Research On Face Detection And Recognition In Video

With the rapid development of artificial intelligence and the wide application of face recognition and verification system,various illegal means of using system defects to forge and deceive have been derived accordingly.The requirement of face biopsy in video becomes more and more urgent.Traditionally,it is necessary for the verifier to cooperate with the system instructions to make corresponding actions to further analyze and distinguish the face in the video.However,due to its inconvenience,this method has many shortcomings,so it has become a new research direction to detect the face in vivo without cooperating with the instructions.On the other hand,with the continuous improvement of face detection and recognition algorithm,the complexity of the algorithm is also increasing.How to reduce the operation cost and improve the recognition efficiency without reducing the recognition accuracy has become a hot research direction in the field of face recognition.This thesis systematically summarizes the research foundation and development direction of face biopsy detection and face recognition technology in video,respectively from three aspects: optical flow motion analysis technology,key frame extraction technology and face feature recognition technology,and achieves the following technological innovations and research results.Firstly,optical flow motion analysis and detection technology is introduced into video face detection.Starting from the different characteristics of motion differences between different regions in real and false face video images,the video face detection technology is studied.Video images are divided into 10 regions according to human physiological characteristics,and Lucas-Kanade optical flow vector is selected according to scene characteristics for video images.The results show that the theoretical basis of this method is clear,and the number of living organisms can be effectively distinguished under specific thresholds.Secondly,the technology of face location and detection in video is studied in depth,and a method of face location and detection based on key frame extraction isproposed,which excludes a large number of redundant frames and invalid frames and only detects the face in the key frame of face posture standard.In the key frame extraction process,according to the application scenario,a key frame extraction algorithm combining compressed sensing sub-shot segmentation and color histogram matching is proposed.After that,the trained YOLO face detection model is used to locate and detect the key frames,and the CLM algorithm is used to align the face to get the standard face image and hand it over to the next stage for face recognition.Experiments show that the proposed algorithm can accurately and efficiently capture a standard face image in a video as a key frame,with good detection rate and adaptability.Compared with the traditional frame-by-frame detection method,the computational overhead is greatly reduced.Thirdly,facial feature recognition is studied.On the basis of Sphereface algorithm,Swish activation function is proposed to replace the original PReLU activation function.After training and testing on LFW,the average recognition accuracy of the original PReLU activation function is 92.15%,and that of Swish activation function is 92.40%.Experiments show that the recognition accuracy can be improved by using Swish activation function.Finally,the work of this thesis is summarized,and the development of face detection and recognition technology in video is prospected,and the future research directions and ideas are pointed out.