Research On Face Super-resolution Algorithms Based On Improved Generative Adversarial Networks

Face images obtained in an uncontrolled environment are affected by external factors,and often have problems with small size and blur.This is a great obstacle to the detection of cases by public security organs.Face image super-resolution algorithms can improve the face images resolution and enhance the visual effect.Therefore,the algorithms have high practical application value in the field of public safety.Deep learning algorithms are becoming more mature and have great advantages in dealing with image problems.In order to improve the quality of face images,this paper mainly applies generative adversarial networks to face image super-resolution algorithms.After analyzing the basic principles and existing problems of the original generative adversarial network,two super-resolution algorithms for face images based on the improved generative adversarial network are proposed.The main work of this article is as follows:(1)The principle of generative adversarial network is analyzed,and various problems such as the disappearance of gradient and variable training in the original generative adversarial network are pointed out.Experiments prove the effectiveness of the improved algorithm.At the same time,some novel convolutional neural network modules have been studied and analyzed.Here,based on these basic networks,the basic generation adversarial network is improved.(2)This paper proposes a super-resolution algorithm for face images based on spectral normalization to generate adversarial networks.This method first improves the generator network.It is proposed to use a dense network(DenseNet)without batch normalization layer(BN)as the core of the generator network.This network can effectively avoid the problem of vanishing gradient,and can realize feature multiplexing,which is conducive to generating higher quality reconstructed images.In the discriminator,this paper uses spectral normalization to improve the discriminator.Spectral normalization is easy to implement and can effectively make network training more stable.In terms of loss function,a hinge loss is introduced.Compared with other algorithms that generate adversarial networks through experiments,the algorithm proposed in this paper verifies the superiority of the algorithm proposed in this paper.(3)A face image super-resolution algorithm based on RaGAN is proposed in this paper.In the design of the generator,this method uses the Res2Net module.This module can represent features on multiple scales and make the receptive field of each network layer larger.The activation function of the generator network uses a more advanced Mish activation function.For the discriminator,this paper uses RaGAN for improvement.Different from the standard discriminator D,which estimates the probability that one input image is real and natural,a relativistic discriminator tries to predict the probability that a real image is relatively more realistic than a fake one.In terms of the loss function,the pixel loss uses the l1 loss function,which converges faster than the l2 loss function.For the perceptual loss function,this paper chooses to use the features before the activation layer.Compared with the experimental results of other super-resolution algorithms,the superiority of the algorithm is verified from objective indicators and visual effects.