Research And Implementation Of Local Occlusion Face Recognition Based On Deep Learning

Face recognition system is widely used in public areas such as communities,schools,stations and hospitals because of their advantages of safety,rapidity and non-contact.However,in practical applications,the recognition accuracy of existing face recognition systems is still affected by factors such as appropriate illumination,removal of masks and glasses by the recognizer,etc.Because of the above problems,this research focuses on how to improve the speed of face recognition and improve the accuracy of face recognition in the case of occlusion.Based on generative adversarial network(GAN)and principal component analysis network(PCANet),the local occlusion face recognition system is designed and implemented combined with face inpainting.The influence of occlusion on face recognition is reduced by repairing occlusion and optimizing recognition.The specific research work is as follows:1.This research proposes a multi-scale face restoration algorithm based on GAN(MRGAN).In order to solve the problem of face feature loss caused by occlusion.Firstly,GAN is used to fill the low-resolution image with low difficulty,and four groups of generative networks and discriminant networks are used to gradually generate the image.Secondly,the generative network is responsible for repairing the missing regions,the discriminant network is used to identify whether the generated distribution is close to the real distribution.Finally,in order to limit the generative network to retain the texture features of the generated face and ensure the real texture and identity information of the generated image,the loss function based on Local Binary Patterns(LBP)is used to minimize the difference between the generated texture and the real face texture.By comparing the experiments with the current popular face inpainting algorithms,the advantages of MRGAN are verified from quantitative,qualitative and computational costs.2.This research proposes a multi-scale and multi-layer feature fusion occlusion face recognition algorithm based on PCANet(MMPCANet).In view of the problem that the traditional PCANet stretches the two-dimensional image into column vectors,which leads to the loss of spatial information of the basic image and the low recognition accuracy when there are few training samples.Firstly,the original image feature and the output feature of the first layer are connected by channels,and then the connection results are used as the input of the second layer to obtain more image feature information.Secondly,in order to avoid the loss of image spatial information,the spatial pyramid is used as the feature pool layer of the network.Finally,the feature vector is sent to the random forest classifier for classification.Experiments show that compared with other similar algorithms,MMPCANet has certain advantages in occlusion face recognition.It almost completely extracts the non-occlusion area features of faces,and saves the computational cost.3.This research implements the partial occlusion face recognition system.Based on MRGAN,MMPCANet and related technologies,the partial occlusion face recognition system is designed and implemented by investigating the requirements of face recognition system and according to the software design and development process.The developed system not only realizes the human-computer interaction of face recognition algorithm but also can freely select the algorithm model.Users can conduct face recognition by uploading images or opening cameras.The system test verifies that the algorithm still achieves good results in the actual occlusion face recognition,reflecting the robustness and portability of the algorithm.