IEEE SIGNAL PROCESSING LETTERS, VOL. 20, NO. 4, APRIL 2013 319
Edge Strength Similarity for
Image Quality Assessment
Xuande Zhang, Xiangchu Feng, Weiwei Wang, and Wufeng Xue
Abstract—The objective image quality assessment aims to model
the perceptual fidelity of semantic information betw een two im-
ages. In this letter, we assume that the semantic information of im-
ages is fully represented by edge-strength of each pixel and propose
an edge-strength-similarity-based image quality metric (ESSIM).
Through investigating the characteristics of the edge in images, we
define the edge-strength to take both anisotropic regularity and ir-
regularity of the edge into account. The proposed ESSIM is consid-
erably simple, however, it can achieve slightly better performance
than the state-of-the-art image quality metrics as evaluated on six
subject-rated image databases.
Index Terms—Edge-strength, image quality assessment, regu-
larity and irregularity.
I. INTRODUCTION
I
N recent years, there has been extensive interest in devel-
oping objective image quality assessment (IQA) metrics.
Such metrics have wide applications in the field of computer vi-
sion and image processing [1]. According to the availability of
the ground-truth images in the assessment process, they can be
classified into full-reference (FR), reduced-reference (RR) and
no-reference (NR) metrics. In this letter, the discussion is con-
fined to FR m etrics, w here the ground-truth images are available
and known as the reference images.
The L2-fidelity related IQA metrics, such as signal to noise
ratio (SNR ) and peak SNR (PSNR), are most w idely used in
image processing. But these metrics do not consider the proper-
ties of human visual system (HVS) and show poor consistency
with subjective evaluations. Si nce the introduction of structural
similarity index (SSIM) [ 1], a fl ood of IQA metrics have been
developed with attempt to achieve high consistency wi th sub-
jective assessment [2]–[7]. The SSIM assumed that the HVS
is highly a dapted for extracting structural infor mation from a
Manuscript received October 11, 2012; revised December 14, 2012; ac-
cepted January 15, 2013. Date of publication January 30, 2013; date of current
version February 1 5, 2013. This work was supported by the National S cience
Foundation of China (Grants 61001156, 61105011, 11101292, 60872138 and
61271294) and by the N ational Science Foundation of Ningxia University
(ZR1206). The associate editor coordinating the review of this manuscript and
approving it for publication was Prof. Weisi Lin.
X. Zh ang is with Departmen t of Applied Mathema tics, School of Science,
Xidian University, Xi’an , China and also w ith School of Mathem atics and Com-
puter Scie nce, Ningxia University, Yinchuan, China (e-m ail: love_truth@ 126.
com).
X. Feng and W. Wang are with Department of Applied Mathematics,
School of Scienc e , Xidian University, X i’an, China (e-mail: xcfeng@ma il.x i -
dian.edu.cn; www ang@mail.xidian.edu.cn).
W. Xue is with Institute of Imag e Processing and Pattern Recognition, Xi’an
Jiaotong University, Xi’an, China (e-mail: x.wolfs@stu.xjt.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Dig
ital Object Identifier 10.1109/LSP.2013.2244081
scene and the local structure similarity reflects the visual fidelity
well.In[6],Zhang.et al. pr oposed a feature similarity index
for image qual ity assessment (F SIM), this m etric evaluated the
visual fide lity in the both phase congruency and gradient fea-
ture space with the a ssum pt ion that the HVS perceives images
according to low-level features. A thorough com parison of the
performance between FSIM and other representative IQA met-
rics are presented in [7], which indicated that FSIM achieves the
best assessm ent performance so far.
An ideal IQA metric should p erfectly mimic the HVS. Un-
fortunately, the HVS itself is not w ell understood until now.
As a consequence, m ost existing IQA metrics are on ly designed
on the basis of certain assumption about the HVS. An apparent
fact is that human semantically perceive the image and evaluate
the quality of the imag e based on th e distortion of sem a ntic in-
formation. Consider the simple example of locally adding t he
same amount of noise to a face image, it is obvious that the
manipulating regio n is deterministic for the quality of the re-
sulting image. If the hair region is contaminated, the resulting
face image still looks quite perf ect. But, if the semantically sig-
nificant region, such as that of eyes, nose or the lips, is con-
taminated, the resulting image would look very unpleasant and
would receive a low subjective quality score. This exam ple ver-
ifies that the human rates the image quality according to the dis-
tortion of semantic information. However, accurately measuring
the semantic i nformation rem ains a very challeng ing task.
In this letter, we assume that the semantic informat ion of
images is mainly represented by t he edge-strength of each
pixel, the edge-strength quantifies the ch ance or possibility
of a pixel belonging to the edg e of a semantic object. We
will analyze the characteristics of semantic edge and define
the edge-strength to fo llo w these characteristics, and then we
propose an image q uality metric b ased on the edge-strength
similarity. The proposed metric is conceptually very simple
and can be implemented with very low complexity. However,
it leads t o promising assessm e nt performance as evaluated on
the p ublically available data sets.
II. E
DGE-STRENGTH SIMILARITY BASED
IMAGE QUALITY METRIC
A. Three Characteristics of the Edge and the Definition of
Edge-Strength
The edg
e is a fund amen tal concept in computer vision, partic-
ularly
in the area of object extraction and image segmentation.
Ideal
ly, t he ed ge is a s et of pixels indicating the bound ary of se-
mant
ic objects in the image. In practice, it is always identified by
coll
ecting the set o f pixels at which the image changes sharply,
or h
as discontinuity. However, discontinuity is only o ne side of
1070-9908/$31.00 © 2013 IEEE