Pyramid GAN Network And Lightweight Network Knowledge Distillation For Image Super-Resolution

Image super-resolution generate high-quality super-resolution images play an important role in the application of images in aviation,security,medical treatment,life and entertainment.At present,the research of image super-resolution mainly changes the feature extraction structure of convolutional neural network to extract better features.SRGAN generated rich texture and high resolution images.Lightweight networks not only have fewer parameters but can also generate high-quality images.This paper proposes a Pyramid GAN network based on SRGAN.In addition,this paper proposes a lightweight network using knowledge distillation algorithm.The main work of this paper is as follows:(1)This paper proposes Pyramid GAN network for image super-resolution.In the study of SRGAN,it is found that although the super-resolution images generated by SRGAN network model have rich texture,but there are still some problems such as poor quality of highresolution images,unreasonable texture and uneven local image.Aiming these problems,this paper proposes a Pyramid multi-stage feedback network structure based on StackGAN multistage generator to improve the quality of super-resolution images.Firstly,according to the relationship between the dataset and the generated image in the process of Pyramid GAN network's training,the loss of intermediate supervision and inter-level feedback is proposed to supervise the generation of intermediate scale super resolution image.Then,combining with DenseNet and attentional mechanism,the transition layer in DenseNet structure is replaced by Saliency Block,which combines channel Saliency and spatial Saliency.Finally,the multi-loss function is used to train the generator,which is the adversarial loss of the high-level image data distribution,the perception loss of the local image texture,and the gradient loss between the low-level pixels.Experiments have proved that the image quality generated by the multi-stage feedback network structure is higher than that generated by the single-stage generator,and the Saliency DenseNet is slightly improved.The gradient loss makes the generated image edge clear and the local image texture more reasonable.(2)This paper proposes a lightweight super resolution model using knowledge distillation.The lightweight network structure can not only reduce the size of the model but also generate better super-resolution images through carefully designed structures.However,there is still a gap with the extremely deep network structure.Based on knowledge distillation,this paper presents a lightweight model for image super-resolution using knowledge distillation.The deep convolutional neural network guides the lightweight model learning feature knowledge,so as to improve the super-resolution effect of lightweight network.In addition,based on the unbalanced data enhancement method,enhance lightweight model learning of difficult images in dataset.The experimental results show that the knowledge distillation can significantly improve and promote the super-resolution of the lightweight model with the appropriate number of layers,and the unbalanced data enhancement method can slightly improve the lightweight model.