Face Recognition Based On Sparse Representation

Biometrics uses the unique physical or behavioral characteristics of a human forindividual identification through computers, biosensors and principles of biostatistics. As amajor aspect of biometric identification, face recognition has application prospects and greatresearch value in record management, security systems, credit card verification and criminalidentification etc.However, the recognition rate will be decrease sharply when it refers to the non-idealimaging environments or the incorporation of users, such as illumination variations, posevariations, expression variations, accessory variations and dressing etc. Besides, it will be alsoinfluence the recognition results when the database is too large or small. The sparserepresentation based classification for face images has been one of efficient approaches forface recognition. It is conducted by evaluating the sparse coding coefficients. Nevertheless,the sparse representation for face recognition has great demand for the number of databaseand may not handle well when facial images appeared in variant illumination.Therefore, in order to deal with above problems, the research of the dissertation isconcentrated by the following aspects:Firstly, a novel method for illumination problem is proposed in this paper. The totalvariation model was introduced to optimize the parameters of Retinex. It can efficiently solvethe uneven illumination for face images.Secondly, the paper proposed a new algorithm based on the improved Retinex algorithmand sparse representation (STR). The original Retinex algorithm will be optimized by partialdifferential equation and then the obtained features will be used as dictionary of sparserepresentation for classification.Thirdly, an algorithm MS (a method based on ModulePCA and sparse representation) forthe single sample face images is proposed. Primarily, the method of ModulePCA is used toexpand the number of samples. Then, the transformed images and the original images will beused as a new training set for sparse representation.In this dissertation, we have done sufficient experiments on ORL, YaleB, CAS-PEAL,IMM and PolyU-NIR Database, and also on the dataset built by our own Lab, the results showthat our algorithms is comparable to others.