Study On Face Recognition Methods Based On Temporal And Spacial Features Extraction

The study of face recognition is a frontier subject in the filed of computer vision which has been paid close atteion by researchers and has been widly applied in many fields such as criminal investigation,public security ans video retrival and other fields.In recent years,the continuous development of video or image collection and storage tools such as cameras,film and television media,monitoring system and web media,has made the acquisition and storage of video and images more convenient,which has brought sufficient realistic conditions to face recognition.Compared with other objects,human faces have unique properties and complex 3D structures.The spatiotemporal features of human faces contain important information about human face structures.The face recognition technology and method based on spatiotemporal features extraction in this thesis has good universality and adaptability,and can be widely used for face recognition in various environments.Temporal features cantain important information in the field of multi-target video based face recognition.The detection and recognition of faces in the video is difficlut due to the dynamic changes of faces and the simultaneous appearance of multiple faces in video.Temporal features can be used to correct the accuracy of face detaction and deduce the false detection and missed detection.Meanwhile,temporal information can be used to gradually improve the recognition accutacy of multi-target face recognition.Video or iamge set cantains more information about various aspects of human face,but the data collected contains many factors that affect the finall result of recognition such as character angle,the image blur,facial shade,expression change,wear dress,light conditions and changes caused by the difference of image capture system,etc.,which brought new challenges to face recognition.Spatial features extraction and be used to improve the discrimination ability of face recognition system by modeling or simulating the 3D structure of human faces.therefore,the in-depth study of this subject is of great significance.The main innovation points of this paper include the following three aspects:(1)This thesis proposes and studies an video-based face enhanced detection method based on temporal feature extraction,as well as a multi-track incremental learning method for identifying multi-person in video.Since the mainstream V&J face detection algorithm has some false detection and missing detection,this paper USES a stable video target tracking method to enhance detection.The experiment shows that the combination of detection and tracking method can effectively improve the results of face detection,thereby improving the accuracy of face detection.The average face detection rate of our method(89.51%)is better than V&J method(64.97%).At present,video-based face recognition mainly focuses on the recognition of single face in video,and the occurrence of multiple faces at the same time often occurs in some scenes.However,there are few researches on this aspect.This paper proposes a multi-track incremental learning algorithm,which combines the above enhanced detection and face recognition based on LBPH feature to conduct video multi-face recognition.The experiment shows that the multi-track incremental learning method can effectively improve the accuracy of video multi-face recognition.The recognition average result of LBPH+MTIL method(94.44%)is better than LBPH method(83.33%)on Honda/UCSD dataset.(2)This thesis proposes and studies a face recognition method based on pose estimation with facial geometric spatial structure.Generally,attitude estimation is applied to face of single picture,but face recognition based on picture set has little work in this aspect.Simply transplanting the traditional Pn P method for attitude estimation of single picture to image set leads to the increase of computational complexity and even failure.This problem can be solved by defining a forward image model for each image set,rather than using a fixed model as a baseline.Experiments show that the method of face feature point attitude estimation can get better recognition results than some advanced methods.Our method(97.44%)exceeds the best method SANP(92.31%)on Honda/UCSD dataset(50 frames).On You Tube Celebrities dataset,our method(77%)exceeds the best method IDLM(76.52%).(3)This thesis proposes and studies a facial points data alignment machine learning method based on spatial feature extraction for image set-based face recognition.There are many methods for feature extraction,such as original pixel,HOG and LBP feature,etc.The features extracted by these methods tend to have higher dimensions and tend to contain a lot of redundant information.Facial feature point is the human visual perception of the nature of human face,it can be used to reduce feature dimension and reduce redundant information,however,for each frame of video or photo collections to extract face feature points is irregular,which requires a method of alignment will be irregular face feature points aligned to a common coordinate space.Experimental results show that this method can improve the accuracy of face recognition in image sets.Our method(94.87%)exceeds the best method SANP(92.31%)on Honda/UCSD dataset(100 frames).