Research On Single-Sample Partial Occlusion Face Recognition Based On Deep Learning

In the field of computer vision,compared with human fingerprints,iris and other features,the faces can be used for identity matching and are more accessible.Face recognition has always been a hot topic in current research.Recently,most face recognition algorithms have performed well based on data driving.It is a difficult problem to collect multiple samples of every person in some special application scenarios.Such as ID card management system,criminal investigation law enforcement system,passport verification and registration port identification,every one can only obtain one training sample(id photo image taken by the camera),which is called a single training sample constraint.When the test face image is occluded under this constraint,the existing algorithms to solve the occlusion face recognition also perform poorly.It results in a single-sample partial occlusion face recognition issue.Aiming at this issue,a singlesample partial occlusion face recognition algorithm based on deep learning is proposed.The main contents of this thesis are as follows:1.Aiming at the problem of insufficient training samples,this thesis adopts the technical route of pre-training deep neural network with open dataset and proposes a single-sample partial occlusion face recognition method based on Residual Shrinkage Attention Network(RSA-Net).Considering that the occluded face image is difficult to identify,the RSBU-CW denoising module in the field of signal processing is applied to face recognition.The branches of the twin network consist of residual shrinkage networks.It can adaptively filter out the features destroyed due to occlusion.In order to highlight the features of the unoccluded area,we embed spatial attention modules in the network to weight at the image space level and improve the saliency of the unoccluded area features.Meanwhile,in order to make the network obtain more discriminative representations of facial features,we embed the channel attention module to scale the importance of the channel level by counting the data characteristics at the channel level.Finally,in order to compensate for the lack of adaptability of the model due to the difference between the source and target domains,the training samples are expanded using image mirroring and symmetric face transformations.The expanded samples are used to fine-tune the model.The experimental results demonstrate the effectiveness of the proposed method.2.Based on the RSA-Net algorithm described above,this theis designs and develops a prototype system for single-sample partial occlusion face recognition.Firstly,we conduct a functional requirements analysis of the system.Secondly,we introduce the system architecture design and the main business processes.Then we design the server side from three aspects: the main class functions,the dynamic model of interaction between the server-side objects,and the database tables used by the system.Next,we have designed and implemented clients based on PC and mobile devices respectively.Finally,we write test cases to test the system functionally,and the test results show that the developed system already has the basic functions of the prototype system and it has a certain practicability.