Face-based Super-resolution Technology Based On Generative Adversarial Networks

Human face,as inherent characteristic,is significant to identify personal identity.During these years,the research on related technologies such as face recognition and image detection has reached unprecedented heights,and has been successfully applied to intelligent transportation,identity authentication,security,criminal investigation,smart home,etc.each field.However,due to the limitations of software and hardware conditions,and such as long distances,imaging distortion,downsampling,and other factors,the face image obtained by the camera may have problems such as blur,low resolution,and loss of facial details,which affects detection and recognition accuracy.In order to solve this problem,this paper carry out research work by using deep learning-related technologies,we proposes face super-resolution technology based on improved generative adversarial networks,and constructs related network models.The specific research content is as follows:1.Based on the introduction of typical generative adversarial network related technologies,this paper analyzes its basic models and training methods,selects appropriate training parameters,and theoretically confirms the application of generative adversarial networks to face super-resolution compared with other traditional super-resolution algorithms and deep learning methods such as SRCNN,DRCN,ESPCN,it has great advantages.The super-resolution face image generated by it increases the magnification and also obtains a better face.high-frequency details.2.This article has made corresponding improvements to typical generative adversarial networks.First of all,some improvements have been made to the network structure,including removing its BN layer in the generating network part and proposing a new convolution-based residual dense block(OctRRDB);in the discriminating network part,replacing the original discriminative network with a discriminator in relatively average generative adversarial networks;a new loss function and a network interpolation method are proposed.Through the above improvements,it is expected to achieve better facial super-resolution effects.3.A multi-scale,multi-class,multi-scenario face dataset was independently constructed,which provided reliable data support and reference for face super-resolution work and low-resolution face recognition work based on theimproved generative adversarial network.Apply the improved generative adversarial network to face super-resolution,design appropriate training methods and training parameters,and train a satisfactory network model.Based on this method,the face super-resolution algorithm makes face images much clearer.Also,this algorithm solves the problem of artifacts in the face super-resolution algorithm based on typical generative adversarial networks.4.The face super-resolution technology based on improved generative adversarial network has been applied and tested in face recognition.The results show that the algorithm can effectively improve the accuracy of face recognition.Therefore,the face super-resolution technology based on this method has a positive impact on face recognition and face detection,it has a good reference value for academic research and its application in the fields of intelligent transportation,identification,security,criminal investigation and smart home.