Research On Face Recognition Method With Illumination Change Based On Generative Adversarial Networks And Fuzzy Masking

Due to the needs of anti-terrorism,homeland security,and social security,identity recognition becomes more and more important,and face recognition is widely used due to its features such as no contact and easy access.However,the face images captured in scenes that do not require cooperation have problems such as illumination,posture,occlusion,etc.,making the accuracy of face recognition low,especially changes in lighting conditions,which will cause a significant drop in the accuracy of face recognition.In this thesis,the following researches are conducted on this issue:(1)Based on the low-light image illumination enhancement model based on the Generative Adversarial Networks,aiming at the task of face recognition with illumination change,a new method of face recognition based on adaptive enhancement of mask region and identity feature constraint is proposed for paired images.The recognition method starts from the pixel level and feature level of the generated image to improve the accuracy of face recognition.Firstly,for the pixel level,in order to make the generated image have better lighting conditions,facial features and face contour,the face analysis module and facial features symmetry prior are introduced to enhance the facial features areas with poor lighting conditions and maintain the facial features and face contour;secondly,for the feature level,in order to make the generated image and the standard image of the corresponding identity have similar features,combined with the characteristics of face data set containing standard images,a pair of images is constructed,and a face feature extraction network is introduced to impose feature level constraints on the generated images.The experiment shows that this method can better enhance the illumination of low-light face images and improve the accuracy of face recognition.(2)For the lack of pixel level constraints of paired images in the above work,which leads to the lack of details in the generated images.Based on the above work,this chapter proposes a face image illumination enhancement method with generative adversarial network based on self supervision and fuzzy masking.Firstly,based on the traditional gamma correction,an adaptive Gamma image illumination enhancement network based on deep learning is proposed to enhance the low illumination image adaptively,and uses the existing paired data set for training,then the trained model is used to enhance the low-light human face image,and the gradient of the enhanced image is taken as the self supervision information;Secondly,draw on the image enhancement method of adaptive fuzzy masking,obtain the mutation information of the image based on U-Net,fuse it with the image generated by the abovementioned generating confrontation network,and use the self-supervised information to constrain the gradient of the fused image to enhance the image details.The experiment shows that the method can improve the quality of the image and improves the accuracy of face recognition.(3)In order to further improve the accuracy and confidence of face recognition,this chapter proposes a face image re-recognition method based on sparse representation with graph convolutional network.This method draws on the idea of few-shot learning to construct query sets and support sets,and uses graph convolutional networks for category derivation to improve the accuracy and confidence of face recognition;In addition,because the existing graph convolution network only uses feature information,does not consider pixel information,and the initialization of opposite edge is completely random,To further improve the accuracy of face recognition,sparse representation is used to introduce pixel information,and use it to initialize the edge of graph convolutional network.The experiment shows that this method can further improve the accuracy and confidence of face recognition.