Академический Документы
Профессиональный Документы
Культура Документы
Sumit Budhiraja
I.
INTRODUCTION
PREVIOUS WORK
1073
isi
(3)
A. Sparse Represenatation
Sparsity intends to convey that the original image shows its
dense form in a particular basis, but after transformation into
other convenient domain, the coefficients of image offer a brief
summary. There are many signals that possess concise
representations when expressed in a suitable basis. In almost
every case, the coefficients of the transform (usually DCT,
Wavelet etc.) of an image provided are small and the transform
domain is sparse. Suppose we have a vector
RN (such as Ndimensional image) which can be expanded in an orthonormal
basis for example wavelet transform domain which can be given
as = [ 1, 2,........ N] as follows:
(2)
(1)
(4)
(5)
Where
1, 2,....... N} is the measurement matrix of
size
, where << . Measurement matrix multiply x to
acquire only the required information from , so each column
of can be seen as a sensing column. Here, is the sensing
matrix that senses i.e. samples the compressed data at single
step [1].
In order to recover an image with compressive sensing there
are various techniques based on Basis pursuit (BP) which works
on the principle of norm and it provides valid results for
reconstruction of image but it exhibit slower running time for
the algorithm. The one of the popular recovery algorithm is
OMP which is quite faster than BP algorithm but also having
computational complexity.With the help of certain methods if
complexity can be lessen, then the OMP algorithm is far more
productive than BP in terms of total elapsed time and the
accurate recovery [4]. In previous research work of Rauhut
which shows the method of the OMP algorithm [15] and its
implementation for reconstructing the image from undersampled
1074
r||}1
(7)
M=max
The condition M<1 is called the Exact Recovery Condition
The condition M<1 is called the Exact Recovery Condition
(ERC) [20]. This condition is enough for the exact recovery of
the image. These conditions guarantee the recovery of the
selected set with higher success rate.
B. Modified OMP Algorithm
In this paper, a modified OMP is presented in recovering
the sparse signal based on the lesser number of linear
measurements, considers the following model as:
y= x +
(8)
where can be random noise vector like i.i.d Gaussian noise
vector. The sparse image is here recovered in the presence of
noise under the implication of stopping rules and their
property. The modified algorithm is shown as follows:
1) Consider an N x N image, Select suitable value of M
and construct the Measurement matrix (M x N).
Disintegrate the image into low frequency sub-band having Li
(i=1,2.....N), and high frequency coefficients Hi, Vi, Di
(i=1,2.....N).
2) The next step is to obtain the sparse form of the low
frequency coefficients only using CS. Recover the low
frequency coefficients using the modified OMP algorithm
under the certain stopping condition and its properties.
3) Set the residual r0 = y and initialize the set of
selected variables = . Suppose the iteration count k=1. The
other parameters can be specified as as Measurement Matrix
(M x N), x as the input image (N x N),
(N x N) as the
transform matrix.
4) The algorithm progresses further if it satisfies these
two conditions at each iteration step:
a) Mutual incoherence for should be a small value
(
1)
b) Estimated Residual > zero_threshold where
zero_threshold is the specified complex integer to ensure the
selection of significant columns only and iteration count
should be k < min {K, M/N}
5) Select the indices {
}i=1,2,3...N corresponding to N
k-1
.
largest entries in
6) Augment the set of selected variables: k = k-1
{ 1 ,..
}. Then, solve a least squares problem to
acquire a new image estimate:
(9)
k
2
7) Update the residual to recover the image:
(10)
Check if, M <1, then retrieve the recovered image:
(11)
2
set
1, go to step 5, otherwise stop the algorithm.
1075
SIMULATION RESULTS
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
Fig.2 Reconstruction result of Modified OMP for image Lena (a) Original
image (b) Image reconstructed with M=170 (c) Image reconstructed
with M= 190 (d) Image reconstructed with M=220
(a)
(b)
(c)
(d)
(e)
(f)
The results for PSNR and the comparative analysis are given
in Table I. The value of PSNR is compared for reconstructed
images at several measurements M x N (N=256). The quality
of image improves with increase in number of measurements,
M. Among these techniques, Modified OMP provides best
PSNR. The Modified algorithm ensures the selection of the
significant columns, thus eliminating the non-significant ones
due to certain stopping conditions. However, the recovery
performance becomes comparable for higher measurements.
Table II compares the running time of different algorithms
comparisons with that of Modified Algorithm. The table
shows that as the sampling rate (M/N) increases, the running
time increases.
1076
(b)
(c)
(d)
(e)
M=128
M=150
M=170
M=190
Subspace Pursuit
25.76
26.49
26.90
27.07
OMP
26.44
28.23
30.72
32.63
25.33
26.45
28.04
ROMP
21.84
25.59
27.36
28.26
Modified OMP
32.09
31.87
32.09
33.67
[1]
[2]
PSNR(db)
REFERENCES
[3]
[4]
[5]
[6]
M=128
M=150
M=170
M=190
Subspace Pursuit
52.38
58.20
74.92
77.39
OMP
4.64
5.02
5.26
5.29
Modified OMP
3.91
4.24
4.30
4.43
[7]
[8]
[9]
CONCLUSION
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
1077
1078