International Conference on Computational Intelligence and Multimedia Applications 2007

Image compression using SVD

Prasantha H S, [email protected]
Shashidhara H L, [email protected]
Balasubramanya Murthy K N, [email protected]
PES Institute of Technology, Bangalore, Karnataka, India
Abstract
It is well known that the images, often used in variety of computer applications, are difficult to
store and transmit. One possible solution to overcome this problem is to use a data compression
technique where an image is viewed as a matrix and then the operations are performed on the
matrix. Image compression is achieved by using Singular Value Decomposition (SVD) technique
on the image matrix. The advantage of using the SVD is the property of energy compaction and
its ability to adapt to the local statistical variations of an image. Further, the SVD can be
performed on any arbitrary, square, reversible and non reversible matrix of m x n size. In this
paper, SVD is utilized to compress and reduce the storage space of an image. In addition, the
paper investigates the effect of rank in SVD decomposition to measure the quality in terms of
MSE and PSNR.

1. INTRODUCTION:
Images, used in Image Processing application, require big storage space and a considerable
bandwidth for transmission. One of the possible solutions to this problem is data compression
whose main goal is to reduce the quantity of data used to represent a digitized image. The data
compression techniques always eliminate the redundant data. The process of compressing (or
decompression) images can be classified into two categories, lossless compression and lossy
compression. Lossless methods produce exact copy of the original data and have a limit of
reduction in terms of the entropy. Lossy techniques usually work in terms of the incapacity of the
human visual system to detect small details and variations in images. The elimination of those
details reduces the amount of space required to store the image. Under certain circumstances,
lossy compression can be used to achieve the requirements of storage space, although the image
quality is reduced by the compression process. A gray scale image can be seen like a matrix of m
x n size. A digital color image has the size of m x n x 3. For a CIF image that has a size of 352 X
288 if each pixel is represented in 8 bits for R, G and B components, the storage requirement is
2.4 MB. This establishes the need for compression.

The paper is organized as follows. Section 2 presents the singular value decomposition and
section 3 discusses the implementation details. The test results are given in section 4 while
section 5 indicates the conclusions.

2. SINGULAR VALUE DECOMPOSITION:

The method refactors matrix A in three new matrices U, S, and V in such way that, U and V are
orthogonal matrices and S is a diagonal matrix
A = U S VT
A= [

A = Ur VrT = Ar

3. IMPLEMENTATION DETAILS
Three different sets of image of size 291 x 240, 214 x 244 and 256 x 256 are considered for the
analysis. Validation of the proposed project is done using MATLAB version 7.0 at PESIT R & D

0-7695-3050-8/07 $25.00 © 2007 IEEE 143

DOI 10.1109/ICCIMA.2007.386

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 11, 2008 at 04:22 from IEEE Xplore. Restrictions apply.
 Centre of Telecommunication Engineering department. For each image S matrix is found and
graph is plotted for different ranks. For each image reconstructed matrix is found from U, S and
V matrix and reconstructed image is displayed. In all the cases, differences between the original
and the reconstructed pixel intensities are plotted. In each case, the MSE and PSNR are evaluated
for different ranks. The compression ratio is calculated and plotted as a function of rank of the
image matrix. A graph is drawn for different ranks versus MSE and PSNR.

4. TEST RESULTS:
To demonstrate the effect of SVD on compression three set of sample images are taken. The sizes
of the images are of approximately same. The sample images are as shown in Fig 1. The
exhaustive surveys are conducted for analyzing the scheme. A sample of the result is reported
below.
Original ORIGINAL IMAGE

Fig. 1 Input Images that are considered correspond to boy, text and camera man
4
x 10 SVD Reconstructed difference Histogram HIS TOOGRAM OF ORIGINAL
3 180 300

160
2.5 250
140

2 120 200

100
1.5 150
80

1 60 100

40
0.5 50

20

0 0
0 60 80 100 120 140 160 180 200 220 240
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 70 80 90 100 110 120 130 140 150 160

Fig. 1a) Energy compaction, reconstructed image, difference between original image and
reconstructed image, histograms of reconstructed and original for rank 2.
4
x 10 SVD Reconstructed differenc e histogram HIS TOOGRAM OF ORIGINAL
3 300 300

2.5 250 250

2 200 200

1.5 150 150

1 100 100

50 50
0.5

0 0
0 60 80 100 120 140 160 180 200 220 240
0 5 10 15 20 25 30 35 60 80 100 120 140 160 180 200 220 240

Fig. 1b) Energy compaction, reconstructed image, difference between original image and
reconstructed image, histograms of reconstructed and original for rank 32.
4
x 10 SVD Recons tructed difference his togram HIS TOOGRAM OF ORIGINAL
3 300 300

2.5 250 250

2 200 200

1.5 150 150

1 100 100

0.5 50 50

0 0 0
0 20 40 60 80 100 120 140 60 80 100 120 140 160 180 200 220 240 60 80 100 120 140 160 180 200 220 240

Fig. 1c) Energy compaction, reconstructed image, difference between original image and
reconstructed image, histograms of reconstructed and original for rank 128.

144

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 11, 2008 at 04:22 from IEEE Xplore. Restrictions apply.
 15
MS E VS RANK
52
PSNR VS RANK
Rank MSE PSNR CR
50 2 14.01 36.66 66:1
10
48
4 11.61 37.48 33:1
46
8 8.035 39.08 16.5:1
MSE

PS NR
44
16 4.931 41.20 8.25:1
42
5
32 2.674 43.85 4.12:1
40

38
64 1.2816 47.05 2.06:1
0
1 2 3 4 5 6 7 36
128 0.5327 50.86 1.03:1
1 2 3 4 5 6 7
rank in powers of 2
rank in powers of 2

Fig. 1d) Graph of MSE and PSNR versus rank Table 1: Variation of PSNR, MSE and
CR versus rank
GRAPHS:
Compression ratio versus rank PSNR versus Compression ratio
CR VS RANK PSNR versus compression ratio
70 60

60
55 IMAGE 2
image 1

50
50

40

P SNR
CR

45
30
image 3
40 IMAGE 1
20
IMAGE 3
image 2 35
10

0 30
1 2 3 4 5 6 7 0 10 20 30 40 50 60 70
rank in powers of 2 Compression Ratio

Fig. 4: Variation of compression ratio versus variation in rank and PSNR against compression
ratio.

5. CONCLUSIONS
The paper presented the process of image compression using SVD. It is observed that the matrix
coefficients move towards Y-axis as the rank of the image matrix increases which indicates that
the maximum energy is concentrated with only first few coefficients. It is also noticed that with
the increase in the rank of the reconstructed image matrix, the number of entries in the matrix
would increase which in turn improve the perceptual picture quality correspondingly. The
perceptual quality of the textual image is better compared to the picture image even for the
smaller rank. It can also be observed that SVD technique provides much better quality for text
data even at lower ranks compared to image objects. More compression ratio can be achieved for
smaller ranks. For text data perceptual quality of the picture is almost same as original for smaller
rank and there by increase the compression ratio. It can be observed from Fig 4 that PSNR falls
as compression ratio increases.

REFERENCES:
1. Dr. Edel Garcia, Singular Value Decomposition (SVD) A Fast Track Tutorial First Published on
September 11, 2006; Last Update: September 12, 2006
2. The singular Value Decomposition and It’s Application in Image processing, Christopher Jason Ogden,
Tammie Huff, Math-45-College of Redwoods, December 18, 1997
3. Jody S. Hourigan and Lynn V. McIndoo, The singular Value Decomposition, Linear Algebra- Math 45
4. Singular value decomposition and principal component analysis in a Practical Approach to Microarray
data Analysis (D.P. Berrar, W. Dubitzky, M. Granzow, eds.) Kluwer: Norwell, MA, 2003. pg. 91-109.
5. Lay, David C. Linear Algebra and its applications. Addison Wesley Longman, INC., 1997
6. Leon, Steve.Atlast: Computer Exercise for Linear Algebra, Prentice-Hall, Inc., 1997
7. Mulachy, Colm & Rossi, John.”Atlast: A Fresh approach to Singular Value Decomposition”

145

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 11, 2008 at 04:22 from IEEE Xplore. Restrictions apply.