Research On Video Face Verification Algorithms Based On Deep CNN And Local Textural Features

With the reduction of image acquisition cost,image verification equipment and central monitoring equipment having become popular,and face recognition technology has gradually emerged.Along with the social media developing sharply in recent years,high-level features such as role and theme based on face recognition in video environment has greater research value than face recognition based on static image.The analysis and research in video have greater universality and use value.With the development of convolutional neural networks,static face recognition has been gradually overcame,but face recognition based on unconstrained video environment is not often concerned.Especially in the case of insufficient of data set size,low resolution of samples,different characteristics and varied background,video face recognition is destined to rely on the precious inter-frame information of the video.This paper analyzes and introduces the outstanding scientific research results of video face recognition at home and abroad in the past five years from various angles,which focuses on the application of deep learning and convolutional neural networks in the field of face recognition.In the past research,in order to achieve "end-to-end" deep learning ideas,more algorithms use RGB or gray scale images as input,and increase the depth or width of the network,although they can greatly improve the performance of the algorithm,but expect too much of the training equipment and testing platform.In view of the above difficulties and considerations,this paper propose a face verification algorithm based on the TPLBP and the Siamese shallow convolutional neural network and a three-tuple deep resnet framework suitable for multi-sample face recognition verification.In order to reduce the calculation requirement of the training and the testing platform,the face recognition algorithm based on the TPLBP of the Siamese shallow convolutional neural network uses the local texture feature named TPLBP that is excellent in the face recognition direction.The TPLBP texture feature is used as the input of the network.In order to extract the inter-frame information of the video,the texture features maps of the multi-frame images are stacked and then used in a Siamese shallow 3D convolutional network,which realizes the feature reduction and classification.The Siamese shallow 3D convolutional network use the paired structure to solve the similarity of the video to the high-level features,and add a linear transformation to the interval of 0 to 1 for the similarity.Finally,the classification result will be obtained with the threshold of 0.5.In order to enable the network to capture more abundant low-level feature information,the triple-feature multi-feature fusion depth video face recognition algorithm uses another FPLBP,whose scope is wider than TPLBP.Furthermore,replace the network with a 3D residual convolutional network structure(with bottleneck structure)and take a triplet loss description that well suited to similarity-than-conceptual recognition tasks.The loss description is a good solution to the problem of insufficient information in the category sample.Face verification algorithm based TPLBP and the Siamese shallow convolutional neural network greatly reduces the network scale and model computation compared to the sameperformance “end-to-end” depth method.The triple loss deep video face recognition algorithm scales larger but have better performance.The above two algorithms have been trained and tested on YouTube Face dataset,and the training process and test results are presented in a multi-angle manner in the form of learning curve,verification curve,etc.,compared with the benchmark algorithm of the database.The effect of structural changes on model performance is discussed.We analyze and forecast the challenges and development direction of video face recognition.