Research On Face Feature Extraction From Small Sample Size

Image acquisition, preprocessing, feature extraction and feature matching andclassification are the major steps of a face recognition system. Face feature extractionplays a very important role in a face recognition system, which this article mainlyfocused on. There are always some conditions for feature extraction, such as multiplesamples, small samples or a single sample, etc. The researches on face featureextraction from small samples or a single sample, not only have the realistic andacademic significance, but also have the extensive application prospect. Afterstudying feature extraction from multiple samples, we further reseach featureextraction from small samples and a single sample, during where we put forwardsome new frameworks and new ideas.First of all, a new framework of feature extraction from small samples isproposed in this paper. In order to represent face features better and reduce dimension,wavelet transform could be firstly used for extracting face features. Then theapproximation coefficients are processed by HOG+LDA technique and mean squaredeviation is employed to handle horizontal, vertical and diagonal detail coefficients,respectively. Experiments show that the proposed menthod can obtain goodrecognition rate and decrease the computations.Secondly, the hybrid method of the two-dimensional wavelet transform andmaximum margin criterion(MMC) is proposed. That is, a source image is firstlyprocessed by the three layers two-dimensional wavelet transform. Secondly, MMC isemployed to handle the approximation coefficients. The Euclidean distance is used todistinguish the category. This method can overcome the singularity or morbidity ofwithin-class scatter matrix with a single training sample in LDA. Finally, using the multi-scale and multi-direction characteristics of Gaborwavelet transform, a new algorithm is proposed to solve the problem that there is nowithin-class scatter matrix with a single training sample. Making full use of thecharacteristics of maximum margin criterion or linear discriminant analysis, we canget higher recognition rates and less transmitted data.The ultimate purpose of this study is to find a way to extract the effective facefeatures as quickly as possible from a single training sample under the condition of ahigher recognition rate and less transmitted data.