Feature Selection And Extraction For Gender Recognition Using Facial Images

Facial images not only can be easily got, but also contain a wealth of information. In recent years, more and more researchers are devoted to gender recognition based on the facial image, which could provide the favorable priori information for expression recognition, gesture recognition and action recognition. It can also bring more commercial benefits.A couple of factors such as the illumination condition, race, age and image quality have a great influence on the facial image. The performance of gender classification based on single facial image is limited, so the main research content of this dissertation is to extract and select out a variety of features that have greater contributions to gender recognition, and to eventually realize the gender recognition system based on facial images by fusing multiple features.This dissertation proposes a facial sub-area partition method based on landmarks. Because of the poor recognition effects of single feature, this dissertation fuses the facial sub-area, PCA and LBP as the fusion feature to perform gender recognition. To avoid losing some important facial information in the process of facial roughness, this dissertation proposed a facial roughness feature based on dividing blocks to perform gender recognition. This method exploits facial grid divides a face image to some blocks, then calculates the facial roughness for each facial block as a gender feature vector. To further improve the computational efficiency and accuracy of the algorithm, this dissertation proposed an improved Elastic Bunch Graph Matching algorithm. The traditional Elastic Bunch Graph Matching algorithm has high complexity during the iterative process which cannot satisfy real-time requirements. We optimized the iterative process to improve the efficiency and accuracy of the algorithm.The emphasis of this dissertation is gender feature selection and extraction. The improved method mentioned in this dissertation overcomes the problems such that the inaccurate divide of facial sub-areas, the omission of feature information in the process of facial roughness extraction and the poor real-time performance of the algorithm. The proposed algorithm obviously improved the gender recognition efficiency and accuracy.