Application Of Face Recognition Based On Deep Learning In Audit Of Video CRBT

Video ring back tone(CRBT)is a China Mobile's value-added service based on VoLTE.Users can upload and manage short video as their own video ring back homepage.How to audit content on a large number of videos to prevent the proliferation of sensitive content has brought new challenges to the video CRBT platform..The review content of the CRBT video is mainly sensitive people.,and the key technologies for its implementation are face detection and face recognition.The traditional manual feature methods cannot meet the practical needs of face detection and recognition tasks in unconstrained environments because of its limited ability to characterize features.With the development of deep learning,face detection technology and face recognition technology based on convolution neural network have become the mainstream.This thesis researchs on this technology in depth.In terms of face detection,this thesis proposes a one-stage face detection algorithm CSSD based on the one-stage general object detection algorithm SSD.The CSSD uses the multi-feature map prediction detection architecture to improve the robustness to face scale.and proposes the feature enhancement module to enhance the expression ability of the prediction feature map to improve the detection performance of small-scale faces.It also introduced a cascade frame for face bounding box regression in the first-stage detection structure to improve the prediction accuracy of the face bounding box from coarse to fine in two steps.CSSD has achieved good results on the public dataset FDDB and WIDER FACE,which proves its effectiveness.In terms of face recognition,this thesis proposes a multi-channel feature fusion face recognition algorithm MCFF-DeepID based on the DeepID method framework.MCFF-DeepID firstly extracts multi-channel features from face and face key organs to obtain information-rich comprehensive features.Then,using multi-channel feature fusion module to complete the fusion and screening of different features;Finally,use Softmax Loss and Center Loss to improve the training loss function in training.Based on these three improvements,MCFF-DeepID achieved 99.38%accuracy on the face verification data set LFW,which proves the effectiveness of the algorithm..With the help of the above two key algorithms,this thesis designs and implements a video content auditing system for the CRBT platform,which realizes the function of video sensitive person detection and specific character retrieval.The system accuracy rate is tested in various scenarios,and reaches 91.2%of the comprehensive accuracy performance,which can meet the actual application needs of the system.