Face Detection Based On Component-based Multi-angle Model Fusion

In the field of computer vision and pattern recognition, face recognition has been very popular research focus. As society progressed, people's life is becoming more and more rich and colorful, with the rapid development of modern high-tech, many fields are involved in face detection, such as intelligent entrance guard system, video conferencing, the subway station, bank monitor. Both the face detection system or the face detection technology is portable to other platforms, face detection improved will naturally improve face detection system. So far, there are many face detection methods, but face detection is still a challenging research topic dues to the face itself has more flexible variability and nature environment is complex.Deformable part-based model(DPM) which from P. Felzenszwalb is able to deal with highly variabled object classes and achieves excellent results in the Pascal object detection challenges. DPM relies on new methods for discriminative training with partially labeled data. DPM can improve the efficiency of face detection using the relationship between face overall information and part information, even the complex face target or in the complex nature environment, the model can work better.In this paper, a multi-view face detection is proposed based on the deformable part-based model. First of all, the cell size of HOG feature of face detection used has been studied, it sets sbin 6,8 and 10. We found that 10 is the best setting through experiments. The second, how do the number of model part make an effect for face detection has been researched. Eventually, when the number of part is 8 we can got the best detection effect. Finally, a multi-view face detection is proposed based on above two research and different kinds of view of face in the complex environment. We train two angel models for front and profile faces. The two models for different view are made fusion to achieve multi-angel face detection according to special fusion strategy. For avoiding undetected, we also make the lower threshold of profile face model detection. Experiments of the proposed metiod which carried out in the FERET databases, PASCAL databases and the face databases constructed by oneself show better performance.