Pose and Illumination Invariant Face Recognition in Video via Motion Analysis and Texture Feature Optimization

ABSTRACT Face recognition and person re‐identification in video remains a challenging area in computer vision community. A variety of texture feature sets are evaluated for better recognition of individuals in video for varying pose and illumination conditions—as found in real world video surveillance. This paper focuses on face recognition in video based changing environment using Fuzzy‐ARTMAP (FAM) neural network classifier. Individual specific facial model of each target subject is structured using two‐class classification to solve a multi‐class classification problem. Two lower dimensional linear subspace techniques, PCA and LDA, are applied. All evaluations are carried out by executing 10 iterations and average results are presented. We split dataset randomly in each permutation: two‐thirds of the data model is reserved for training of the classifier, and rest for testing. Experimental results of our study are determined on two publicly available databases: the ChokePoint database and the Extended YaleB database. Our findings show that FAM classifier performs better with the feature set “Pixel intensity combined with Local Binary Patterns (LBP) and Local Phase Quantization (LPQ)” in varying illumination and pose scenarios during face recognition problem.