Lightweight Generative Adversarial Networks For Image Super-resolution Reconstruction

Since the 1990 s,in order to solve the image quality problem,image superresolution techniques,as an important branch of computer vision research,aiming to recover high-resolution images from low-resolution images,have been widely used in daily life,aerospace and military,medical imaging and many other fields.From traditional methods such as interpolation-based image super-resolution to the current deep learning-based technology,it has evolved and improved.The image superresolution technology has greatly improved the quality of the generated images,and at the same time has put higher demands on the computational power.The evolution of deep learning started to use generative adversarial networks for image super-resolution enhancement in 2017,which is the most famous one is Super Resolution Generative Adversarial(SRGAN).SRGAN models have problems such as high computational complexity and need of large amount of training data.These problems limit the application of SRGAN models in some resource-constrained scenarios,such as mobile devices,embedded applications,etc.Although SRGANgenerated images can obtain more details,they also have the problem of unreasonable texture detail features due to the use of feature maps to compute Euclidean distances.In this paper,we propose a lightweight improved SRGAN model Shuffle 2 SRGAN,which achieves the reduction of network complexity and optimization of generated image quality by optimizing the network structure and modifying the loss function.The specific study is as follows:(1)Using ShuffleNet-V2 cells as the backbone network to replace the residual cellbased structure inside the original SRGAN generator,and simultaneously constructing a lightweight SRGAN using ShuffleNet-V1 cells in the same way,we experimentally compare the parameters of the generator,the computational effort,and the processing speed and generation quality of the model under 4-fold super-resolution.The results show that the number of SRGAN parameters and computational effort after lightweighting are significantly reduced.The best performance was achieved with ShuffleNet V2 lightening,which reduced the number of parameters and computation by74% and 51% compared with the original SRGAN,and accelerated the reconstruction speed of the same low-resolution image by 20% and 5% on CPU and GPU,respectively.However,the lightweight SRGAN has the highest PSNR decrease of 2.61% in terms of generated image quality.(2)To address the problems of degraded quality of SRGAN generated images after light weighting and the existence of anomalous texture in SRGAN generated images,a loss function based on Ra GAN and Canny edge detection is proposed and the network structure is improved.The results show that the PSNR of the improved network is increased by 2.01% and SSIM is increased by 4.77% compared with the original SRGAN,and the anomalous texture is also well suppressed without any significant change in processing speed.