
become more like the chosen sample vector. From this step the After wavelet decomposition, the lower frequencies bands are
number of neighbors and how much each weight can learn adopted as the bands with noteworthy features. Hence these
decreases over time. This whole process is repeated a large bands are collected together to form the wavelet coefficients.
number of times, usually more than 1000 times. Then the coefficients are quantized by vector quantization
technique with respect to an expected compression rate. The
In sum, learning occurs in several steps and over much iteration: obtained coefficients are capable of reproducing all the image
1. Each node's weights are initialized. data values at the time of reconstruction by inverse wavelet
transform process. In this work of hybrid compression, this
2. A vector is chosen at random from the set of training data. property of wavelet transform has been used to compress the
code vectors generated by SOFM neural network.
3. Every node is examined to calculate which one's weights are
most like the input vector. The winning node is commonly
known as the Best Matching Unit (BMU). 4. COMPRESSION BY HYBRID METHOD
4. Then the neighborhood of the BMU (r) is calculated. The The hybrid approach combines the advantage of both SOFM
amount of neighbors decreases over time. neural network and wavelet transform to enhance the
5. The winning weight is rewarded with becoming more like the performance of the compression[16],[17]. The SOFM neural net
sample vector. The neighbors also become more like the sample will be used to obtain the code vectors and then those code
vector. The closer a node is to the BMU, the more its weights vectors are again compressed by DWT This method will be
get altered and the farther away the neighbor is from the BMU, implemented with the help of following steps:-
the less it learns.
a) Pre processing of image. And converting image data blocks
6. Repeat step 2 for N iterations. into vectors.
b) Training of SOFM neural network to generate code vectors.

36
 International Journal of Computer Applications (0975 – 8887)
Volume 95– No. 11, June 2014

c) Apply wavelet transform to reduce the size of code vector. Figure 4 and Table 2 shows results of compression using hybrid
d) Encoding of wavelet coefficients. method.
e) Decompression of image in decoder using codebook &
inverse wavelet transform.
f) The quality analysis of the decompressed images is carried
out using performance parameters.

5. IMPLEMENTATION AND RESULTS

In order to evaluate the performance of the proposed hybrid
approach of image compression, standard images like Lena and
Cameraman (256 x 256) are used. The work is implemented
using MATLAB. Performance of the hybrid approach is Figure 4. Image -Lena_gray.tif ,(left) Original image
evaluated using various measures. They are used to judge the (right)Decompressed image using hybrid method with
quality of image compression. This work uses statistical PSNR=28.21dB.
parameters like compression ratio(CR), Peak Signal-to-Noise
Ratio (PSNR) and some the co-relational based quality Table 2. Results of hybrid ANN -DWT method
measures like AD (average difference) & IF(Image fidelity)
[18]. These measures look at correlation features between the Image PSNR IF AD CR
pixels of original and reconstructed image. The experiments are Block
carried out with the number of clusters as 64, 128 and 256 etc. size
The method first uses SOFM neural network to generate the
desired codebook. The image is blocked (4x4) and is fed the 2 x2 32.15 0.9969 -2.4851 256:1
network for training up to 100 epoch to generate the code 4 x4 26.21 0.9929 -4.1854 128:1
vectors. Each code vector will be of 16 elements . To reduce the 4 x4 27.21 0.9958 -3.8413 64:1
size of each vector Wavelet transform is used. The DWT will 4 x4 28.83 0.9964 -3.6959 32:1
reduce the size of the code vector from 16 elements to 4
8 x8 25.55 0.9920 -5.8840 32:1
elements. The mother wavelet used for DWT was ‘db1’.
8x8 26.01 0.9932 -5.3699 16:1
The experimentation with codebook generated using same
image was carried out and it has been observed that proposed 6. CONCLUSIONS
method gives higher compression ratio over SOFM-neural In this paper image compression using hybrid method of ANN
network(NN) method. The reconstructed image quality is also and DWT is discussed. The compression ratio using this method
retained. Following tables shows the results obtained and is almost doubled. The PSNR for Lena image by proposed
comparisons of both approaches using above mentioned quality method is 1-2 dB more as compared to SOFM neural network
measures. Figure 3.and table1 shows results of compression method. It has been observed that if the block size is reduced
using SOFM neural network only. Using hybrid method, for a then PSNR and CR increase. But the value of average difference
256 x 256 Lena image for example, if the block is chosen is 4x4 is increased. Thus the combination with discrete wavelet
and number of clusters or code vectors equal to 128, the PSNR transform can provide higher compression ratio along with
of 28.83 dB is obtained. The PSNR for same configuration, but slightly higher PSNR values than SOFM Neural network based
using SOFM–NN method is about 27.12dB. If the number of method.
code vectors are increased the PSNR value increases for both
methods. The important thing is compression ratio with hybrid The neural networks like SOFM have capability to compress the
approach is improved images at higher compression rates. Wavelet transform shows
faithful reconstruction of compressed vectors using
approximation coefficients from LL sub band. The
hybridization using both these benefits shows significant
improvements in the performance of the compression algorithm
which is used to enhance the performance of ANN based
compression methods. The future approach can make the use of
some statistical parameters like mean and standard deviation to
select the blocks of image or region of interest, the quality
measures can be further improved.

7. REFERENCES
Figure 3. Standard Image Lena_gray.tif (left) Original [1] Cottrell et.al.“Learning internal representation from gray-
image (right) Decompressed image by SOFM-NN scale images:, An example of extensional programming.”
In Proceedings of the Ninth Conference of the Cognitive
Table.1 Results for Image Lena, using SOFM- NN Science Society, 1988.
Image PSNR IF AD CR
[2] R.D. Dony, S. Haykin, “Neural network approaches to
Block dB
image compression”, Proc. IEEE, proc. Vol. 83, No. 2,
size
Feb. 1995, pp 288-303.
2 x2 27.33 0.9970 0.0364 128:1
4 x4 26.81 0.9953 0.1133 64:1 [3] A. Namphol, S. Chin, M. Arozullah, “Image compression
4 x4 27.12 0.9911 -0.0534 32:1 with a hierarchical neural network”, IEEE Trans.
4 x4 28.68 0.9917 -0.0542 16:1
8 x8 22.46 0.9839 -0.0123 16:1
8 x8 22.87 0.9853 -0.0541 8:1
8 x8 23.88 0.9861 -0.0534 4:1
37
 International Journal of Computer Applications (0975 – 8887)
Volume 95– No. 11, June 2014

Aerospace Electronic Systems, Vol 32, No.1, January [12] Amjan Shaik, et.al, “Empirical Analysis of Image
1996, pp326-337. Compression through Wave Transform and Neural
Network”. IJCSIT Vol. 2 2011
[4] Christophe Amerijckx et.al, “Image Compression by Self-
organized Kohonen maps”, IEEE Trans. on Neural [13] G.Boopathy.et.al, “Implementation of Vector Quantization
Networks,Vol.9, No.3, May 1998. for Image Compression - A Survey” Global Journal of
Computer Science and Technology Vol. 10 Issue 3 April
[5]J. Jiang, “Image Compression with neural networks – A 2010.
survey”, Signal Processing: Image communication,
Elsevier Science B.V., Vol.14, 1999, pp 737-760. [14] Robina Ashraf et.al “Adaptive Architecture Neural Nets for
Medical Image Compression” IEEE Proceeding, 2006.
[6] Antonini, M.; Barlaud, M.; Mathieu, P.; Daubechies, I,
“Image coding using wavelet transform” IEEE Trans. on [15] Rafael C. Gonzalez, Richard E. Woods, Digital Image
Image processing Vol.2 ,2002,pp 205-220. Processing, Third Edition, Pearson Education 2008.
[7] S. Anna Durai, E. Anna Saro, “Image Compression with [16] G.Boopathi,“An Image Compression Approach using
Back- Propagation Neural Network using Cumulative Wavelet Transform and Modified Self Organizing Map”,
Distribution Function”, International Journal for Applied IJCSI International Journal of Computer Science Issues,
Science and Engg. Technology 2006. Vol. 8, Issue 5, No 2, September 2011.
[8]S.S.Panda et.al, Image compression using back propagation [17] Chaabouni,I.“An improved image compression approach
neural network”, IJESAT 2012 Vol 2. with combined wavelet and self organizing maps” Electro
technical Conference (MELECON), Mar 2012 16th
[9] Ren-Jean Liou, et.al “Quadtree Image Compression Using IEEE Mediterranean.
Sub-Band DCT Features and Kohonen Neural Networks
by” IEEE Proc. 2008 . [18] Marta Mrak, et.al “Picture quality measures in image
compression systems”. EUROCON, Ljuijana, Slovenia.
[10]Y. Linde, A. Buzo, and R. M. Gray, “An algorithm for 2003.
vector quantizer design,” IEEE Trans. Commun.,
vol.COMM-28, pp. 84–95, Jan. 1980. [19] A.K. Krishnamurthy et al., “Neural Networks for VQ of
speech and Images”, IEEE journal On selected areas in
[11]Ivan Vilovic' “An Experience in Image Compression Using comm., Vol.8, No.8,sept.2006
Neural Networks”, 48th International Symposium
ELMAR-2006, 07-09 June 2006. [20] Dandawate et al. "Neuro-Wavelet based vector quantizer
design for image compression”, Indian Journal of Science
and Technology Vol.2 No. 10 Oct 2009.

38

View publication stats