The Design And Implementation Of Face Clustering In Videos

Nowadays,in the information society,the quantity of video has a geometric growth on the Internet.How to mine useful information from those videos has become an interesting research field.In the paper,a system of face clustering from the videos is designed and implemented.This system aims to retrieve all the faces through a video and made these faces clustered into different categories to find when the person will appear in the video and how long it would last.The system consists of three modules:face tracking,face pre-clustering based on face image sets and face re-clustering using more accurate face verification.Face tracking aims to capture all the faces inside the video and divides them to different face image sets.After obtaining the faces,LBP feature from these faces and calculate the distance between two face image sets using a method based on SVM,then hierarchical clustering is adopted to make these image sets clustered.In order to solve the problem of low recall in the pre-clustering module,we use a more advanced face verification algorithm.The algorithm is based on deep learning and joint Bayesian.This module could re-cluster those same faces that can’t be clustered in the pre-clustering module.At last,the whole system was tested in two video datasets:CINESER and BigBang,the result shows that the mean accuracy and mean recall of the system could both be around 90%;in the mean time,the time consumption of different types of videos processed by the system was recorded.