Research On Sparse Classification Method Of Facial Image

Facial feature extraction and classification via sparse representation is a characteristic method in the field of face recognition. The generation of redundant dictionary is the key issue in the sparse representation-based face recognition, and yet the transform of the facial image plane considerably affects the recognition result.Based on the above analysis, the thesis concentrates on the dictionary generating and reducing the influence of the transform (concentrating on the shift) on the recognition. The main work is as follows:①In the light of the facial sparse representation model, we propose a sparse representation-based classification method based on multi-dictionaries which are formed by the 2D Double-Density Dual-Tree Complex Wavelet Transform(DD-DT CWT). The generation of the redundant dictionaries is of key importance, which here reflects in: 1) selecting some of the sub-bands which preserve the most energy from multi-scale transformation to form the redundant dictionaries of the corresponding scale according to the energy distribution of coefficients, utilizing the different identification information from different scales; 2) selecting the different sub-bands from the same scale form the multi dictionaries, utilizing the different identification information from different orientations within the same scale; 3)the recognition result is obtained uniting sparse representations which are obtained through sparse decomposition on multi dictionaries.②With a view to the restraint from the sparse representation-based classification model, which requires strict alignment between the training and testing face image plane, the research on the model which combines the facial sparse classification framework with the transform frame has been done. The work is done in two steps. Firstly the misaligned testing image is given in the Taylor's expansion and we get the estimation of the feature of the test in the standard space (here it equals the aligned space where training samples lie) and the feature of the variation which is the linear representation of directional gradient. Secondly, we get the sparse representation in the standard space and shift parameter after carrying out the dimension reduction, decomposition and shift parameter estimation. Compared with the conventional method without estimation of the shift parameter, the proposed method efficiently reduce the restraint in the requirement of strict alignment, sequentially the recognition rate is enhanced.③The sparse factor (SF) is chosen as the evaluation criterion of sparse coefficients, whose threshold is set according to the sparse decomposing process of the facial image . It is an efficient evaluation measurement for the sparse coefficients, and yet an indirect evaluation measurement for the generation method for redundant dictionary.