Research On Fast Face Detection Technology In Video Streaming Based On Deep Learning

Face detection,as a prerequisite for many applications such as face recognition,expression analysis and so on,has attracted widespread attention.The promotion and application of cameras provide safety guarantee for people's life and work.Applying face detection and face recognition to video processing to assist people in completing functions such as video surveillance and person search,which can reduce the investment in human resources and improve work efficiency.When applying face detection to video processing,it is necessary to ensure not only the detection accuracy,but also the detection speed.Therefore,this dissertation mainly studies the fast face detection of video,and applies it to the actual video scene.Firstly,this dissertation introduces the object detection network YOLOv4 and applies it to face detection tasks.Through the analysis of face attributes,based on the bounding box regression loss of YOLOv4,we propose six improvement schemes.We perform cluster analysis on the annotation information in the WIDER FACE dataset.According to the clustering results,we adjust the prior setting and network structure of YOLOv4 to improve detection performance.In addition,the proposed improvement schemes are applied to the face detection network Retina Face to improve the detection performance of the model.Through experimental verification,when using YOLOv4 and Retina Face to perform face detection on the video stream,it can not only ensure the detection accuracy,but also ensure the real-time detection speed.YOLOv4 cannot guarantee real-time detection speed after improving the network structure.However,in practical applications,the model can be simplified by model pruning to improve the detection speed.Secondly,when performing face recognition on video,a frame-by-frame processing method is usually adopted,which ignores the correlation between frames and leads to the phenomenon of face recognition for multiple faces of the same person.In order to reduce the number of face recognition of the same person,we study two face deduplication algorithms,namely HOG feature de-duplication algorithm and IOU de-duplication algorithm.The HOG feature de-duplication algorithm extracts HOG features from the detected faces,and calculates the cosine similarity between the HOG features of different faces to determine whether they belong to the same person.However,the HOG feature extraction takes a lot of time and is not suitable for application in actual scenes.The IOU de-duplication algorithm considers the correlation between the face position of the current frame and the face position of the previous frame.When the moving speed of the person is limited,the face position of the current frame and the face position of the previous frame will partially overlap.We use this as an indicator to determine whether the face images belong to the same person.Experiments have found that IOU deduplication ratio reaches more than 90%,and it will not affect the overall operating speed.Finally,we introduce the application of the model in actual scenarios.Face detection and face recognition are combined to form a complete face recognition system,which is used in video to realize face recognition tasks.For the real-time video,a blacklist database can be established to check whether there are blacklisted persons in the video,and realize the security warning function.For historical video,we establish a database of persons to be searched,and use face recognition algorithms to find relevant video clips of persons to be searched in the massive video data.It reduces the investment of human resources and improves processing efficiency.It also provides new technical support for public security prevention and control,personnel control,and case reduction.