Research And Application Of Pose-Robust Gender Recognition Method Based On Face Images

Gender recognition of face image is one of the basic demographic classification requirements, while in behavioral monitoring, authentication, human-computer interaction, video search, marketing strategies and other fields also has broad potential applications. In the past few decades, the gender recognition of front face images has basically reached a satisfactory discrimination, but the identification of face images which have changes in position is still a problem which is worthy of study, and at the same time, the pose problem in itself is an inevitable problem while in the application of face recognition. Whether the real-time monitoring, or the database retrieval, there will be a lot of position changed face images. In order to get better gender recognition results from these face images, we first proposed an improved face feature point extracting way, and then focused on enhancing the robustness of the characteristics of the face images, finally achieved the following three aspects of works:1. Proposed two types of more efficient face feature extraction methodsBy combining the Gabor filter and LBP operator we get more applicable LGBP features of gender identification. At the same time, we also proposed another feature extraction method based on sparse coding. Through a set of experiments on the CAS-PEAL database, we get80%and85%accuracy respectively. Considering of the difference of the correct rate gap between the two methods is not large, so we will use a shorter time-consuming way, the LGBP feature extraction methods for the next step.2Proposed an improved ASM method based on self-adaption multiple templates to strengthen the robustness of the key points extracting way of the human face imagesIn the study of extracting the key points of the face images, ASM as a simple and effective method obtained widely used, but the original ASM algorithm can’t process the problem of the position changed human face images, and thus we proposed the improved ASM method based on self-adaption multiple templates matching. By mapping the triangulation based texture, we solved the problem of extracting key points of a position changed human face image, and according to the experiments on the CMU-PIE database, this method is proved effective. 3Proposed a robust way of the gender classifier by using sparse representationWhen we come to a large library, classify with the large libraries to express themselves must be the sparsest, based on this idea, we propose the method of using the sparse representation to the classification approach. After the extraction of the database image feature and dimensionality reduction, we will construct an over-completed-dictionary, and then the test sample only need to use over-complete-dictionary to sparse express, the kind of gender is the minimal residual of test sample. Because sparse expression solving process is the process of solving the optimization, thus the classification constructed will naturally has robust of the pose problem, and through a set of experiments on the CAS-PEAL database, we got95.5%accuracy finally.At last, we built a practical system to meet the needs of the currently bayonet and other customs, and put the proposed robust gender recognition method into the system to verify the feasibility of it.