Sparse Representation And Clustered Dictionary Learning Based Face Recognition

Face recognition is a hot research topic in the pattern recognition and computer vision; it is widely used in various applications such as military, remote intelligent monitoring, public security and human-computer interaction. Recently, Sparse Representation-based Classification(SRC) has attracted more and more researchers’ interest due to its simplicity and effectiveness. Moreover, it has also been successfully applied to a variety of problems in image processing and computer vision. This thesis puts forward three enhanced recognition algorithms based on the analysis and summary of the relevant researches about SRC.1. This paper propose the Locality-constrained Group Kernel Sparse Representation basedClassification algorithm(LG-KSRC). This approach can select neighboring training samples from the same class with the test samples to reconstruct the test samples by introduce a locality-constrained term and group-sparse term into original SRC, such that the data locality and group sparsity can be better preserved. In order to further enhance the discriminability of the algorithm, it is also extended to nonlinear space. Experimental results verify the effectiveness of the algorithm.2. This paper propose the Half Local Binary Pattern Kernel Sparse Representation based Classification algorithm(HLBP-KSRC). First of all, we integrate a kernel coordinate descent into the SRC framework, which can generate a novel kernel SRC framework(KCD-SRC). And then, we employ local image features and develop a HLBP-Hamming kernel for KCD-SRC. The whole framework is discriminative and robust against illumination variations, noises and occlusions. Experimental results demonstrate the high performance of the algorithm.3. This paper propose the Semi-supervised Competitive Agglomeration based on classified K-SVD algorithm(SCA-KSVD). Firstly, the algorithm reduces the number of dictionary atoms by clustering algorithm, and construct optimal dictionaries for better learning. Secondly, the optimized dictionaries turn more discriminative and representative by iteratively learning with K-SVD algorithm.This thesis performs extensive experiments on three public face databases, including the Extend YaleB, ORL and AR to demonstrate the effectiveness of the proposed method, and compares our appraoch with many classical sparse representation and dictionary learning algorithms. Experimental results verify that our algorithm outperforms other algorithms.