Research On Super-Resolution Reconstruction Method Of Face Based On Generative Adversarial Network

Facial features are the key information of human identity recognition,so high-definition face images are vital for face recognition.Face image resolution is low due to acquisition equipment,weather,shielding and other reasons,which reduces the accuracy of face recognition technology.In recent years,the research of face image transformation from low resolution to high resolution reconstruction by using interpolation method,reconstruction method and deep learning method has attracted the attention of many researchers.In order to solve the problem of low resolution of face image,after in-depth understanding of the related technologies of deep learning method,this paper proposed the method research.Firstly,thesis proposed a Dense Attention Block to replace the residual blocks in SRGAN generators.The residual network is used in the deep feature extraction module of the original generator.The information interaction by means of jump link will lead to incomplete information mining between modules.Therefore,thesis proposed a dense attention block to replace the original residual block.In thesis,dense attention blocks can effectively reconstruct face information by ignoring unimportant information by introducing attention mechanism.At the same time,the dense attention block adopts the dense connection mode to fully extract the information between modules,and reduce the network parameters of the generator,effectively.Secondly,for the improved SRGAN network,a combination of Spectral Normalization(Spectral Normalization)is proposed to optimize the discriminator network.Since the SRGAN network belongs to the branch of the generative confrontation network,the SRGAN network also suffers from unstable training.Therefore,thesis introduces spectral standardization technology based on the improved SRGAN network to stabilize the training of the discriminator.By introducing spectral standardization to control the Lipschitz constant of the discriminator network in the SRGAN network,the instability of training is reduced.Thereby,the generator of the improved SRGAN network is optimized to improve the performance of the model.Finally,the proposed method is implemented based on Tensor Flow and Keras frameworks.The public dataset Celeb A European face images are used as the network training dataset.And a self-built 3500 Chinese face image dataset is used to verify the generalization of the model.The model proposed in this article is compared with network models such as SRGAN,SRCNN,VDSR,etc.The improved model in this article has improved PSNR and SSIM compared with the other models,and subjectively feel better.