Image compression based on

The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.1, February 2015
DOI : 10.5121/ijma.2015.7101 1
IMAGE COMPRESSION BASED ON
COMPRESSIVE SENSING USING WAVELET
LIFTING SCHEME
Sheikh Md. Rabiul Islam1
, Xu Huang2
, and Keng Liang Ou3
1,2
Faculty of Education, Science, Technology and Mathematics
University of Canberra, Australia
3
College of Oral Medicine, Taipei Medical University, Taiwan
ABSTRACT
Many algorithms have been developed to find sparse representation over redundant dictionaries or
transform. This paper presents a novel method on compressive sensing (CS)-based image compression
using sparse basis on CDF9/7 wavelet transform. The measurement matrix is applied to the three levels of
wavelet transform coefficients of the input image for compressive sampling. We have used three different
measurement matrix as Gaussian matrix, Bernoulli measurement matrix and random orthogonal matrix.
The orthogonal matching pursuit (OMP) and Basis Pursuit (BP) are applied to reconstruct each level of
wavelet transform separately. Experimental results demonstrate that the proposed method given better
quality of compressed image than existing methods in terms of proposed image quality evaluation indexes
and other objective (PSNR/UIQI/SSIM) measurements.
KEYWORDS
Compressive Sensing, CDF9/7, Basis Pursuit, Orthogonal Matching Pursuit, Lifting scheme.
1. INTRODUCTION
Many researches are on orthogonal transforms; such as discrete cosine transform (DCT) and
discrete wavelets transform (DWT) etc., which provides a unique representation of a given signal.
These transforms have been widely employed in signal processing. Also, these transforms
facilitate their practical applications, e.g., fast algorithm design for transforms and rate distortion
optimization for image and video coding [1]. On the divergent, redundant transforms or
dictionaries do not have a unique representation for a given signal. Expanding a signal raises an
ill-posed problem under a redundant transform or dictionary, but provides an optimized solution,
e.g., a sparse one, for a specific application. This property has benefited many applications such
as time-frequency analysis [2] [3] [4], signal denoising [4] [5], image and video coding [6] [7]
[8] , and compressive sensing [9].
In traditional imaging systems, images are uniformly sampled first at a high rate, and then most of
the sampling data are thrown away for the purpose of compression. A common question is why
we acquire this precious data, just to use a little part of it and discard the most. Over the past few
years, a new sampling theory called compressed sensing (CS) has emerged. CS is a new pattern
of obtaining signals, different from identical rate digitization followed by compression and often
used for transmission or storage [10] [11]. CS delivers us a basis for acquiring and compressing
signals simultaneously. The CS theory [11] states that if signals are sparse in some basis, then
they will be recovered from a small number of random linear measurements via attractable
convex optimization techniques. Innovation in CS have the potential to reduce the sampling rate

2
significantly in many signal processing and its applications such as cameras, medical scanners,
fast analog to digital converter, and high-speed radar are widely used.
Compressive sensing comprehends three key problems: sparse representation, measurement
matrix and reconstruction algorithm. In compressive sensing, the signal is passed first through
some sort of transformed (Fourier transform (FT), wavelet transform (WT), discrete cosine
transom (DCT)) into sparse domain or compressible, so that a measurement matrix irrelevant to
transform basis can be designed to measure the signal, and the measures values can realize the
exact or approximate signal reconstruction by solving numerical optimization techniques.
The main contribution of this paper is the development of a novel method for image compression
based on compressive sensing using wavelet lifting scheme which is faster, simpler, and also
keeping strong edge preservation. The method uses proposed sparse representation based on
CDF9/7 wavelet transform. To guarantee exact recovery of every ‫-ݔ‬ sparse signal, the
measurement matrix needs to be one-to-one on all ‫-ݔ‬sparse vectors. Under this circumstance,
Candes and Tao showed a slightly stronger condition RIP (Restricted Isometric Property) [12] . In
this paper, compare the best fit of sparse representation of image by CDF9/7 wavelet transform,
we use Gaussian measurement matrix [13], Bernoulli measurement matrix [14] and Random
orthogonal measurement matrix [15] and image reconstruction by convex optimization techniques
for reconstruction of image such as ݈ଵ norm which is called Basis Pursuit (BP) [16], and
Orthogonal Matching Pursuit (OMP) algorithm [17] . For comparison purposed of this novel
work, we have also used sparse basis DWT and DCT of this paper. Experimental results on open
sources database [18] show that our proposed approach with sparse basis on CDF9/7 wavelet
transform in CS have higher qualities compressed image than traditional approach with sparse
DWT and DCT in terms an image quality assessment scheme [19] [20].
This paper is organized as follows. Section 2 gives a review of the compressive sensing problem
statement. Section 3 describes sparse image representation by wavelet lifting scheme. Section 4
describes our proposed algorithm for image compression based on compressive sensing.
Experimental results are reported in Section 5. The paper ends with a brief conclusion.
2. COMPRESSIVE SENSING
Shannon’s Nyquist sampling theorem specifies that a signal should be sampled at a rate higher
than twice the maximal frequency of the signal for fidelity of signal reconstruction. For high
bandwidth signals, such as image and video, the required sampling rate becomes very high. Some
small coefficients of the discrete cosine transform (DCT) or wavelet transform (WT) coefficients
can be discarded with little affection the quality of the reconstructed signal significantly. This
fundamental idea was used in most existing signal compression techniques. The concept of
compressive sensing (CS) is to acquire significant information directly without first sampling the
signal in the traditional sense. It has been shown that if the signal is “sparse” or compressible,
then the acquired information is sufficient to reconstruct the original signal with a high
probability [8] [10] [11]. Sparsity is defined with respect to an appropriate basis, such as DCT or
WT for that signal. The theory of CS is also acquired measurements of the signal through a
process that is incoherent with the signal. Incoherence makes convinced that the information
developed is randomly spread out. In CS, a sensing technique should provide a sufficient number
of CS measurements in a non-adaptive manner, so that enables near perfect reconstruction.
According to compressive sensing theory three main steps for CS application:
a) Sparse representation of the signal.
b) Design ‫ܯ‬ × ܰ measurement matrix unrelated to transform basis to measure the signal
and develop an ‫-ܯ‬ dimensional measurement vector.

3
c) Reconstruction the signal by ‫-ܯ‬dimensional measurement vector.
The three criteria of CS can be described as below:
The following notation [8] as we have ݂ = ሼ݂ଵ, … … … , ݂ேሽ be ܰ real-valued samples of a signal,
which can be represented by the transform coefficients, ‫.ݔ‬ That is,
݂ = Ψ‫ݔ‬ = ∑ ‫ݔ‬௜߰௜
ே
௜ୀଵ (1)
where Ψ = ሾ߰ଵ, ߰ଶ, … … … ߰ேሿ is an ܰ × ܰ transform basis matrix, which determines the domain
where the signal is sparse and also ‫ݔ‬ = ሾ‫ݔ‬ଵ, ‫ݔ‬ଶ, … … … ‫ݔ‬ேሿ is an ܰ-dimension vector of
coefficients with ‫ݔ‬௜ =< ‫,ݔ‬ ߰௜ >. We assume that ‫ݔ‬ is ܵ-sparse, meaning that there are only
significant elements in ‫ݔ‬ with ܵ << ܰ.
Suppose a general linear measurement process computes inner products ‫ܯ‬ < ܰ between ݂ and a
collection of vectors, ߶௝ giving ‫ݕ‬௜ =< ݂, ߶௝ >; ݆ = 1 … … ‫.ܯ‬ If Φ denotes the ‫ܯ‬ × ܰ matrix
with ߶௝ as row vectors, then the measurements ‫ݕ‬ = ሾ‫ݕ‬ଵ, ‫ݕ‬ଶ, … … … ‫ݕ‬ெሿ are given by:
‫ݕ‬ = Φ݂ = ΦΨx = Θx (2)
where ‫ݕ‬ is M-dimensional observation vector and Φ is the ‫ܯ‬ × ܰ random measurement matrix
and Θ = ΦΨ is called sensing matrix. For reconstruction ability ‫ݔ‬ is ܵ-sparse, if Θ satisfies the
Restricted Isometric Property (RIP) [12]. Define the restricted isometric property constant
ߜ௞ ∈ ሺ0,1ሻ for the sparse signal ‫ݔ‬ for any ܵ as the minimum value for the constitution of the
following formula:
ሺ1 − ߜ௞ሻ‖‫‖ݔ‬ଶ
ଶ
≤ ‖Θ‫‖ݔ‬ଶ
ଶ
≤ ሺ1 + ߜ௞ሻ‖‫‖ݔ‬ଶ
ଶ
(3)
Currently, researchers developed some measurements matrix are follows: Gaussian random
matrix [13], binary random matrix (Bernoulli matrix) [14], Fourier random matrix [21],
Hadamard matrix [22], random orthogonal matrix [15].
The signal reconstruction problem involves using ‫ݕ‬ to reconstruct the ܰ-length signal, ‫ݔ‬ that is ܵ-
sparse, given Φ and Ψ. Since, ‫ܯ‬ < ܰ, this is an ill-conditioned problem and there are infinite
many ‫ݔ‬ᇱ
that is satisfy Θ‫ݔ‬ᇱ
= ‫.ݕ‬ The conventional approach to solving ill-conditioned problems
of this kind is to minimize the ݈௣ norm. Define the ݈௣ norm vector ‫ݔ‬ as ‖‫‖ݔ‬௣ = ∑ |‫ݔ‬௜|௣ே
௜ୀଵ for
1 ≤ ‫݌‬ < ∞ and ‖‫‖ݔ‬௣ = ݉ܽ‫ݔ|ݔ‬௜|. In practice, signals are often encountered that are not exactly
sparse, but whose coefficients decay rapidly. As mentioned, compressible signals are that
satisfying power law decay:
‖‫ݔ‬∗‖ ≤ ܴ݅
ଵ
௤
Considered ݈ଶ norm to solve this ill-condition and the optimization problem is given by:
‫ݔ‬ො = ܽ‫ݔ‖݊݅݉݃ݎ‬ᇱ‖ଶ such that Θ‫ݔ‬ᇱ
= ‫ݕ‬ (4)
It has been demonstrated that this ݈ଶ minimization can only produce a non-sparse ‫ݔ‬ො [10]. The
purpose is that the ݈ଶ norm measures the energy of the signal, and signal sparsity properties could
not be combined in this measure. The ݈଴ norm counts the number of non-zero entries and,
therefore, allows us to specify the sparsity requirement. The optimization problem using this
norm can be stated as:
‫ݔ‬ො = ܽ‫ݔ‖݊݅݉݃ݎ‬ᇱ‖଴ such that Θ‫ݔ‬ᇱ
= ‫ݕ‬ (5)

4
There is a high likelihood of obtaining a solution using only ‫ܯ‬ = ܵ + 1 independent and
identical distribution (i.i.d.) Gaussian measurements [11]. However, the produced solution is
numerically unstable [10] . It turns out that optimization based on the ݈ଵ norm is able to exactly
recover S-sparse signals with a high probability using only ‫ܯ‬ ≥ ݈ܿܵ‫݃݋‬ሺܰ/ܵሻ i.i.d. Gaussian
measurements [10] [24]. The convex optimization problem is given by:
‫ݔ‬ො = ܽ‫ݔ‖݊݅݉݃ݎ‬ᇱ‖ଵ such that Θ‫ݔ‬ᇱ
= ‫ݕ‬ (6)
This can be reduced to a linear program. Algorithms based on basis pursuit (BP) [4] can be used
to solve this problem with a computational complexity of ܱሺܰଷ
ሻ .
A number of different algorithms have been developed to solve these three CS reconstruction
problems. They include linear programming (LP) techniques [10] [24] , greedy algorithms [2]
[25] [26] [27] , gradient-based algorithms [9] [28] and iterative shrinkage algorithms [29] [30].
Although ݈ଵ norm has strong guarantees of exact recovery, it has disadvantages in computational
cost and implementation complexity. As a result, another line of research that seems valuable to
explore is Orthogonal Matching Pursuit (OMP) algorithm. This recovery scheme is especially
simpler to implement and potentially faster than Basis Pursuit (BP). In this paper we have used
gradient-based algorithms as OMP algorithm [17] and basis pursuit (BP) algorithm [4] for
proposed image compression applications.
3. SPARSE IMAGE REPRESENTATION USING WAVELET LIFTING
SCHEME
DCT and DWT are the two most commonly used sparse basis in CS algorithms. Due to its
periodicity, DCT can well describe the texture characteristics of an image, but it easily may lead
to the block effect. DWT can be an excellent description of the edge characteristics of an image.
Therefore, it can compensate for this shortcoming of DCT. The reconstruction effects of DWT
sparse processing are mainly affected by wavelet function type and decomposition levels.
According to the contrast experiments of different wavelet basis, references [10] [14] found that
choosing symmlet wavelet function for image wavelet decomposition can produce the best
reconstructed image. Currently, multi-resolution pyramid decomposition and synthesis algorithm,
namely the Mallat algorithm, is the most commonly used in wavelet research area.
Cohen-Daubechies-Feauveau 9/7 (CDF 9/7) wavelet transform (WT) [31] is a lifting scheme
based wavelet transform which can reduce the computational complexity. A lifting is an
elementary modification of perfect reconstruction filters, which is used to improve the wavelet
properties. The lifting scheme is a flexible technique that has been used in several different
settings for easy construction and implementation of traditional wavelets and for construction of
wavelets on arbitrary domains such as bounded regions, second generation wavelets or surface,
spherical wavelets. To optimize the approximation and compression of signals and images, the
lifting scheme has also been widely used to construct adaptive wavelet basis with signal-
dependent lifting. The principle of lifting scheme is described as follows: consider an input image
x fed into a ℎ෨ (low pass filter) and ݃෤ (high pass filter) separately. The outputs of the two filters are
then down sampled by 2ሺ↓ 2ሻ. The resulting low-pass subband ‫ݕ‬௅ and high-pass subband ‫ݕ‬ு are
shown in Figure 1. The original signal can be reconstructed by synthesis filters h(low pass) and
g(high pass), which take the up-sampled by 2 ሺ↑ 2ሻ for ‫ݕ‬௅ and ‫ݕ‬ு as inputs [32] [33]. An analysis
and synthesis system has the perfect reconstruction property if and only if x ' = x.
The mathematical representations of ‫ݕ‬௅ and ‫ݕ‬ு can be defined as
ቊ
௬ಽሺ௡ሻୀ∑ ௛෩ሺ௜ሻ௫ሺଶ௡ି௜ሻ
ಿಽషభ
೔సబ
௬ಹሺ௡ሻୀ∑ ௚෤ሺ௜ሻ௫ሺଶ௡ି௜ሻ
ಿಹషభ
೔సబ
(7)

5
Where ܰ௅ and ܰு are the lengths of ℎ෨ and ݃෤ respectively.
)(
~ 1−
zh
)(~ 1−
zg
2↓
2↓
)( 1−
zh
)( 1−
zg
2↑
2↑
Ly
Hy
x 'x
Figure 1: Discrete wavelet transform (or subband transform) analysis and synthesis system.
In this paper, we developed multi-layer lifting scheme WT with sparse basis, which decomposes
an image ‫ݔ‬ into 4 parts for each layer: low-low frequency LL1, high-low frequency HL1, low-
high frequency LH1, and high-high frequency HH1. LL1 is a sub-band corresponding to low
frequency in both vertical and horizontal direction; HL1 is a sub-band corresponding to high
frequency in vertical and low frequency in horizontal direction; LH1 is a sub-band corresponding
to low frequency in vertical and high frequency in horizontal direction; HH1 is a sub-band
corresponding to high frequency in both vertical and horizontal direction. For this first layer, an
image ‫ݔ‬ is decomposed into 4 parts LL1, HL1, LH1 and HH1 .This concept is also applied to the
second and third level decompositions based on the principle of multiresolution analysis. For
example the LL1 subband is decomposed into four smaller subbands: LL2, HL2, LH2, and HH2.
The three layer subbands are sparse so we can adopt Orthogonal Matching Pursuit (OMP)
algorithm [17] or Basis Pursuit (BP) to rebuild these parts directly.
4. PROPOSED METHOD
For image processing application, we proposed an image compression based on compressive
sensing using wavelet lifting scheme framework as shown in Figure 2. The objective of novel
image compression based on compressive sensing process is to estimate the original image ‫ݔ‬ with
dimension ܰ × ܰ pixels from the function ݂ as show in Equation (1). The parameters are denoted
as ‫ݔ‬ is an image, Ψ is an ܰ × ܰ transform basis matrix, Φ is the ‫ܯ‬ × ܰ measurement matrix,
OMP is Orthogonal Matching Pursuit algorithm, BP is Basis Pursuit algorithm and ‫ݔ‬ᇱ
reconstruction image. Our proposed algorithm is designed as shown in Figure 2.
Figure 2: Proposed image compression framework based on compressive sensing.
Proposed Algorithm:
Stage 1: Compute ߖ :
a. Takes sparsity of an image ݂ = Ψ‫ݔ‬ = ∑ ‫ݔ‬௜߰௜
ே
௜ୀଵ , where ‫ݔ‬ is the sparse vector with only
ܵ << ܰ and ‫ݔ‬௜ is none zero elements.
b. Perform sparse domain CDF9/7 wavelet transform to signal ݂ into several sparse
directions to CDF9/7 wavelet subband ݂௝.௟, where ݆ is the decomposition level and ݈ the
number of direction level at each scale. In general the image ݂ is transformed into a ‫ݔ‬
sparse signal with only ‫ݔ‬௜ non-zero elements. We are developed into three levels of the
CDF9/7 wavelet transform coefficients are respectively expressed as

6
݂መଵ
= ܹଵ
݂ = ൣ݂መ௅௅
ଵ
݂መ௅ு
ଵ
݂መு௅
ଵ
݂መுு
ଵ
൧
்
݂መଶ
= ܹଶ
݂መ௅௅
ଵ
= ൣ݂መ௅௅
ଶ
݂መ௅ு
ଶ
݂መு௅
ଶ
݂መுு
ଶ
൧
்
and
݂መଷ
= ܹଷ
݂መ௅௅
ଶ
= ൣ݂መ௅௅
ଷ
݂መ௅ு
ଷ
݂መு௅
ଵ
݂መுு
ଷ
൧
்
Where ܹ௝
, ݆ ∈ ሼ1,2,3ሽ represents the 2D CDF9/7 wavelet transform (WT) matrix of
level ݆.
Stage 2: Compute Φ:
a. Design ‫ܯ‬ × ܰ dimension observation matrix Φ , where ‫ܯ‬ < ܰ , then use Φ to
measure ݂ே×ே and develop the observation vector ‫ݕ‬ெ×ଵ .
Stage 3:
a. Apply CS scheme to each direction and decomposition level as follows:
‫ݕ‬ = Φ௝,௟Ψ‫ݔ‬௝,௟ (8)
Stage 4: Reconstruction:
a. Apply ݈ଵ norm approach is called Basis Pursuit (BP) or Orthogonal Matching Pursuit
(OMP) to reconstruct the signal ‫ݔ‬ from ‫ݕ‬ of Equation (8).
b. Reconstruction of subbands is transformed back to the spatial domain by inverse
CDF9/7 wavelet transform to recovered image ‫ݔ‬ᇱ
.
Stage 5: Evaluation:
To evaluate the compressed image using proposed quality indexes [19] [20] and
others evaluation of image quality (EIQ) indexes.
Stage 6: Comparison:
We use sparse basis DWT or DCT to decompose the image into feature by replacing
of stage 1(b).
5. EXPERIMENTAL RESULTS
In this section, we illustrate the reconstruction performance of compressed sensing image; we
deploy both the OMP and Basis Pursuit on proposed sparse basis CDF9/7 wavelet transform
(WT), DWT and DCT using Gaussian matrix, Bernoulli matrix and random orthogonal matrix
measurements. To evaluate the performance of proposed algorithm we used open source
databases [18]. The images are for 8-bit 256 × 256 pixels and 8-bit 64 × 64 pixels. We used
different image quality indexes: PNSR, UIQI, Q(Skewness), Q(Kurtosis), SSIM,
SSIM(Skewness), SSIM(Abs-Skewness), SSIM(Kurtosis) [19] [20] to evaluate the qualities of
the compressed images. In order to make clearer comparisons, besides our proposed sparse basis
CDF9/7 WT with three different measurements matrix’s and Basis Pursuit (BP) algorithms or
greedy optimization techniques as OMP and are also used. We made experiments on 15 images in
our proposed image compression based on CS methods. Comparing the evaluation of image
quality (EIQ) indexes for the compressed image at several measurements ‫ܯ‬ < ܰ (ܰ = 256), we
note that the higher measurements ‫,ܯ‬ the better quality of image compression will be. Figures 3
& 4 show the quality indexes of image compression obtained with proposed sparse CDF 9/7
wavelet transform (WT) and sparse DWT with OMP algorithm performed on the human body

7
muscle cells and cellulase cells image from novel image compression framework based on CS as
shown in Figure 2 . As example, in Figure 3 (a), the PSNR values of proposed sparse basis CDF
9/7 with three measurements matrix and OMP in proposed framework is showing higher values
than sparse basis DWT with three measurements matrix and OMP in the range [150-250].
Similarly, we also observes that other quality index values including new image quality indexes
of proposed sparse basis CDF 9/7 WT are higher values than others. However, the new image
quality indexes also show higher value than original UIQI and SSIM. Comparing the evaluation
of image quality (EIQ) indexes for the compressed image at several measurements ‫ܯ‬ < ܰ ሺܰ =
256ሻ, we note that the higher the measurements ‫ܯ‬ given the better the quality of image
compression will be. Figures 5 & 6 also show the quality indexes of image compression obtained
with proposed sparse basis CDF9/7 WT and other techniques of sparse basis DWT or DCT with
BP algorithm were also used on the Brain (MR)-1 and Brain (MR)-2 images.
(a) (b)
(c) (d)
(e) (f)
150 200 250
20
25
30
35
40
45
No. of Measurements(M)
PSNR(dB)
Proposed-CDF9/7-RandomOrtho-OMP
DWT-RandomOrtho-OMP
Proposed-CDF9/7-Gaussian-OMP
DWT-Gaussian-OMP
Proposed-CDF9/7-Bernoulli-OMP
DWT-Bernoulli-OMP
150 200 250
0.4
0.6
0.8
1
UIQI
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.4
0.6
0.8
1
Q(Skewness)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.5
0.6
0.7
0.8
0.9
1
Q(Kurtosis)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.5
0.6
0.7
0.8
0.9
1
SSIM
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.5
0.6
0.7
0.8
0.9
1
SSIM(Skewness)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP

8
(g) (h)
Figure 3: Plots of eight image quality indexes for human body muscle cells image versus number of
measurements ሺ‫ܯ‬ሻሺ‫ܯ‬ < ܰ, ܰ = 256ሻ in the range [150-250] for proposed sparse basis CDF9/7 wavelet
transform with three measurements matrix and OMP including the sparse basis DWT.
(a) (b)
(c) (d)
(e) (f)
150 200 250
0.5
0.6
0.7
0.8
0.9
1
SSIM(Abs-Skewness) Proposed-CDF9/7-RandomOrtho-OMP
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.5
0.6
0.7
0.8
0.9
1
SSIM(Kurtosis)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
30
35
40
45
PSNR(dB)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.65
0.7
0.75
0.8
0.85
0.9
UIQI
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.65
0.7
0.75
0.8
0.85
0.9
Q(Skewness)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.65
0.7
0.75
0.8
0.85
0.9
Q(Kurtosis)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.8
0.85
0.9
0.95
1
SSIM
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.8
0.85
0.9
0.95
1
SSIM(Skewness)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP

9
(g) (h)
Figure 4: Plots of eight image quality indexes for human body celulas cell image versus number of
measurements ሺ‫ܯ‬ሻሺ‫ܯ‬ < ܰ, ܰ = 256ሻ in the range [150-250] for proposed sparse basis CDF9/7 wavelet
transform with three measurements matrix and OMP including the sparse basis DWT.
The experimental results clearly show that the proposed method based on sparse basis CDF 9/7
WT by OMP with three different measurements matrix as Gaussian, Bernoulli and random
orthogonal is out-performed than all other methods with various number of measures ሺ‫ܯ‬ሻሺ‫ܯ‬ <
ܰ, ܰ = 256ሻ in the range [150-250]. However the proposed sparse basis CDF9/7 WT with BP
algorithm is also better over all other methods with different number of measurements ሺ‫ܯ‬ሻሺ‫ܯ‬ <
ܰ, ܰ = 4096ሻ in the range [1500-3500].
Figure 7 shows the visual comparison of compressed images. The comparison is clearer with data
plotted in Figure 3, which shows the relationship between the image quality index UIQI,
Q(Skewness), Q(Kurtosis), SSIM, SSIM(Skewness), SSIM(Absolute Skewness),
SSIM(Kurtosis), for the different number of measurements (M) values performed on an human
body muscle image. Similarly for Brain (MR)-1 image, the visual comparison of the compressed
image in Figure 8 with data plotted in Figure 5.
(a) (b)
(c) (d)
150 200 250
0.8
0.85
0.9
0.95
1
SSIM(Abs-Skewness)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
150 200 250
0.8
0.85
0.9
0.95
1
SSIM(Kurtosis)
DWT-RandomOrtho-OMP
DWT-Gaussian-OMP
DWT-Bernoulli-OMP
1500 2000 2500 3000 3500
20
25
30
35
40
45
PSNR(dB)
Proposed-CDF9/7-RandomOrtho-BP
DCT-RandomOrtho-BP
Proposed-CDF9/7-Gaussian-BP
DCT-Gaussian-BP
Proposed-CDF9/7-Bernoulli-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
UIQI
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
Q(Skewness)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
Q(Kurtosis)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP

10
(e) (f)
(g) (h)
Figure 5: Plots of eight image quality indexes for brain (MR)-1 image versus number of measurements
ሺ‫ܯ‬ሻሺ‫ܯ‬ < ܰ, ܰ = 4096ሻ in the range [1500-3500] for proposed sparse basis CDF9/7 wavelet transform
with three measurements matrix and BP including the sparse basis DCT.
In Figure 7, represent the visual comparison why does need proposed algorithm based on
compressive sensing. The visual effect is showing the compressed image with the proposed
algorithm (CDF9/7 WT with OMP) algorithm) in Figure 7(b) random orthogonal measurement
matrix, Bernoulli measurement matrix in Figure 7(c), and Gaussian measurement matrix (where
ܰ = 256 and ‫ܯ‬ = 190) in Figure 7(d) higher quality as compared with compressed image of
DWT with OMP algorithm for Figure 7(e) - (g).
Similarly, we can also compare of compressed image with proposed algorithm (CDF9/7 WT with
BP algorithm) for Figure 8 (b) – (d) (where ܰ = 4096 and ‫ܯ‬ = 3000), therefore showing higher
quality as compared with compressed image of DCT with BP algorithm for Figure 8(e) -(g).
(a) (b)
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM(Skewness)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM(Abs-Skewness)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM(Kurtosis)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
10
20
30
40
50
PSNR(dB)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.2
0.4
0.6
0.8
1
UIQI
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP

11
(c) (d)
(e) (f)
(g) (h)
Figure 6: Plots of eight image quality indexes for brain-2 (MR) image versus number of measurements
ሺ‫ܯ‬ሻሺ‫ܯ‬ < ܰ, ܰ = 4096ሻ in the range [1500-3500] for proposed sparse basis on CDF9/7 wavelet transform
with three measurements matrix and BP including the sparse basis on DCT.
1500 2000 2500 3000 3500
0.4
0.5
0.6
0.7
0.8
0.9
Q(Skewness)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
Q(Kurtosis)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM(Skewness)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM(Abs-Skewness)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP
1500 2000 2500 3000 3500
0.4
0.6
0.8
1
SSIM(Kurtosis)
DCT-RandomOrtho-BP
DCT-Gaussian-BP
DCT-Bernoulli-BP

(b)
(e)
Figure 7: Compressed of human body
with proposed algorithm (CDF9/7 WT with OMP) for (b)
Bernoulli measurement matrix; (d) Gau
OMP for (e) random orthogonal measurement
measurement matrix (where ܰ = 256
As can be seen from all figures, proposed algorithm has increased
The values of image quality index
algorithm. These shows the proposed algorithm is
others. As we all know, image quality indexes value reflects the difference between
image and reconstructed image and the highe
algorithm is not only deals with higher
ensure reconstruction quality.
(a)
(c) (d)
(f) (g)
uman body muscle cells image (grey) (a) Original image ; compressed
with proposed algorithm (CDF9/7 WT with OMP) for (b) random orthogonal measurement matrix;
(d) Gaussian measurement matrix; and compressed image of DWT with
rthogonal measurement matrix; (f) Bernoulli measurement matrix
256 and ‫ܯ‬ = 190) .
As can be seen from all figures, proposed algorithm has increased quality images
e values of image quality index are better than others such as sparse DWT or DCT based
s the proposed algorithm is feasible and has distinct advantages
. As we all know, image quality indexes value reflects the difference between
image and reconstructed image and the higher is quality indexes, the better. The proposed
algorithm is not only deals with higher frequency part, but also deals with low-frequency
12
compressed image
rthogonal measurement matrix; (c)
image of DWT with
(f) Bernoulli measurement matrix; (g) Gaussian
quality images significantly.
or DCT based
advantages over
. As we all know, image quality indexes value reflects the difference between the original
. The proposed
frequency part to

13
(a)
(b) (c) (d)
(e) (f) (g)
Figure 8: Compressed of human brain (MR) image (grey) (a) Original image ; compressed image with
proposed algorithm (CDF9/7 WT with BP) for (b) random orthogonal measurement matrix; (c) Bernoulli
measurement matrix; (d) Gaussian measurement matrix; and compressed image of DCT with BP
algorithm for (e) random orthogonal measurement matrix (f) Bernoulli measurement matrix (g) Gaussian
measurement matrix (where ܰ = 4096 and ‫ܯ‬ = 3500) .
The main motivation was to develop a good image compression using proposed sparse CDF9/7
wavelet transform based on compressive sensing as much as possible and its quality evaluates by
proposed image quality indexes and others evaluation of image quality (EIQ) indexes. This
proposed algorithm can be used for compressive sensing image processing application.
6. CONCLUSION
In this paper, we proposed image compression based on compressive sensing using wavelet
lifting scheme framework that addresses the best-compressed image components and preservation
of high-frequency details in medical images. The proposed method is also compared with three
different matrix, i.e. Gaussian, Bernoulli and random orthogonal measurement matrix and image
reconstruction by convex optimization technique for reconstruction of the image via ݈ଵ − ݊‫݉ݎ݋‬
which is called Basis Pursuit (BP) and greedy pursuit such as Orthogonal Matching Pursuit
(OMP) algorithm. Experimental results demonstrate that the proposed sparse basis CDF9/7

14
wavelet transform can better compressed images than sparse DWT or DCT with OMP or Basis
Pursuit (BP) algorithm in proposed framework with CS.
REFERENCES
[1] S. Mallat, A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way, 3 edition.
Amsterdam ; Boston: Academic Press, 2008.
[2] S. G. Mallat and Z. Zhang, “Matching pursuits with time-frequency dictionaries,” IEEE Trans.
Signal Process., vol. 41, no. 12, pp. 3397–3415, Dec. 1993.
[3] S. G. Mallat, “Adaptive time-frequency decompositions,” Opt. Eng., vol. 33, no. 7, p. 2183, Jul.
1994.
[4] S. Chen, D. Donoho, and M. Saunders, “Atomic Decomposition by Basis Pursuit,” SIAM J. Sci.
Comput., vol. 20, no. 1, pp. 33–61, Jan. 1998.
[5] J. Jin, B. Yang, K. Liang, and X. Wang, “General image denoising framework based on compressive
sensing theory,” Comput. Graph., vol. 38, pp. 382–391, Feb. 2014.
[6] S. Fischer, G. Cristobal, and R. Redondo, “Sparse overcomplete Gabor wavelet representation based
on local competitions,” IEEE Trans. Image Process., vol. 15, no. 2, pp. 265–272, Feb. 2006.
[7] B. Wang, Y. Wang, I. Selesnick, and A. Vetro, “Video Coding Using 3D Dual-Tree Wavelet
Transform,” EURASIP J. Image Video Process., vol. 2007, no. 1, p. 042761, Mar. 2007.
[8] M. Y. Baig, E. M.-K. Lai, and A. Punchihewa, “Compressed Sensing-Based Distributed Image
Compression,” Appl. Sci., vol. 4, no. 2, pp. 128–147, Mar. 2014.
[9] M. A. T. Figueiredo, R. D. Nowak, and S. J. Wright, “Gradient Projection for Sparse Reconstruction:
Application to Compressed Sensing and Other Inverse Problems,” IEEE J. Sel. Top. Signal Process.,
vol. 1, no. 4, pp. 586–597, Dec. 2007.
[10] R. G. Baraniuk, “Compressive Sensing [Lecture Notes],” IEEE Signal Process. Mag., vol. 24, no. 4,
pp. 118–121, Jul. 2007.
[11] E. J. Candes and M. B. Wakin, “An Introduction To Compressive Sampling,” IEEE Signal Process.
Mag., vol. 25, no. 2, pp. 21–30, Mar. 2008.
[12] E. J. Candes and T. Tao, “The Power of Convex Relaxation: Near-Optimal Matrix Completion,”
ArXiv09031476 Cs Math, Mar. 2009.
[13] Y. Yu, A. P. Petropulu, and H. V. Poor, “Measurement Matrix Design for Compressive Sensing
Based MIMO Radar,” IEEE Trans. Signal Process., vol. 59, no. 11, pp. 5338–5352, Nov. 2011.
[14] Z. Zhang and J. Wang, “Bernoulli matrix and its algebraic properties,” Discrete Appl. Math., vol.
154, no. 11, pp. 1622–1632, Jul. 2006.
[15] G. Stewart, “The Efficient Generation of Random Orthogonal Matrices with an Application to
Condition Estimators,” SIAM J. Numer. Anal., vol. 17, no. 3, pp. 403–409, Jun. 1980.
[16] S. Chen and D. Donoho, “Basis pursuit,” in 1994 Conference Record of the Twenty-Eighth Asilomar
Conference on Signals, Systems and Computers, 1994, 1994, vol. 1, pp. 41–44 vol.1.
[17] J. A. Tropp and A. C. Gilbert, “Signal Recovery From Random Measurements Via Orthogonal
Matching Pursuit,” IEEE Trans. Inf. Theory, vol. 53, no. 12, pp. 4655–4666, Dec. 2007.
[18] “TEST IMAGES,” http://decsai.ugr.es/cvg/dbimagenes/index.php. .
[19] S. M. R. Islam, X. Huang, and K. Le, “A Novel Image Quality Index for Image Quality Assessment,”
in Neural Information Processing, M. Lee, A. Hirose, Z.-G. Hou, and R. M. Kil, Eds. Springer Berlin
Heidelberg, 2013, pp. 549–556.
[20] Sheikh Md. Rabiul Islam, Xu Huang, and Kim Le, “Novel Evaluation Index for Image Quality”, The
International Conference on Digital Image Computing: Techniques and Applications (DICTA), pp:
293-300, Wollongong, New South Wales, Australia, 2014.
[21] G. Xu and Z. Xu, “Compressed Sensing Matrices from Fourier Matrices,” ArXiv13010373 Cs Math,
Jan. 2013.
[22] M. Ozeki, “Generalized Hadamard matrices and related matrices and their applications to the
construction of the positive definite integral hermitian forms,” J. Comb. Theory Ser. A, no. 3, pp.
270–275, 1981.
[23] “Compressive Sensing: The Big Picture - igorcarron2.” [Online]. Available:
https://sites.google.com/site/igorcarron2/cs. [Accessed: 09-Sep-2014].
[24] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles: exact signal reconstruction
from highly incomplete frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–
509, Feb. 2006.

[25] J. A. Tropp and A. C. Gilbert, “Signal recovery from partial information via Orthogonal Matching
Pursuit,” IEEE TRANS Inf. THEORY
[26] D. L. Donoho, Y. Tsaig, I. Drori, and J.
Linear Equations by Stagewise Orthogonal Matching
2, pp. 1094–1121, Feb. 2012.
[27] D. Needell and J. A. Tropp, “CoSaMP: Iterative Signal Recovery from Incomplete and Inaccurate
Samples,” Commun ACM, vol. 53, no. 12, pp. 93
[28] E. van den Berg and M. P. Friedlander,
[29] S. J. Wright, R. D. Nowak, and M. A. T. Figueiredo, “Sparse Reconstruction by Separable
Approximation,” IEEE Trans. Signal
[30] J. M. Bioucas-Dias and M. A. T. Figueiredo, “A New TwIST: Two
Shrinkage/Thresholding Algorithms for Image Restoration,”
no. 12, pp. 2992–3004, Dec. 2007.
[31] W. Sweldens, “Lifting scheme: a new philosophy in biorthogonal wavelet constructions,” in
1995, vol. 2569, pp. 68–79.
[32] T. Acharya and A. K. Ray, “Introduction
1–16.
[33] E. Kofidisi, N.Kolokotronis, A. Vassilarakou, S. Theodoridis, and D. Cavouras, “Wavelet
medical image compression,”
AUTHORS
Sheikh Md. Rabiul Islam received the B.Sc.in Engg. (ECE) from Khulna University,
Khulna, Bangladesh in December 2003, and M.Sc. in Telecommunication Engineering from
the University of Trento, Italy, in October 2009 and currently doing an PhD by Research
under Faculty of Education Science Technology & Mathematics
Australia. He joined as a Lecturer in the department of Electronics and Communication
Engineering of Khulna University of Engineering & Technology, Khulna, in 2004, where
he is joined an Assistant Professor in the same department in the eff
chapter, 25 referred Journal and 15
Wireless communications, signal & image processing, and biomedical e
Professor (Dr) Xu Huang has received the B.E. and M.E. degrees and Ph.D. in Electrical
Engineering and Optical Engineering prior to 1989 and the second Ph.D. in Experimental
Physics in the University of New South Wales, Australia in 1992. He
Graduate Certificate in Higher Education in 2004 at the University of Canberra, Australia.
He has been working on the areas of the telecommunications, cognitive radio, networking
engineering, wireless communications, optical communications,
processing more than 30 years. Currently he is the Professor at the
Science Technology & Mathematics
Society since 1989 and a Fellow of Institution
Engineering (CPEng), a Member of Australian Institute of Physics. He is a member of the Executive
Committee of the Australian and New Zealand Association for Engineering Education, a member of
Committee of the Institution of Engineering Australia at Canberra Branch. Professor Huang is Committee
Panel Member for various IEEE International Conferences such as IEEE IC3PP, IEEE NSS, etc. and he has
published about two hundred papers in high level of t
conference; he has been awarded 9 patents in Australia.
Professor Keng-Liang Ou obtained Ph.D. degree from Graduate Institute of
Engineering, National Chiao Tung University, Taiwan. He came to Taipei Medical
University, Taiwan to pursue cutting edge research for biomaterials and currently holds the
position of elected Dean of College of Oral Medicine. He is also in charge of Graduate
Institute of Biomedical Materials and Tissue Engineering, Research Center for Biomedical
Implants and Microsurgery Devices and Research Center for Biomedical Devices and
Prototyping Production. In addition to institutional appointment, Prof. Ou also serves as the
President of Institute of Plasma Engineering
treatment and the Head of Taiwan Oral Biomedical Engineering Association.
energy to novel research in fields of b
A. Tropp and A. C. Gilbert, “Signal recovery from partial information via Orthogonal Matching
IEEE TRANS Inf. THEORY, 2005.
, I. Drori, and J.-L. Starck, “Sparse Solution of Underdetermined
Linear Equations by Stagewise Orthogonal Matching Pursuit,” IEEE Trans. Inf. Theory
D. Needell and J. A. Tropp, “CoSaMP: Iterative Signal Recovery from Incomplete and Inaccurate
, vol. 53, no. 12, pp. 93–100, Dec. 2010.
E. van den Berg and M. P. Friedlander, Probing the Pareto frontier for basis pursuit solutions
S. J. Wright, R. D. Nowak, and M. A. T. Figueiredo, “Sparse Reconstruction by Separable
IEEE Trans. Signal Process., vol. 57, no. 7, pp. 2479–2493, Jul. 2009.
Dias and M. A. T. Figueiredo, “A New TwIST: Two-Step Iterative
Shrinkage/Thresholding Algorithms for Image Restoration,” IEEE Trans. Image Process.
3004, Dec. 2007.
W. Sweldens, “Lifting scheme: a new philosophy in biorthogonal wavelet constructions,” in
T. Acharya and A. K. Ray, “Introduction in Image Processing," John Wiley & Sons, Inc., 2005,
E. Kofidisi, N.Kolokotronis, A. Vassilarakou, S. Theodoridis, and D. Cavouras, “Wavelet
medical image compression,” Future Gener. Comput. Syst., vol. 15, no. 2, pp. 223–243, Mar. 1999.
received the B.Sc.in Engg. (ECE) from Khulna University,
Khulna, Bangladesh in December 2003, and M.Sc. in Telecommunication Engineering from
the University of Trento, Italy, in October 2009 and currently doing an PhD by Research
on Science Technology & Mathematics at University of Canberra,
Australia. He joined as a Lecturer in the department of Electronics and Communication
Engineering of Khulna University of Engineering & Technology, Khulna, in 2004, where
ant Professor in the same department in the effect of 2008. He has published
Journal and 15 top international conferences. His research interests include VLSI,
Wireless communications, signal & image processing, and biomedical engineering.
has received the B.E. and M.E. degrees and Ph.D. in Electrical
Engineering and Optical Engineering prior to 1989 and the second Ph.D. in Experimental
Physics in the University of New South Wales, Australia in 1992. He has earned the
Graduate Certificate in Higher Education in 2004 at the University of Canberra, Australia.
He has been working on the areas of the telecommunications, cognitive radio, networking
engineering, wireless communications, optical communications, and digital signal
processing more than 30 years. Currently he is the Professor at the Faculty of Education
Science Technology & Mathematics. He has been a senior member of IEEE in Electronics and in Computer
Society since 1989 and a Fellow of Institution of Engineering Australian (FIEAust), Chartered Professional
mmittee of the Institution of Engineering Australia at Canberra Branch. Professor Huang is Committee
published about two hundred papers in high level of the IEEE and other Journals and international
conference; he has been awarded 9 patents in Australia.
obtained Ph.D. degree from Graduate Institute of Mechanical
National Chiao Tung University, Taiwan. He came to Taipei Medical
f Biomedical Materials and Tissue Engineering, Research Center for Biomedical
Institute of Plasma Engineering Taiwan, the leader of the Taiwan society for metal heat
and the Head of Taiwan Oral Biomedical Engineering Association. Professor Ou devotes
energy to novel research in fields of biomaterials, bioengineering, biosensing and bioimaging, as well as
15
A. Tropp and A. C. Gilbert, “Signal recovery from partial information via Orthogonal Matching
Underdetermined Systems of
IEEE Trans. Inf. Theory, vol. 58, no.
D. Needell and J. A. Tropp, “CoSaMP: Iterative Signal Recovery from Incomplete and Inaccurate
Probing the Pareto frontier for basis pursuit solutions. 2008.
S. J. Wright, R. D. Nowak, and M. A. T. Figueiredo, “Sparse Reconstruction by Separable
2493, Jul. 2009.
Step Iterative
IEEE Trans. Image Process., vol. 16,
W. Sweldens, “Lifting scheme: a new philosophy in biorthogonal wavelet constructions,” in SPIE,
," John Wiley & Sons, Inc., 2005, pp.
E. Kofidisi, N.Kolokotronis, A. Vassilarakou, S. Theodoridis, and D. Cavouras, “Wavelet-based
243, Mar. 1999.
ect of 2008. He has published 2 book
nternational conferences. His research interests include VLSI,
. He has been a senior member of IEEE in Electronics and in Computer
of Engineering Australian (FIEAust), Chartered Professional
mmittee of the Institution of Engineering Australia at Canberra Branch. Professor Huang is Committee
he IEEE and other Journals and international
Mechanical
National Chiao Tung University, Taiwan. He came to Taipei Medical
f Biomedical Materials and Tissue Engineering, Research Center for Biomedical
Taiwan, the leader of the Taiwan society for metal heat
Professor Ou devotes
, bioengineering, biosensing and bioimaging, as well as

16
translation medicine. In addition, he has extensive collaborations with industry and has played a major role
in developing medical devices for health professionals worldwide. He is the leader and organizer for the
biomedical product design, production, manufacturing, testing, legalization and market planning, with
supports from team of scientists and researchers with different expertise. These efforts were well
recognized by the community and Professor Ou was awarded with the 49th
Ten Outstanding Young Persons
of Taiwan at the year of 2011.

Image compression based on

More Related Content

What's hot (19)

Similar to Image compression based on (20)

Recently uploaded (20)

Image compression based on