A Complete Dual-Cross Pattern for Unconstrained Texture Classification
Swalpa K. Roy, Bhabatosh Chanda,
Bidyut B. Chaudhari
Indian Statistical Institute, Kolkata
swalpa@ieee.org, {chanda,bbc}@isical.ac.in
Dipak K. Ghosh
NIT Rourkela, Orissa
dipak@ieee.org
Shiv Ram Dubey
IIIT Chittoor, Sri City
srdubey@iiits.in
Abstract
In order to perform unconstrained texture classifica-
tion this paper presents a novel and computationally effi-
cient texture descriptor called Complete Dual-Cross Pat-
tern (CDCP), which is robust to gray-scale changes and sur-
face rotation. To extract CDCP, at first a gray scale normal-
ization scheme is used to reduce the illumination effect and,
then CDCP feature is computed from holistic and compo-
nent levels. A local region of the texture image is repre-
sented by it’s center pixel and difference of sign-magnitude
transform (DSMT) at multiple levels. Using a global thresh-
old, the gray value of center pixel is converted into a binary
code named DCP center (DCP C). DSMT decomposes into
two complementary components: the sign and the magni-
tude. They are encoded respectively into DCP-sign (DCP S)
and DCP-magnitude (DCP M), based on their correspond-
ing threshold values. Finally, CDCP is formed by fusing
DCP S, DCP M and DCP C features through joint distribu-
tion. The invariance characteristics of CDCP are attained
due to computation of pattern at multiple levels, which
makes CDCP highly discriminative and achieves state-of-
the-art performance for rotation invariant texture classifi-
cation.
1. Introduction
Texture classification is one of the active research topics
due to scientific challenges and potential use in a wide range
of practical applications such as medical image analysis,
remote sensing, fabric inspection, segmentation, content-
based image retrieval [18], and iris based biometric recog-
nition [19]. In the past satisfactory performance has been
obtained by various techniques only in controlled environ-
ment. However, classification of unconstrained texture im-
age is a crucial problem due to wide variation of view-
points, illumination changes and degraded quality of tex-
ture image. Therefore, the design of efficient descriptor is
a fundamental problem in texture image classification. Ba-
sically, texture representation can be categorized in terms
of the employed approaches, e.g. geometrical, structural,
model-based, statistical, and signal processing. Earlier tex-
ture classification methods focus on the statistical analy-
sis of texture images which include the co-occurrence ma-
trix based approach [7] and filtering based techniques [16].
These methods provide good classification performance as
long as both training and test sample images have identi-
cal orientations. However, arbitrary rotations which could
occur in a real-world scene, affect the performance of the
methods. Thus, rotation invariance is a crucial issue to
be addressed and attention has been focused on the design
of geometrically and photometrically invariant local texture
representation [24, 15, 21]. At first Kashyap and Khotan-
zad proposed circular autoregressive dense approach [8] for
the rotation invariance texture classification. Earlier, many
other models have been explored for rotation invariance
classification, including multi-resolution, hidden Markov
model, and Gaussian Markov model. Recently Varma and
Zisserman proposed VZ-MR8 [20] and VZ-Patch [21] to
learn a texton dictionary from a set of training images,
which are rotation and scale invariant and then classified
the unknown sample images using learned texton distribu-
tions. The downside of these methods are feature extraction
and matching complexity which is not favourably good.
In order to gain more robustness, feature extraction is
often performed over local region of the image. In 1996 a
simple and computationally efficient texture representation,
called local binary pattern (LBP) was proposed by Ojala et
al. [14] for gray scale and rotation texture classification.
Other variants of LBP such as DLBP [9], LBP variance
(LBPV)[6], completed LBP (CLBP)[5], local derivative
pattern (LDP)[23], local wavelet pattern LWP [2], local di-
rectional derivative pattern (LDDP)[4], local ternary pattern
(LTP)[17] etc. were proposed due to numerous application
of LBP in the field of computer vision and pattern recogni-
tion such as texture segmentation, face recognition, shape
localization and object recognition [11]. However, most de-
scriptors are based on the same basic idea of LBP and ex-
tracts only circular isotropic micro structure of the texture
image at one level, which is not enough to describe the tex-
2017 4th IAPR Asian Conference on Pattern Recognition
2327-0985/17 $31.00 © 2017 IEEE
DOI 10.1109/ACPR.2017.160
741