Research On Key Technologies Of Identity Recognition Based On Deep Convolution Features

With the continuous advancement of technology,identity has become inseparable from our lives and has been widely used in our production and life.Real-life access control systems,face recognition systems and surveillance systems are all closely related to identity recognition.The wide application of identity has brought great convenience to people's lives,and it also provides a strong guarantee for our security and plays an important role in real life.However,as an important technology in the field of security,identity feature extraction and classification still faces considerable difficulties and challenges,and there are many problems that need to be solved.On the one hand,whether face information can be captured normally is a key factor in the identification process.Under the condition that enough face information can be captured,how to solve the diversity of facial expressions,the change of angle and the problem of face occlusion become the key to identification.In the absence of the ability to capture face information or lack of face information,pedestrian variability,clothing diversity,and image occlusion and blurring make identification more challenging.On the other hand,due to the complex background environment of surveillance video,the influence of camera distance and resolution also increase the difficulty of identification.Therefore,this puts higher demands on the robustness of the identity feature.In order to solve these problems better,based on the existing identification,this paper studies how to extract more robust features to identity recognition and classification for different scenarios.Firstly,we usually capture enough face information for identification in the case of identity recognition in conventional controlled scenario.In order to improve the robustness of the feature,deep feature of the image is extracted for recognition.Compared with the traditional manual features,deep feature is more discriminative.At the same time,some inter-class and intra-class information are not fully considered in the process of considering deep feature extraction.We propose joint cooperative representation method using deep features to make the features more robust.The results show that the proposed method has achieved state-of-the-art performance on several benchmark datasets.Secondly,it is difficult to obtain face information or insufficient face information for identity recognition in an uncontrolled scene.This paper proposes tree branch network module with combining local and global features to learn feature extraction methods.The main contribution of this method propose a more efficient partial block method based on original image,and at the same time,the importance of the global features are considered in entire recognition process.In addition,compared with some existing methods,the proposed method does not require any additional information to assist the identification,avoiding the error caused by the introduction of additional information,and also reducing the extra calculation.Finally,based on the tree branch module,we propose a method of mutual learning between local and global features.In order to make the features more robust to adapt to the identification in complex scenes,the method combines local and global features for mutual learning.The most important thing is that we propose to combine all local features for mutual learning,which greatly enhances the identity recognition effect.