Research On Low-resolution Face Recognition With Joint Multi-loss Function Based On GAN

Face recognition technology is the basis of data acquisition and decision-making methods for autonomous vehicles.With the application and popularization of video surveillance in traffic and life,acquiring face image data from surveillance video has become the main face image acquisition method at present.Face recognition based on low-resolution face images obtained from surveillance video is a difficult problem in the field of face recognition.Therefore,this thesis studies low-resolution face recognition.At present,super-resolution technology mainly focuses on improving image quality,and does not optimize face recognition performance.This thesis proposes a low-resolution face recognition method based on GAN（Generative Adversarial Network,Generative Adversarial Network）.The specific research contents are as follows:（1）Low-resolution face recognition method.The influencing factors of low-resolution face images and how to degrade high-resolution images into low-resolution face images close to the real are analyzed in detail.Three basic methods of super-resolution reconstruction are introduced and analyzed.The high-resolution face recognition process and the low-resolution face recognition method provide a theoretical basis for the subsequent research on low-resolution face recognition.（2）A face super-resolution reconstruction model based on the attention mechanism generative adversarial network.The basic structure of GAN is expounded,and it is analyzed that GAN has the disadvantage that the reconstructed image texture details are not rich enough,resulting in insufficient high-frequency information.Aiming at this shortcoming,a face super-resolution reconstruction model based on the attention mechanism generative adversarial network is proposed.The attention module and the residual block are introduced into the GAN generator respectively,and the weight value is increased in the part containing the feature information in the face image,and the high-resolution face image is reconstructed.An attention module is introduced into the discriminator,and the Adam algorithm is used for iterative optimization to improve the performance of the super-resolution reconstruction model.（3）GAN-based joint multi-loss function low-resolution face recognition model.Use the super-resolution model to reconstruct low-resolution face images,and then perform face recognition on the reconstructed high-resolution face images.A GAN-based joint multi-loss function low-resolution face recognition model is proposed.The cosine distance between face features is used as the identity loss L_id,and the total loss of the super-resolution reconstruction model is used to construct the total loss L_AGANof low-resolution face recognition.Loss to constrain the super-resolution reconstruction model to learn face identity information,reconstruct images with more face feature information,and improve the performance of low-resolution face recognition models.（4）Experimental analysis and verification.Taking the CASIA-Face V5 face database as an example,the validity of the GAN-based joint multi-loss function low-resolution face recognition model constructed in this thesis is verified,and three comparison models are designed,namely the bicubic difference algorithm Bicubic,The enhanced deep super-resolution network algorithm EDSR（Enhanced Deep Super-Resolution Network）and the enhanced super-resolution generative adversarial network algorithm ESRGAN（Enhanced Super-Resolution Generative Adversarial Networks）are used to compare and analyze the super-resolution reconstruction effect and the accuracy of low-resolution face recognition from qualitative and quantitative aspects,respectively.Compared with the high-resolution face images reconstructed by the three control models,the low-resolution face recognition model has better evaluation indicators,and the recognition accuracy is higher in low-resolution face recognition.