Genetic Algorithm Optimized Structured Dictionary for Discriminative Block Sparse Representation

This paper presents a few innovations towards learning a discriminative block-structured dictionary. The learning process of such a dictionary is broadly divided into two steps: block formation and dictionary update. In the existing works on block structure estimation, it is assumed that the maximum block size is known a priori. In real-world problems, such an assumption may be sub-optimal. For addressing that, a genetic algorithm optimized K-means clustering based block formation approach is proposed in this work. We also propose a novel dictionary learning approach that incorporates three attributes, namely, reconstruction and discriminative fidelities, block-wise incoherence, and l(2,1)-norm regularization. To further enhance the discriminative ability of the sparse codes, the class-specific and class-common information are modeled separately in the dictionary. The l(2,1)-norm regularization enhances the consistency among sparse codes belonging to the same-class data. In the proposed approach, the dictionary is updated block-wise by employing the singular value decomposition of the composite error matrix obtained through the weighted combination of the component errors. The proposed innovations are evaluated on several public image databases for super-resolution and classification tasks. Along with those image databases, speech based speaker verification task is also evaluated the proposed approach in a few different domains to validate the generalizability. The experimental results obtained on these different databases demonstrate the effectiveness of the proposed approaches when compared with the respective state-of-the-art.