| The rapid development of computer technology,as well as the potential demand in the field of security,promotes the birth of computer vision algorithms.Cameras that can be seen everywhere in our lives provide a lot of video data,and lay the foundation for the development of video processing.In the field of authentication,face images are the most readily available video data.Faces are more direct,friendly and reliable than fingerprints and iris.So a large number of face recognition algorithms have emerged.The traditional image-based face recognition has come to its bottleneck,while videos contain more spatial and temporal information.At the same time,face recognition in video suffers more variability.In different videos,the clarity and the light condition vary.Also the gesture and expression of faces are complex.What's more,it is easy to encounter occlusion in dynamic videos,which challenges the face recognition in video.This paper discusses two kinds of common face recognition algorithms,and analyzes their advantages and disadvantages.Based on the existing feature extraction method of 2D sparse coding and pooling,a sequence-based 3D video face feature extraction algorithm is proposed.This algorithm divides the face sequence into a series of sub-sequences with fixed length,which can be regarded as three-dimensional cubic pixel blocks.Then these blocks can be cut into small standard cubes.A dictionary can be made from parts of these cubes by K-SVD algorithm.The sparse coefficient calculated based on the dictionary will be the feature of each cube.Due to the high dimensionality of the sparse feature,pooling with multi-level is performed to reduce the dimension,which preserves both the global and local features.The final feature is very discriminative,and it can achieve good results with linear SVM classifier.Experiments show that,on our own dataset,accuracy can be improved by 14.66% compared with the common 2D sparse coding method when taking 10% of the total sample as training set. |
PDF Full Text Request