Is local dominant orientation necessary for the classification of rotation invariant texture? Zhenhua Guo a , Qin Li b,n , Lin Zhang c , Jane You d , David Zhang d , Wenhuang Liu a a Shenzhen Key Laboratory of Broadband Network & Multimedia, Graduate School at Shenzhen, Tsinghua University, Shenzhen, China b Shenzhen Key Laboratory of Sensor Technology, College of Physics Science and Technology, Shenzhen University, Shenzhen, China c School of Software Engineering, Tongji University, Shanghai, China d Department of Computing, the Hong Kong Polytechnic University, Hong Kong, China article info Available online 24 October 2012 Keywords: Texture classification MR8 Image patch Texton Rotation invariance abstract Extracting local rotation invariant features is a popular method for the classification of rotation invariant texture. To address the issue of local rotation invariance, many algorithms based on anisotropic features were proposed. Usually a dominant orientation is found out first, and then anisotropic feature is extracted by this orientation. To validate whether local dominant orientation is necessary for the classification of rotation invariant texture, in this paper, two isotropic statistical texton based methods are proposed. These two methods are the counterparts of two state-of-the-art anisotropic texton based methods: maximum response 8 (MR8) and gray value image patch. Experimental results on three public databases show that local dominant orientation plays an important role when the training set is less; when training samples are enough, local dominant orientation may not be necessary. & 2012 Elsevier B.V. All rights reserved. 1. Introduction Texture analysis is a hot research topic in the fields of computer vision and pattern recognition. It includes four funda- mental problems: classifying texture images based on content [1–4,34,35]; segmenting an image into regions with homoge- neous texture [5]; synthesizing textures for computer graphics [6]; and establishing shape information from texture cue [7]. Among them, texture classification has been widely studied because of its wide range of applications, such as medical image analysis [1], remote sensing [2], surface inspection [3], biometrics [4] and plant image classification [34,35]. In the early stage, extracting statistical features to classify texture images is the main stream. The representative methods include the co-occurrence matrix method [8] and the filtering based methods [9]. Their classification results are good as long as the training and test samples have identical or similar orientations. However, in real situations, the rotations of textures could vary arbitrarily, severely affecting the performance of the statistical methods and raising the classification issue of rotation invariant texture. To address the issue of rotation invariance, many algorithms were proposed. Kashyap and Khotanzad [10] were among the first researchers to study rotation-invariant texture classification by utilizing a circular autoregressive model. After that, many other models were explored, including the multiresolution autoregres- sive model [11], hidden Markov model [12], and Gaussian Markov random field [13]. In general, there are three kinds of methods for the classifica- tion of rotation invariant texture: computing global rotation invariant features [14,15], extracting local rotation invariant features [16–19], and global matching scheme with local rotation variant features [20,21]. The local rotation invariant feature is intuitive and simple, as it processes the image directly and does not require complicated operations, such as thresholding [36] and moment computation [37]. And it could get good results [16–19], especially for small size images [14]. In Ojala et al. [16] proposed to use the local binary pattern (LBP) histogram for the classifica- tion of rotation invariant texture. LBP is a simple but efficient operator to describe local image patterns. Using a group of filter banks, Varma and Zisserman [17] proposed a statistical learning based algorithm, namely maximum response 8 (VZ_MR8), with which a rotation invariant texton library is first built from a training set and then an unknown texture image is classified according to its texton distribution. Later, Varma and Zisserman [18,19] extended their work by proposing a new texton, VZ_Joint, using image patch to represent features directly. Similar to VZ_MR8, an image is classified by its texton distribution. LBP, VZ_MR8 and VZ_Joint are three typical local rotation invariant features, while their underlying local invariance is different: LBP extracts an isotropic feature, as it does not consider any local dominant orientation; VZ_MR8 and VZ_Joint select anisotropic features, as VZ_MR8 defines a dominant orientation from six Contents lists available at SciVerse ScienceDirect journal homepage: www.elsevier.com/locate/neucom Neurocomputing 0925-2312/$ - see front matter & 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.neucom.2011.11.038 n Corresponding author. Tel.: þ86 755 26941509; fax: þ86 755 26941509. E-mail addresses: zhenhua.guo@sz.tsinghua.edu.cn (Z. Guo), csyjia@comp.polyu.edu.hk (J. You), Kenneth_lee_qin@gmail.com (Q. Li). Neurocomputing 116 (2013) 182–191