978-1-5090-5316-2/16/$31.00 ©2016 IEEE VCIP 2016, Nov. 27 - 30, 2016, Chengdu, China
Perceptual Sparse Representation for Compressed
Sensing of Image
Jian Wu
2
, Yongfang Wang
1,2
Kanghua Zhu
2
, Yun Zhu
2
1
Key Laboratory of Advanced Display and System Application, Ministry of Education, Shanghai, 200072, China
2
School of Communication and information Engineering, Shanghai University, Shanghai, 200072, China
Abstract—Most traditional image coding schemes based on
compressed sensing exploited the sparse domain in fixed bases
and less consider the image non-stationary characteristic and
human visual characteristic, which leads to poor performance of
the reconstruction. In this paper, we proposed a novel sparse CS
scheme combined with just-noticeable difference (JND) Model
and random permutation. Firstly, the DCT-based JND profile
has been utilized to remove the perceptual redundancies which
also makes the signal sparser, then the random permutation is
adopted to balance the sparsity of each block in image.
Experimental results show that the proposed perceptual sparse
algorithm outperforms some existing approaches, and it can
achieve better subjective and objective image quality compared
to other algorithms when the sampling rate is above 0.3.
Index Terms
—
Compressed Sensing, Random Permutation,
Just-noticeable Distortion, Sparsity, Discrete Cosine Transform
I. I
NTRODUCTION
Compressed sensing (CS) [1] has been proposed by
Donoho, Candes and Tao in 2006, which has already been
widely used in many application areas [2] such as signal
processing, pattern recognition, communication and etc. As a
new signal processing theory, CS can acquire a signal at a
sampling rate much lower than Nyquist rate via linear
projection onto a random basis, and the original signal can be
reconstructed through optimization method with high
probability from some random measurements under certain
conditions. CS has a great potential in video and image
applications for its low complexity and low power
consumption in digital data acquisitions, and less sensitive to
the channel error. Due to its great practical potentials, it has
stirred great excitements both in academic and industries in
past few years. However, CS faces several challenges
including a computationally expensive reconstruction process
and huge memory required to store the random sampling
operator.
CS based image/video coding scheme decreases the burden
of capturing and encoding at the encoder side, while the
decoder tends to be rather computationally complex. In many
cases, we cannot apply the CS into image compressing
directly, due to a large computational cost associated with
multidimensional signal reconstruction and a huge memory
burden. To address these problems, the block-based CS (BCS)
sampling of images coupled with iterative smooth projected
Landweber (BCS-SPL) [3] is proposed to reduce the size of
measurement matrix by dividing the image into some non-
overlapped blocks and speed up the process of reconstruction,
which achieves a more efficient reconstruction than full-image
CS sampling. However, the method mentioned above adopted
the same measurement rate for all the blocks, which ignores
the fact that images are typically non-stationary signals and
blocks may have different characteristics. Zheng et al. [4] and
Liran et al. [5] proposed the idea that allocate different
sampling rate for different blocks according to the blocks’
characteristics. In [4], authors proposed to assign a sampling
rate depending on texture complexity. In [5], authors use the
CS measurements acquired at the collector to estimate the
sample variance of each block directly and then adaptively
assigns a sampling rate of each block in terms of their sample
variances. In [6], the DCT coefficients random permutation
(CRP) method is proposed to make the sparsity of every block
to tend to be same, so that the non-adaptive projection
representation for the natural images can lead to good
performance.
In [1], authors presented the idea that the more the signal
sparse, the less measurement required, and the sparsity of the
natural image may not be enough in some domains. In this
paper, we proposed a novel sparse representation method for
image consisting of preprocessing based on JND and random
permutation to exploit the sparsity. In preprocessing based on
JND, we consider the human visual characteristics in CS,
which not only removes the perceptual redundancies to
improve the compression ratio but also makes the signal
sparser to achieve better reconstruction image quality. And
then we extend CRP method [6] by permutated the DCT
coefficients randomly to balance the sparsity of each block
after preprocessing for increasing the BCS sampling
efficiency.
The rest of the paper is organized as follows. Section II
describes the brief idea about CS. Section III proposed CS
scheme based on JND model and random permutation.
Experiments and results are given in Section IV, and we
conclude this paper in Section V.
II. C
OMPRESSED
S
ENSING
In this section, we briefly review the CS principles for
signal acquisition and recovery. CS allows the signal to be
acquired via linear projection into a measurement vector
whose dimension is much lower than the origin. A signal
with length
can be called
-sparse in a domain
, if there
are
entries in its transform coefficients
T
θ = Ψ
are nonzero.