Researches On Face Recognition Method Based On Kernel Sparse Representation

Face recognition technology is a technology which uses the computer to extractfacial features and then perform authentication according to these features. As one ofbiometric identification technology based on physical characteristics, it has broadapplication prospects in field of information security, criminal detection, civil andeconomic matters, gateway control and so on. Although people can easily recognizethe face and its expression changes, but it is rather difficult for the computer toautomatically recognize face image. Therefore, face recognition is still the researchhotspot in field of pattern recognition and computer vision.Different from traditional face recognition algorithm, sparse representationbased classification algorithm adopts over-complete dictionary to sparsely representsignals and then classifies images by the coefficients. It has a good classificationperformance and now has been successfully applied to the face recognition. With theexisting face recognition technology and the latest sparse representation theory as theresearch background and aiming at solving the problem that face recognitionalgorithms are of poor robustness to changes such as expressi on, pose, light conditionand occlusion, this paper further studies face recognition algorithm based on kernelsparse representation. Some meaningful results are obtained. The work of this papermainly includes:1. Inspired by prior knowledge of face images‘approximate symmetry, a symmetricGabor features and sparse representation based algorithm is proposed in this paper. Inthe framework of the proposed algorithm, we firstly perform mirror transform to faceimages to get their mirror image, with which the face images can be decomposed intoodd-even symmetric faces. Then, Gabor features are extracted from both odd facesand even faces to get the Gabor odd-even symmetric features, which can be fused viaa weighting factor to generate the new features. At last, the newly obtained featuresare combined to form an over-complete dictionary which is used by sparserepresentation to classify the faces.2. The distribution of face images can be highly nonlinear under a perceivablevariation in viewpoint, illumination or facial expression. In order to improve thealgorithm’s capacity to process nonlinear data, this paper introduces kernel trick to theGabor feature and sparse representation based face recognition algorithm and proposes Gabor feature and kernel sparse representation based face recognitionalgorithm. Furthermore, we use the existing sparse representation reconstructionalgorithm to obtain the coefficients of the kernel sparse representation.3. As for the problem that the parameters of kernel methods ar e difficult to determine,multiple kernel learning method is proposed to choose kernel parameters of kernelsparse representation in the paper.This paper conducts contrast experiments on Yale、AR、FERET face databases. Theexperiment results shows that our proposed methods improves robustness of the facerecognition algorithm for they can achieve better face recognition results eve n whenface images are under variation in expression, pose, illumination and occlusion.