Research On Deep Neural Network Model Optimization For Face Detection

The intention of face detection technology is to determine whether there are faces in target images.If faces are detected,the positions of the faces will be the output of system.It is regarded as the basic task of face recognition,facial expression recognition and face alignment.In practical application,the technology poses stringent requirements on the detection speed,accuracy and the size of classifier..Deep learning-based face detection methods are mainly divided into two categories.One is the two-step method,which generates candidate windows on the original image,and then classifies and locates them.The other type is the one-step method which classifies and locates directly on the original images.The former has higher detection accuracy and the latter has faster detection speed.Although the accuracy given by both strategies can meet the requirements of most applications,the large size of the classifier is less portable and thus limits their application scenarios.In view of this situation,the main work of this thesis described as followed.(1)This thesis adopts the strategy of cascade neural network and use a three-stage cascade structure to simplify the detector.In the first stage,the network nominates candidate windows;in the second stage,the network modifies the selected windows;in the third stage,the network outputs the window positions with human faces.(2)In this thesis,a convolutional optimization module is proposed,which can deeply mine the limited information without excessively increasing the number of parameters.However,this also paid some prices when pursuing smaller models.Therefore,the existing three-stage cascade convolutional neural network needs careful optimization.(3)This thesis also proposes an end-to-end cascade neural network structure,which can simplifies the training of the cascade neural network.Not only does it reduce operation works,it also reduces the training time by two-thirds.(4)Finally,for the three-stage cascade structure,the three models: 1)model before the improvement;2)model after the improvement;3)the end-to-end models are compared based on the two metrics of recall and accuracy.The improved effectiveness of this method was verified on the public WIDER FACE detection dataset.For the face candidate window extraction stage,a 3% recall improvement(from 80% to 83%)was given by the method in this paper,while a 1%(from 84% to 85%)increase of the recall was rendered during the revision phase of the selected windows.At the third stage of outputting face window positions,recall and accuracy are both improved by 2%(accuracy: from 90% to 92%,recall: from 88% to 90%)at the same time.