1 LBP Based Edge-Texture Features for Object Recognition Amit Satpathy, Member, IEEE, Xudong Jiang, Senior Member, IEEE, and How-Lung Eng, Member, IEEE Abstract—This paper proposes two sets of novel edge-texture features, Discriminative Robust Local Binary Pattern (DRLBP) and Ternary Pattern (DRLTP), for object recognition. By in- vestigating the limitations of Local Binary Pattern (LBP), Local Ternary Pattern (LTP) and Robust LBP (RLBP), DRLBP and DRLTP are proposed as new features. They solve the problem of discrimination between a bright object against a dark back- ground and vice-versa inherent in LBP and LTP. DRLBP also resolves the problem of RLBP whereby LBP codes and their complements in the same block are mapped to the same code. Furthermore, the proposed features retain contrast information necessary for proper representation of object contours that LBP, LTP and RLBP discard. Our proposed features are tested on 7 challenging data sets - INRIA Human, Caltech Pedestrian, UIUC Car, Caltech 101, Caltech 256, Brodatz and KTH-TIPS2- a. Results demonstrate that the proposed features outperform the compared approaches on most data sets. Index Terms—object recognition, local binary pattern, local ternary pattern, feature extraction, texture. I. I NTRODUCTION C ATEGORY recognition and detection are 2 parts of object recognition. The objective of category recognition is to classify an object into one of several predefined cate- gories. The goal of detection is to distinguish objects from the background. There are various object recognition challenges. Typically, objects have to be detected against cluttered, noisy backgrounds and other objects under different illumination and contrast environments. Proper feature representation is a crucial step in an object recognition system as it improves per- formance by discriminating the object from the background or other objects in different lightings and scenarios. Furthermore, a good feature also simplifies the classification framework. Object recognition features are categorized into two groups - sparse and dense representations [7]. For sparse feature representations, interest-point detectors are used to identify structures such as corners and blobs on the object. A feature is created for the image patch around each point. Popular feature representations include Scale-Invariant Feature Trans- form (SIFT) [16], [30], Speeded Up Robust Feature [3], Local Steering Kernel [49], Principal Curvature-Based Regions [9], Copyright (c) 2013 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to pubs-permissions@ieee.org. A. Satpathy is with Institute for Infocomm Research, Agency for Science, Technology & Research, 1 Fusionopolis Way, Connexis, Singapore 138632. E-mail: satpathya@i2r.a-star.edu.sg X. Jiang is with the School of Electrical and Electronics Engineering, Nanyang Technological University, Nanyang Link, Singapore 639798. E-mail: exdjiang@ntu.edu.sg H.-L. Eng is with Zweec Analytics, 67 Ayer Rajah Crescent, 03-23/24, Singapore 139950. E-mail: howlungeng@zweec.com Region Self-Similarity features [33], [50], Sparse Color [51] and the sparse parts-based representation [1]. A comprehensive evaluation of sparse features can be found in [34], [35]. Dense feature representations, which are extracted at fixed locations densely in a detection window, are gaining popularity as they describe objects richly compared to sparse feature rep- resentations. Various feature representations such as Wavelet [40], Haar-like features [55], Histogram of Oriented Gradients (HOG) [8], [56], Extended Histogram of Gradients [44]– [46], [48], Feature Context [57], Local Binary Pattern (LBP) [2], [22], [47], Local Ternary Pattern (LTP) [52], Geometric- blur [59] and Local Edge Orientation Histograms [25] have been proposed over recent years. Dense SIFT has also been proposed to alleviate the sparse representation problems [4], [24], [53]. LBP is the most popular texture classification feature [18], [20], [21], [27], [38], [41], [42], [62]. It has also shown excellent face detection performance [2], [19], [26], [52], [61]. It is robust to illumination and contrast variations as it only considers the signs of the pixel differences. Histogramming LBP codes makes the descriptor resistant to translations within the histogramming neighbourhood. However, it is sensitive to noise and small fluctuations of pixel values. To handle this, Local Ternary Pattern (LTP) has been proposed [52]. In comparison to LBP, it has 2 thresholds which creates 3 different states as compared to 2 in LBP. It is more resistant to noise and small pixel value variations compared to LBP. Like LBP, it has also been used for texture classification and face detection [13], [23], [28], [42], [52], [60]. However, for object recognition, LBP and LTP present two issues. They differentiate a bright object against a dark back- ground and vice versa. This increases the object intra-class variations which is undesirable for most object recognitions. Nguyen et al. [37] propose Robust LBP (RLBP) to map a LBP code and its complement to the minimum of both to solve the problem. However, in the same block, RLBP also maps to the same value. For some different local structures, a similar feature is obtained. Hence, it is unable to differentiate them. Different objects have different shapes and textures. It is therefore desirable to represent objects using both texture and edge information. However, in order to be robust to illumination and contrast variations, LBP,LTP and RLBP do not differentiate between a weak contrast local pattern and a similar strong one. They only capture texture infor- mation. Object contours, which also contain discriminatory information, tend to be situated in strong contrast regions. Therefore, by discarding contrast information, contours may not be effectively represented.