Research Of Video Face Recognition Based On Keyframe Information Integration

In the procedure of video-based face perception technology, the problem of howto steadily and exactly recognize faces which have substantial changes in posture andscale in video is inevitable. In such situation, the traditional face recognitionalgorithm only recognize faces with a single static video frame, depending on theframe we have chosen. Obviously, this way is neither reliable nor accurate. Eventhough for the sake of improving system stability, adding up several video framesequences to identify results using mathematical methods and finally work out aconclusion, there is often more harm than good because of the increasing computationcomplexity. As a newly developing topic, Compressed Sensing (CS as anabbreviation) provides solution seeking mindset in numerous fields. Although facerecognition algorithms based on CS always aim at static image at present, consideringthe direct information sampling characteristic of CS and the frame-to-framecorrelation feature of video, we take both into account. In detail, randomizing thevideo frames’ information in advance, then use it in face recognition based on CS inorder not to exceed system computational burden too much, but get a better systemperformance.By studying the face perception technology, especially the video face recognitioncurrently, this paper puts forward a video-based face recognition system usingkeyframes information integration in the light upon understanding and comparingseveral face recognition arithmetic. Firstly, extracting keyframes from videosequences which are suitable for face identification in accordance to the similarities ofcolor histogram. Then Adaboost+Haar algorithm is used to detect the face inkeyframes, followed by denoising and normalizing the detected face region images.Finally, with the knowledge of Compressed Sensing theory, proposing a CS method,which is appropriate for video-based face recognition on the basis of keyframeinformation integration that replaces the single frame by randomly sampling multiplekeyframes image information to make improvement. Theory analysis andexperimental consequences prove that our system has relatively higher recognitionrate, preferable robustness and feasibility, thereby meet the requirements oflow-resolution video applications.