Research And Application Of Face Detection Algorithm Based On Deformable Model

The facial image analysis of human is one of the hottest research fields in computer vision.Its applications require face detection as first step.Currently,frontal face detection methods have achieved dramatically success,and they play an important role in face recognition,biological information security,virtual reality and etc.However,the fast face detection is relatively weak in unconstrained capture conditions.In this paper,we surveyed the most real world advances based on traditional deformable parts-based model,and bought in face alignment methods.Then we proposed a two cascade fast face detection framework.What's more,traditional support vector machine and boosting learning approach were implemented in this framework.Experimental results showed that our methods achieve fast detection speed,high accuracy and low false alarms on some benchmarks.Our contributions are as follows:(1)As the traditional deformable parts-based model improved detection accuracy by serval trained multi-view models,it cannot meet real-time requirement.To handle these challenges,this paper proposed a two cascade face detection framework.The first layer detector quickly supplies stable facial regions,and the second layer checks these candidates and filters invalid faces.The second detector learns discriminately local patches features that generated with face shape that provided by alignment method.This framework includes three advantages:cascade structure speeds up detection;one model for multi-view faces detection;samples are easily collected.(2)We proposed a fast face detection method based on two stage cascade SVM with above-mentioned framework.In the first layer,sparse face features effectiveness was analyzed in theory.In the second layer,sophisticated local features,such as SIFT or SURF,were extracted surrounded by facial landmarks.We conducted hard samples mining to speed up convergence in the training procession.We applied the program with C++ and obtained a real-time face detector by several approaches.Compared with recently advance methods,the proposed face detector ran fast in VGA video,and achieved comparative precision.It reduced false positive rate,and provided an effective approach for unconstrained condition faces detection.(3)To handle the problems of traditional cascade boosting methods that met high false positive rate and a large of positive samples collected inconveniently in joint multi-task training process,the cascade face detection based on face alignment awareness was proposed.Inspired by the fact that myopia's capacities of distinguishing non-face regions without glasses,the first layer detector imitates this procession and rejects the most background sub-windows.In the second layer,local features are extracted exactly from diamond patches that surrounded facial landmarks.Furthermore,we also tried to handle complicated challenge when face shape regression machine trained in supervised way failed to obtain face shape.Experiments showed that the proposed method achieves high detection accuracy and low false positive rate,with handled the problems of face pose variations and occlusion.