Research On A Simple Video Static Summarization Based On Human Face

Generally,video summarization is generated from key frames.This paper focuses on the analysis of human face in the video,and key frame sets are extracted based on faces.But since a large number of faces can be detected in a video including redundancy,and the dimension of facial features is usually very high,therefore,in order to improve the accuracy and speed of video summarization systems,this paper presents a method for describing face-track which is a face set generated by tracking a person's face in a video.In the proposed face-track description,all the face images are projected into a pre-trained feature space,resulting in a low-dimensional and compact representation.Act in response,in this way,original features of faces in a video are compressed not only from quantity but also dimension.In experiments,several feature extraction methods are compared,and several clustering methods are analyzed,results show that proposed face-track description method can achieve better performance not only in time but also in accuracy of clustering.Generally,video summarization is generated from key frames based on clustering algorithms.However,common approaches need to predefine the number of clusters and the time order of frames is ignored in these approaches as well.In order to improve the accuracy of video summarization,a new approach named face video static summarization based on sparse subspace clustering is proposed in this paper.Different with common approaches that always extract features of whole frames and analysis the content of frames,this approach focuses on faces in frames.After detecting and tracking faces from frames,face-tracks will be generated.Then,sparse subspace clustering thatno need to predefine the number of clusters,is used to classify face-tracks into several clusters,resulting in that each cluster depict a person.Therefore,a video can be summarized as a set of people.When the user choose one person`s face,the system will quickly shows all the keyframes including this person.It means that proposed approach not only can be used to get summarization of video,but also can be used to search videos containing an interested person.Experiments compared proposed approach with several state-of-art approaches,results show that proposed method can achieve better video summary.