Completed robust local binary pattern for texture classification

Abstract

Original Local Binary Pattern (LBP) descriptor has two obvious demerits, i.e., it is sensitive to noise, and sometimes it tends to characterize different structural patterns with the same binary code which will reduce its discriminability inevitably. In order to overcome these two demerits, this paper proposes a robust framework of LBP, named Completed Robust Local Binary Pattern (CRLBP), in which the value of each center pixel in a 3×3 local area is replaced by its average local gray level. Compared to the center gray value, average local gray level is more robust to noise and illumination variants. To make CRLBP more robust and stable, Weighted Local Gray Level (WLG) is introduced to take place of the traditional gray value of the center pixel. The experimental results obtained from four representative texture databases show that the proposed method is robust to noise and can achieve impressive classification accuracy.

Introduction

Texture classification plays an important role in computer vision and image processing. In the past decades, numerous algorithms for texture feature extraction have been proposed, many of which focus on extracting texture features that are robust to noises, rotation and illumination variants [1]. Davis [2] exploited polarograms and generalized co-occurrence matrices to obtain rotation invariant statistical features. Duvernoy [3] proposed Fourier descriptors to extract texture feature on the spectrum domain. Goyal et al. [4] proposed a method by using texel property histogram. Eichmann and Kasparis [5] presented texture descriptors based on line structures extracted by Hough transform. Kashyap and Khotanzad [6] developed a circular simultaneous autoregressive (CSAR) model for rotation invariant texture classification. Cohen et al. [7] characterized texture as Gaussian Markov random fields and used the maximum likelihood to estimate rotation angles. Chen and Kundu [8] addressed rotation invariant by using multichannel sub-bands decomposition and hidden Markov model (HMM). Porter and Canagarajah [9] exploited the wavelet transform for texture classification by using the Daubechies four-tap wavelet filter coefficients. Recently, Varma and Zisserman [19], [21], [22] proposed to cluster a rotation invariant texton dictionary from a training set, and then form the textural histogram based on these textons. Later, Xu et al. [23], [24], [25] presented scale invariant texture classification methods by using a multi-fractal spectrum (MFS).

In [10], Ojala et al. proposed to use the Local Binary Pattern (LBP) for rotation invariant texture classification. As shown in Fig. 1, LBP code is computed by comparing a pixel with its neighbors. After the LBP code of each pixel in the image is defined, a histogram will be built to represent the texture image. LBP is a simple yet efficient operator to describe local texture, and has been proven to be invariant to monotonic gray scale transformations.

Since Ojala's work, a lot of variants of LBP have been proposed. For example, Heikkila et al. [11] proposed center-symmetric LBP (CS-LBP) by comparing center-symmetric pairs of pixels instead of comparing neighbors with central pixels. Liao et al. [12] presented Dominant LBP (DLBP), in which dominant patterns were experimentally chosen from all the patterns. Tan and Triggs [13] proposed Local Ternary Pattern (LTP), which extends original LBP to 3-valued codes. Guo et al. [14] proposed completed LBP (CLBP) by combining the conventional LBP with the measures of local intensity difference and central gray level. Recently, Khellah [15] presented a new method for texture classification, which combines Dominant Neighborhood Structure (DNS) and traditional LBP.

Although LBP and its variants have achieved impressive classification results on representative texture databases, there still remain some potential flaws of LBP. For example, LBP is sensitive to noise, and often classifies many different patterns into a same class. This paper attempts to solve these potential difficulties by proposing a robust framework of LBP, named Completed Robust Local Binary Pattern (CRLBP). In CRLBP, the value of each center pixel in a 3×3 local area is replaced by its average local gray level. Compared to gray value, average local gray level is more robust to noise and illumination variants. Experimental results illustrate that CRLBP achieves higher classification rates than other variants of LBP, and is insensitive to noise and illumination variants.

The rest of this paper is organized as follows: Section 2 briefly introduces two main flaws of LBP andtwo improved versions of LBP, i.e., LTP and CLBP. Section 3 presents the framework of CRLBP. Experimental results are reported in Sections 4 and Section 5 concludes the whole paper.