Design And Implementation Of Video-Based Face Gender Recognition System Using Manifold Learning

With the requirement of access control, video surveillance, and network video streaming, video-based face recognition has become one of the hottest topics in face recognition. As a branch of face recognition, the video-based face-gender recognition influences the development of face recognition and access control, video surveillance, and so on. However, there are lots of difficulties in video-based face-gender recognition. First, the most critical issue is how to overcome these difficulties:the low face resolution, illumination changes, poses changes and so on. Second, time-space information hidden in the video stream often ignored by using this way:selecting the most representative picture in video. The main purpose of this thesis is how to solve above problem in video-based face-gender recognition, and establish a system to achieve face-gender recognition. The main work of this thesis is as follows:1. Select an appropriate Video-Modeling method to obtain more time-space information in face images. The purpose of the Video Modeling method is to choose a appropriate way to represent pictures in video and keep the time-space information. In this thesis, k-means clustering method is used, which belongs to manifold model.2. Select an appropriate Feature Extraction algorithm to overcome the face morphology changes in video. The purpose of Feature Extraction algorithm is to select the most valuable feature to represent the face image. The common Feature Extraction algorithms regard face image as a vector. But this way ignores the spatial continuity of information of face images. In this thesis, Tensor Space Analysis is used to extract feature. This algorithm regards face image as a two-dimensional tensor (matrix). The contrast experiments show that Tensor Space Analysis is better than other algorithm. Also Tensor Space Analysis is a manifold learning algorithm. It is able to learn the structure of the low-dimensional manifold, which exists in the face image.3. VG-TSA gender recognition algorithm is used to achieve video-based face-gender recognition. The steps as follows:first, using Adaboost algorithm to detect face image; then using the k-means clustering algorithm to model video; then using Tensor Space Analysis to extract feature; finally, using the support vector machine classifier to classify gender.Depends on this way, a video-based face-gender system is realized.