2018 International Conference on Computational and Characterization Techniques in Engineering & Sciences (CCTES)

Integral University, Lucknow, India, Sep 14-15, 2018

Compression using Thresholding on Signal/Image

by applying Wavelet Analysis
Hema Tekwani Krishna Raj
Research Scholar, Department of Electronics Engineering Professor, Department of Electronics Engineering
Harcourt Butler Technical University Harcourt Butler Technical University
Kanpur, India Kanpur, India
[email protected] [email protected]

Abstract— Wavelet transform due to its exceptional II. THRESHOLDING, COMPRESSION AND DECOMPOSITION
properties is now-a-days a tremendous and exceptional tool
that is used for compression, de-noising, of image and signal. A. Fundamentals of Image Compression
Decomposing an image and then analyzing the basic behavior The goal of image compression is to find an image
of various parameters involved in approximation and details. representation in which pels are less correlated. Redundancy
In this paper a framework has been developed using and irrelevancy are the fundamentals used in image
daubechies vanishing moments and then analyzing the results
compression. Joint Photographic Experts Group (JPEG)
of contained energy in percentage has been done. This paper
proposes a simple or modified scheme for daubechies wavelet
provides a compression method with reasonable speed and
transform for signal or image compression. efficiency that is suitable of compressing continuous-tone
image data with a pixel depth of 6 to 24 bits [8]. To yield a
Keywords— Image Compression, Image De-noising, Soft better representation a signal has structural correlations.
Thresholding, Image Decomposition, Short Time Fourier White noise does not have redundancies and thus it is easily
Transform, Continuous Wavelet Transform, Discrete Wavelet compressable. A compression method is adaptable if it
Transform, Multiresolution Analysis, Discrete Cosine provides a distinguishing feature between signal and noise.
Transformation. To achieve simultaneous de-noising and compression, the
non-zero threshold wavelet coefficients need to be quantized.
I. INTRODUCTION Since the works of Donoho and Johnstone, there has been
The skill of diminishing the quantity of data required to much research on finding thresholds for nonparametric
present a given information is compression. To maintain an estimation in statistics. There are two thresholding methods
appreciable fidelity, the important goal of compression is to frequently used. The shrinkage function also called soft
decrease the bit rate for transmission or storage [3]. In thresholding and the other is hard threshold function [11].
practical, theoretical and in applied areas wavelets have B. Wavelet Thresholding
generated an amazing interest. In 1987, wavelets were first
shown to be the authority of a powerful new approach to Wavelet thresholding is one of the important technique
signal processing and analysis called multi-resolution theory. for removing or killing the noisy parameters by removing the
With multi-resolution theory, the analysis of signals/images coefficients that are insignificant. Thresholding have the
became quite simple at more than one resolution. From a capability of signal de-noising, depends heavily on the
mathematical point of view, images are two-dimensional choice of thresholding parameter. The choice of threshold
arrangement of depth values with locally varying data that depends on the choice of thresholding parameters and the
result from different combinations of abrupt features. In both choice of threshold determines to great extent [9]. In
time and frequency domains wavelets are bounded, so that signals/images energy is mostly fixed in a small number of
an image/signal can be decomposed at multiple levels of proportion so that the noise components are eliminated by
resolution and can be finally refined from low resolution to thresholding the wavelet coefficients. The coefficients of
high resolution [2], [8]. Hence it is simple to catch local these proportions are relatively large compared to other
features in a signal. With wavelet the varying window size is proportions or to any other signal/image that has its energy
allowable. Decomposition of an image/signal into wavelets advancement over a large number of coefficients. To
involves a combination of waveforms. The assumption of the preserve the original signal information, the smaller
wavelet decomposition is to convert the original raw image coefficients are set to zero. Wavelet de-noising using
into several components of approximation and details, with thresholding include three basic footsteps, to compute the
single low-resolution part called approximation and the other wavelet transform of the noisy image, second one to apply
parts called details. Approximation and details at different thresholding on noisy wavelet coefficient and then to
levels of image helps in analyzing its various parameters and compute inverse wavelet transform of altered wavelet
the percentage of energy contained in it, using the coefficients [4]. Wavelet coefficient of noisy signal/image
thresholding gives the approximate details involved in the are gathered based on certain threshold value and threshold
compression. function. Wavelet de-noising using thresholding algorithm is
also known as wavelet shrinkage. The frequency behavior of

978-1-5386-4254-2/18/$31.00 ©2018 IEEE 17

the signal is reflected by wavelet decomposition. The Approximation Level 1 Horizontal Details
extreme point is based on the essential property of the
wavelet transform: father and mother functions, and these
20 20
are well bounded in spatial domain [11].
40 40
Original Soft Thresholding
5 5
60 60
4 4
20 40 60 20 40 60
3 3
Vertical Details Diagonal Details
2 2

1 1 20 20

0 0 40 40

-1 -1
60 60
-2 -2 20 40 60 20 40 60

-3 -3 Fig 2. Four level decomposition of tartan, defining approximation and

details, using db3.
-4 -4

-5 -5
It is called that only LL sub-band has relevant
0 200 400 600 0 200 400 600 information and all other sub-bands have less relevant data.
Fig 1. Signal showing soft threshold with thr = 0.5 The mathematics involved in wavelet analysis is
hierarchically decomposing the various functions and thus it
C. Decomposition is suitable to approximate the discontinuities. The function at
Wavelet transform has become a unique framework for various scales will be influenced by the wavelet
compression, de-noising etc due to its extreme performance. decomposition of a function consists of approximation
There is two band filtering, the image is high and low pass together with detail coefficients [2]. At different resolutions
filtered along the rows and the columns, down sampling and and places wavelets can be scaled and shifted to analyze the
up sampling processes by two are done. The high and low frequency contents of an image or signal and therefore it
frequency components are represented by these two sub- analyzes the image at multiple resolutions and then making it
bands. The easiest form of decomposition is pyramidal efficient tool for multi-resolution.
decomposition. The image and signal quality retains by not Input Signal
distorting their features. Filtering is the foremost approach of 10
image de-noising. These methods are mainly focused on
thresholding the discrete wavelet transform coefficients, 5

which have been affected by additive white Gaussian noise.

Amplitude

The decomposition breaks the signal into lower-resolution 0

components. Data is filtered at every level of decomposition. -5

A Discrete Wavelet Transform (DWT) which provides a Channel 1

compact multi-resolution representation of the image where -10

multi-resolution supplies a simple hierarchical framework Offset=0
for interpreting the image information. The DWT
(a)
decomposes an image into a set of smaller orthonormal
images. The original image is decomposed at different Compressed Signal

scales, to obtain a set of biorthogonal subscales of images. 10

The wavelet decomposition of data is derived from 2-

channel sub-band filtering with two filter sequences, the 5

smoothing or scaling filter and the detail or wavelet filter. In

Amplitude

0
reality the decomposition is truncated after a given number
of steps [3]. The hierarchical wavelet decomposition will be -5
used for images suggested by mallat. The low pass (L) and Channel 2

high pass (H) filters are applied to the image or signal in the -10

horizontal and vertical directions, and the filter outputs sub- Offset=0
sampled by a factor of 2, and generating the orientations HH, (b)
HL, LH and LL corresponding to high and low pass filters.

18
 Residual Signal The discrete wavelet transform of image f(x, y) of size M X
10
N is then
5 ெିଵ ேିଵ
ͳ
Amplitude

0
ܹʣ ሺ݆଴ ǡ ݉ǡ ݊ሻ ൌ  ෍ ෍ ݂ሺ‫ݔ‬ǡ ‫ݕ‬ሻʣ௝బ ǡ௠ǡ௡ ሺ‫ݔ‬ǡ ‫ݕ‬ሻሺ͹ሻ
ξ‫ ܰܯ‬௫ୀ଴ ௬ୀ଴
-5 ெିଵ ேିଵ
ͳ
-10 Channel 3
ܹ᫣௜ ሺ݆ǡ ݉ǡ ݊ሻ ൌ ෍ ෍ ݂ሺ‫ݔ‬ǡ ‫ݕ‬ሻ᫣௜௝ǡ௠ǡ௡ ሺ‫ݔ‬ǡ ‫ݕ‬ሻሺͺሻ
ξ‫ ܰܯ‬௫ୀ଴ ௬ୀ଴
0 2 4 6 8 10 12
Offset=0 Time (secs)
In signal processing DWT plays an important role. Signal
(c) reconstruction and decomposition are major issues in DWT
and inverse DWT. The elementary trick behind
Fig 3. Motivation for threshold. (a) show a sinusoidal pulse function with its
corrupted version, (b) its compressed version and (c) the residual signal of decomposition and reconstruction is low-pass and high-pass
the corrupted version. filtering with the use of down sampling and up sampling
respectively. The conclusion of wavelet decomposition is
III. WAVELET TRANSFORM IN TWO DIMENSIONS hierarchically formulated decompositions. One can prefer
In engineering there have been tremendous developments the level of decomposition j based on a desired cut-off
in the analysis of signals, especially in the form of the Short- frequency [7], [6]. Through filtering of the signal using high
Time Fourier Transform (STFT). In the advancement of pass and low pass filters, decomposition is achieved. In the
wavelet, mother wavelets place an important role. Data interest of the signal is fragmented into discrete frequency
reduction and feature extraction are basic applications of bands of varying resolution, the filtering technique is known
wavelet transform. Time and frequency cannot be resolved as multi-resolution analysis (MRA) [5].
simultaneously, according to heisenberg’s uncertainty
principle, i.e. it is not possible to realize certainty in both IV. CONCLUSION
time and frequency at the equal time, the resolution issue There are tremendous issues to be solved in image or
associated with STFT has been found to be a result of this signal compression. A vast research has also been done in
principle [2]. Windows have to join neighboring windows wavelet based techniques like image or signal compression,
smoothly, the uncertainty principle provides path on how to de-noising etc. So the approach of this paper is to achieve
select windows to resolve adverse effects. The window high compression in signal or image using Daubechies
dilemma remains- a narrow window yields poor frequency Wavelet Transform. The experimental results shows the
resolution, while a wide window provides poor localization. retained energy for image at level 2 and 3. Using db3 at level
The wavelet transform goes further than the Short Time 2 and 3 the percentage in energy retained is improved.
Fourier Transform. Wavelet transform analyzes the signal or
image by multiplying it by a window function and Level db1 db2 db3
performing an orthogonal expansion, analogously to other Level 2 87.73% 87.07% 88.02%
linear integral transformations. The shape of the wavelet
allows to represent different detail or resolution, to represent Level 3 89.33% 89.18% 91.14%
the location of the events in time, the location of the wavelet
allows. To represent a wide variety of signals wavelets can REFERENCES
be adapted. The one dimensional transforms are easily [1] Zhao Hong-tu, Yan Jing, “The Wavelet Decomposition and
extended to two-dimensional functions like images. In two Reconstruction Based on the Matlab”, College of Computer Science
dimensions, a two dimensional scaling function, Ф(x, y) and and
three two-dimensional wavelets, ѱH(x, y), ѱV(x, y) and ѱD(x, [2] Technology, Henan Polytechnic University, JioZou, China, July
2010, pp. 143-145.
y), are required [7], [10]. All are the production of two one-
dimensional functions. Excluding products that produce one- [3] Karan Lees, “Image Compression using Wavelets”, May 2002.
dimensional results, like Ф(x)ѱ(x), the four remaining [4] A.M. Atto, D. Pastor, G. Mercier, “Wavelet Shrinkage: Unification of
Basic Thresholding Functions and Thresholds”, 14 September, 2012.
products produce the separable scaling function [7], [10]:
[5] M. Antonini, M. Barlaud and I. Daubechies, “Image Coding using
ʣሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ ʣሺ‫ݔ‬ሻʣሺ‫ݕ‬ሻሺͳሻ Wavelet Transform”, IEEE Transaction on Image Processing Vol. 1,
No.2, pp. 205-220, APRIL 1992.
and separable, directionally sensitive wavelets [6] S. Kaewpijit, J. Le Moigne, T. El-Ghazawi, “Spectral Data Reduction
᫣ு ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ ᫣ሺ‫ݔ‬ሻʣሺ‫ݕ‬ሻሺʹሻ Via Wavelet Decomposition”, USA.
᫣௏ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ ʣሺ‫ݔ‬ሻ᫣ሺ‫ݕ‬ሻሺ͵ሻ [7] R. Haddadi, E. Abdelmounim, M. El Hanine and A. Belaguid,
“Discrete Wavelet Transform Based Algorithm for Recognition of
᫣஽ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ ᫣ሺ‫ݔ‬ሻ᫣ሺ‫ݕ‬ሻሺͶሻ QRS Complexes”, World of Computer Science and Information
Technology Journal, ISSN: 2221-0741, Vol. 4, No. 9, 127-132, 2014.
These wavelets calculate functional variations, intensity [8] R.C. Gonzalez and R.E. Woods, “Digital Image Processing”, Pearson
variations for images, along different directions. The Publication, Third Edition.
directional sensitivity is a natural consequence of the [9] Amir Averbuch, Danny Lazar and Moshe Israeli, “ Image
separability, it does not increase the calculated involvement Compression using Wavelet Transform and Multiresolution
Decomposition”, IEEE Transactions on Image Processing, Vol. 5,
of the 2-D transform [7], [11]. We first define the scaled and No. 1, January 1996.
translated basis functions: [10] S. Grace Chang, Student Member, IEEE, Bin Yu, Senior Member,
ʣ௝ǡ௠ǡ௡ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ  ʹ௝Τଶ ʣ൫ʹ௝ ‫ ݔ‬െ ݉ǡ ʹ௬ ‫ ݕ‬െ ݊൯ሺͷሻ IEEE, and Martin Vetterli, Fellow, IEEE, “Adaptive Wavelet
Thresholding for Image Denoising and Compression”, IEEE
᫣௜௝ǡ௠ǡ௡ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ  ʹ௝Τଶ ᫣௜ ൫ʹ௝ ‫ ݔ‬െ ݉ǡ ʹ௝ ‫ ݕ‬െ ݊൯ǡ ݅ ൌ ሼ‫ܪ‬ǡ ܸǡ ‫ܦ‬ሽሺ͸ሻ Transactions on Image Processing, Vol. 9, No. 9, September 2000.

19
[11] Stephane G. Mallat, “A Theory for Multiresolution Signal [12] Stephane Mallat, “A Wavelet Tour for Signal Processing: The Sparse
Decomposition: The Wavelet Representation”, IEEE Transactions on Way”, Third Edition.
Pattern Analysis and Machine Intelligence, Vol. 11, No. 7, July 1989.

20