Nested Sparse Classification Method for Hierarchical Information Extraction

Abstract

Use of sparse representation method for classification of visual data has proved its efficiency for challenges encountered in unconstrained face identification. The main identified constrains which heavily affect classification accuracy are variation in expression, pose and lighting. In this paper, a novel classification method is developed as nested sparse classification (NSC) method which incorporates the advantages of sparse representation manifold. In NSC method, sparse representation-based classification is implemented in a nested manner which allows the extraction of hierarchical relationship information between test and training samples. The hierarchical relationship information helps the classifier to be discriminative to inter-personal face changes while robust to intra-personal variations. The implementation not only improves the classification performance but also makes the system capable of being scaled. The improved accuracy of proposed NSC method is demonstrated by experiments carried out on two standard databases (ORL database and YALE database). The performance is analysed with an improvement of more than 2% in terms of classification accuracy.