Effective Facial Feature Representation Based on Directional Micro-Patterns

Abstract

Automatic recognition and analysis of human face allow many interesting appli- cations in biometrics, human-computer interaction, and security industry, such as person identification, age estimation, gender classification, and facial expres- sion analysis. Deriving an efficient and effective feature representation is the fundamental component for any successful facial recognition system. However, the inherent variability of facial images caused by different factors like variations in illumination, pose, facial expression, alignment, and occlusions make classifi- cation a challenging task. Therefore, the aim of the ongoing research in facial recognition is to increase the robustness of the underlying feature representation against these factors. In this thesis, we work toward developing a simple, yet effective appearance- based feature descriptor for representing human facial image. A new local texture pattern, namely the directional ternary pattern (DTP) has been introduced that can effectively capture the texture properties of a local neighborhood and exhibits robustness against illumination variations and random noise. The proposed DTP operator encodes the texture information of a local neighborhood by labeling the edge response values in all eight directions around the center point using three different levels. Our encoding scheme employs a threshold in order to differentiate between uniform and high-textured face regions, and thus, ensures the generation of ternary micro-patterns consistent with the local texture property. The location and occurrence information of the DTP micro-patterns within the facial image is then represented with a spatial histogram, which functions as the facial feature descriptor. i

Abstract ii The performance of the proposed method has been evaluated in three differ- ent applications, which are i) facial expression analysis, ii) face recognition, and iii) gender classification from facial image. Different publicly-available bench- mark image datasets were used for the experiments. We have also compared the performance of the proposed method with some widely-used local pattern-based feature descriptors. Experimental results demonstrate that, the DTP descriptor is more robust in extracting facial information and provides higher classification rate compared to some existing feature representation techniques.