MEDIAN ROBUST EXTENDED LOCAL BINARY PATTERN FOR TEXTURE
CLASSIFICATION
1
Li Liu
∗
,
2
Paul Fieguth,
3
Matti Pietik
¨
ainen,
4
Songyang Lao
1,4
School of Information System and Management, National University of Defense Technology, Changsha, China 410073
2
Department of Systems Design Engineering, University of Waterloo, Waterloo, Canada N2L 3G1
3
Center for Machine Vision Research, Department of Computer Science and Engineering, University of Oulu,90014 Oulu, Finland
Email:lilyliu nudt@163.com, pfieguth@uwaterloo.ca, matti.pietikainen@ee.oulu.fi,feiyunlyi@hotmail.com
ABSTRACT
Local Binary Patterns (LBP) are among the most computation-
ally efficient amongst high-performance texture features. How-
ever, LBP is very sensitive to image noise and is unable to capture
macrostructure information. To best address these disadvantages,
in this paper we introduce a novel descriptor for texture classifica-
tion, the Median Robust Extended Local Binary Pattern (MRELBP).
In contrast to traditional LBP and many LBP variants, MRELBP
compares local image medians instead of raw image intensities.
We develop a multiscale LBP-type descriptor by efficiently com-
paring image medians over a novel sampling scheme, which can
capture both microstructure and macrostructure. A comprehensive
evaluation on benchmark datasets reveals MRELBP’s remarkable
performance (robust to gray scale variations, rotation changes and
noise) relative to state-of-the-art algorithms, but nevertheless at a
low computational cost, producing the best classification scores of
99.82%, 99.38% and 99.77% on three popular Outex test suites.
Furthermore, MRELBP is also shown to be highly robust to im-
age noise including Gaussian noise, Gaussian blur, Salt-and-Pepper
noise and random pixel corruption.
Index Terms— Texture classification, Feature extraction, Local
binary pattern, Local descriptors, Median filtering
1. INTRODUCTION
Texture is an important characteristic of many types of images. It
can be seen in images ranging from multispectral remotely sensed
data to microscopic images. Local Binary Patterns (LBP) [1] have
emerged as one of the most prominent texture descriptors and have
attracted significant attention in the field of computer vision and im-
age analysis due to their outstanding advantages: (1) ease of im-
plementation, (2) invariance to monotonic illumination changes, and
(3) low computational complexity. Although originally proposed for
texture analysis, the LBP method has been successfully applied to di-
verse problems including dynamic texture recognition, remote sens-
ing, fingerprint matching, image retrieval, biomedical image anal-
ysis, face image analysis, motion analysis, and environment model-
ing [2–7]. Due to this progress, the division between texture descrip-
tors and more generic image or video descriptors has been disappear-
ing. Consequently a large number of LBP variants have been devel-
oped to improve its robustness, discriminative power, and breadth of
applicability.
In terms of discriminative power, significant LBP variants in-
clude the Completed Local Binary Pattern (CLBP) [8], Extended
∗
This work has been supported by the National Natural Science Founda-
tion of China under contract No. 61202336 and No. 61201339.
(a) Original Pattern (b) Binary Pattern
(c) Weights (d) Decimal Value
Fig. 1. (a) A typical (r, p) neighborhood type used to derive a LBP
like operator: central pixel x
c
and its p circularly and evenly spaced
neighbors x
0
, · · · , x
p−1
on a circle of radius r .
Local Binary Pattern (ELBP) [6], Discriminative Completed Local
Binary Pattern (disCLBP) [9], Pairwise Rotation Invariant Cooccur-
rence Local Binary Pattern (PRICoLBP) [7] and the combination
of Dominant Local Binary Pattern (DLBP) and Gabor filtering fea-
tures [10]. However, despite the increase in discriminativeness, these
LBP variants suffer in terms of robustness, as they have minimal
tolerance to image blur and noise corruption, as well as increased
computational complexity and feature dimensionality. Similar, with
respect to LBP sensitivity to image degradation caused by blurring
and random noise, significant variations include the Local Ternary
Pattern (LTP) [11], Median Binary Pattern (MBP) [12], Local Phase
Quantization (LPQ) [13], Fuzzy Local Binary Pattern (FLBP) [14],
Noise Tolerant Local Binary Pattern (NTLBP) [15], Robust Local
Binary Pattern (RLBP) [16] and Noise Resistant Local Binary Pat-
tern (NRLBP) [17]. Although more robust than traditional LBP, the
noise tolerance of these methods is still unsatisfactory, and in some
cases the methods compromise discriminative power or impose a
heavy computation burden.
We previously developed the ELBP approach [6], in which four
LBP-like descriptors — Center Intensity based LBP (ELBP CI),
Neighborhood Intensity based LBP (ELBP NI), Radial Differ-
ence based LBP (ELBP RD) and Angular Difference based LBP
(ELBP AD)
1
— were proposed. It was shown that the joint probabil-
ity distribution of ELBP CI, ELBP NI and ELBP RD (collectively
referred as ELBP) produces good texture classification performance,
however with disadvantages of (i) sensitivity to image blur and noise,
(ii) failing to capture texture macrostructure, and (iii) having high
feature dimensionality. In order to overcome these shortcomings,
in this paper we propose a theoretically very simple, high-quality,
yet efficient multiresolution approach, the Median Robust Extended
Local Binary Pattern (MRELBP). The proposed MRELBP approach
possesses some attractive attributes: (1) Gray-scale invariance, (2)
1
In the original work [6], ELBP CI, ELBP NI, ELBP RD and ELBP AD
are referred to as CI-LBP, NI-LBP, RD-LBP and AD-LBP respectively.
978-1-4799-8339-1/15/$31.00 ©2015 IEEE