Research On The Methods Of Real-time Face Detection And Posture Robust Feature Points Detection

With the rapid development of advanced theories and technologies in the field of visual computing and the great improvement of computer computing performance, the automatic analysis of facial images by computer has become a reality. Face detection and facial feature points detection are the basic steps of facial image analysis, and have important influence on the effect of follow-up work. Starting from the real time of face detection and the pose robustness of facial feature points detection, this paper puts forward some effective solutions.In view of the real-time requirements of face detection, this paper uses a two level structure face detection method to improve the traditional AdaBoost face detection. This method uses a coarse to fine manner,applys the BING feature to face description. Training the corresponding coarse classifier to filter most of the non-face regions for the detection and localization of human face. Then, the coarse inspection area of the first stage is used as the input of the second stage, and then further accurate detection is carried out. In this way, the AdaBoost algorithm can be used to detect faces in a more efficient region, thus reducing the original search range and improving the detection speed. Experimental results show that the method can effectively reduce the detection time and speed by 14.4% while maintaining the original detection rate at the same time.Aiming at the pose robustness of facial feature points detection,a cascade convolutional neural network model is designed in this paper. The first level is a global model,which is used to realize coarse location of feature points, and all network layers are connected with ReLu activation function to improve the convergence of the model. In the convolution process, the method of edge extension is adopted to avoid the small response map, and the multi-scale feature extraction is introduced to improve the input of the full connection layer and enrich the extracted feature information. The second stage is a local model for each key point,and its function is to fine tune the location of the feature points based on the coarse location information, so as to obtain the best detection point estimation. The detection based on two levels convolution neural network model not only use the global information of the human face to provide the reliable initial estimation points but also give full play to the effectiveness of local information, make the extraction of the feature information controll in a more efficient area. Experiments on several public databases (LFW, LFPW) show that the proposed model is robust to feature points detection under attitude deflection.