MATCHING TEXTURE UNITS FOR FACE RECOGNITION Bangpeng YAO, Haizhou AI Computer Science and Technology Department, Tsinghua University, Beijing 100084, China Shihong LAO Sensing and Control Technology Laboratory, Omron Corporation, Kyoto 619-0283, Japan ABSTRACT For an image, texture unit (TU) is a small complete unit which characterizes the local texture of a given pixel and its neigh- borhood. Recently, TU-based approaches have been widely used in face recognition. This paper proposes a novel face representation and recognition approach based on TU. We make three major contributions: (1) we introduce a novel TU feature, Local Gabor Quarternary Pattern (LGQP), which in- corporates both Gabor magnitude and phase information in a single TU code; (2) similarity measure of two TU images is treated as a tracking problem between two images, and we present a novel point-to-point matching (PPM) approach for TU similarity measure; (3) based on an integral histogram technique, the PPM similarity can be computed very rapidly. Experimental results on CMU-PIE and FERET data sets show that our method is able to reach very promising results. Index Terms— Texture Unit, Local Gabor Quarternary Pattern, Point-to-Point Matching, Face Recognition 1. INTRODUCTION Due to its wide range of commercial and law enforcement ap- plications, face recognition received various attentions during the past decade. Although great advances have been achieved, recognition of face images with large variations of pose, illu- mination and expression is still a challenging task. There are mainly two approaches to represent face im- ages. One is the holistic methods such as PCA and LDA. The other is local features, such as Gabor wavelet and texture unit (TU) [2]. TU is a small unit which characterizes the local texture of a given pixel and its neighborhood. Recently TU features have drawn increasing attention in face recognition because they can capture small appearance details as well as describing the overall face structure. Methods in this category include Local Binary Pattern (LBP) [1], Local Gabor Texton (LGT) [3], Local Gabor Binary Pattern (LGBP) [11], etc. TU face similarity is usually measured by histogram op- erations, which divide the face image into many grid of cells and compute the histogram similarity in each cell. Such meth- ods, although have been widely used in face recognition, in- evitably cause some loss of discriminative power, because po- sition within each cell is not coded. In [8], a Distance Trans- form (DT) method [8] was presented to alleviate this problem. Given two images X and Y , DT first generates a DT image for each TU code in X, and then uses these DT images to match each code in Y . Pixel deviations between the two im- ages are penalized with Gaussian metric or truncated linear distance. Although DT takes more advantage of spatial in- formation, it is still sensitive to image spatial deviations due to misalignment or pose variations. For example, if Y is ob- tained by slightly shifting X, the DT similarity of X and Y will be greatly decreased because of the penalty function. In this paper, we propose a novel Point-to-Point Match- ing (PPM) method to measure TU based face similarity. The PPM metric can consider detailed face spatial information and is also robust to image spatial deviations. It is based on the idea that, two TU images can be regarded as two consecutive frames in an image sequence, and thus the approaches in ob- ject tracking, especially correspondence based point tracking [9] can be used for TU similarity measure. Moreover, with integral histogram [5] for an intermediate face representation, PPM similarity can be calculated fast. Furthermore in this paper, we introduce a novel TU code, Local Gabor Quarternary Pattern (LGQP). LGQP combines the discriminative information in both Gabor magnitude and phase in one code, and therefore is more powerful than the other Gabor filter based face representations that separately treat magnitude and phase. In the remaining part of this paper, we will firstly intro- duce the LGQP feature and PPM method in Section 2 and 3 respectively, and then give experimental results in Section 4. 2. LOCAL GABOR QUARTERNARY PATTERN Gabor filter is one of the most successful features in face rep- resentation. There are mainly two ways to use it. One is to use only Gabor magnitude features [3] or Gabor phase fea- tures [10], the other is to concatenate magnitude and phase features, such as the Extended LGBP in [11], which shows that the discriminative information in Gabor magnitude and that in Gabor phase are complementary to each other. Rather than using the above two approaches that sepa- rately treat magnitude and phase features, in this work we introduce a new TU feature, Local Gabor Quarternary Pat- tern (LGQP), which makes use of both magnitude and phase 1920 978-1-4244-1764-3/08/$25.00 ©2008 IEEE ICIP 2008