Research On Face Recongnition Based On Sparse Representation

Face recognition is widely used in human-computer interaction, public safety,national security and other fields as a friendly biometric technology. It is a hotresearch topic in the field of pattern recognition and machine vision，because of itsvulnerable to illumination, expression, shade, posture, and the influence of suchfactors, high-precision face recognition is still a challenging task.Sparse representation is the key theory of compression perception. The data ofsparse representation can be essentially reduced processing cost, improve thecompression efficiency. Sparse representation used in image processing is a ind ofeffective method. The main content as follows:Studied two commonly used principal component analysis for face recognitionalgorithms Principle Component Analysis and method Fisherface methods. Theexperimental results show the two algorithms have a good effect on the amount ofdata classification on a small sample library, but it has poor classification results on alarge sample library.Analyzed the dictionary learning and sparse decomposition based on the classicsparse expression face recognition framework. In order to carry out the sparseexpression effectively, we will combine random face with KSVD algorithm to carryout dictionary learning and identification based on the idea of random projection. Theexperimental results show can obtain better recognition results whether on a smalllibrary and on a large library compared to conventional face recognition methods.Since KSVD training algorithm does not have discriminative power, we studiedD-KSVD algorithm by adding classification error to the objective function. In orderto further improve the D-KSVD algorithm classification accuracy, increaseinner-class constraints, we studied the LC-KSVD algorithm by adding discriminativesparse code error to the objective function.In order to optimize iterative effect in the dictionary learning processing, weimprove discriminative dictionary training method,that is, we train the dictionary in the class firstly and then cascaded within initialized dictionary. The dictionary retain amore complete sample information. It reduced the training dictionary error, improvedthe recognition rate. The experimental results show that it does good performance onD-KSVD and LC-KSVD algorithm.