Research On Image Super Resolution Based On Generative Adversarial Network

Image super-resolution algorithm is one of the important technologies in the field of area enhancement.The specific process is appling image processing methods to a low-resolution image and convert it into a high-resolution image.While high-resolution images contain more information,super-resolution technology is of great significance in terms of imagestransmission,remote sensing image enhancement,traffic surveillance,medical diagnosis,and etc.The research on image super-resolution reconstruction began in the 1960 s.After decades of development,methods based on interpolation,reconstruction,and deep learning have emerged successively.This paper improves the existing deep learning based super-resolution reconstruction methods.The thesis proposes a new super-resolution algorithm MRA-GAN,which reduces the amount of network parameters with Cycle GAN.Moreover,we design one image transmission systembased onsuper-resolution technology.The main research contents of this article are reflected in the following three aspects:(1)An improved multi-level residual network MRAN based on attention mechanism is proposed for super-resolution reconstruction.By constructing the attention module,MRAN realizes the redistribution of attention resources,so as to improve the feature utilization and learning ability of the network.Through multi-level residual structure and residual aggregation,MRAN improves the feature extraction capability of the networksignificantly.Experimental results on five benchmark test sets show thatreconstruction results of MRAN are better than that of existing other methods while reducing the amount of network parameters.(2)Based on MRAN,a new super-resolution reconstruction network named as MRA-GAN is proposed by applying Cycle GAN.Compared with SRGAN,the first image super-resolution algorithm that applies generative adversarialnetwork,MRA-GAN applies MRAN as generator,which improves learning ability and feature extraction ability of network significantly.The proposed MRGAN uses relative discriminators,real images can also play a positive role in trainingprocess;Improves the perceptual loss function,uses the feature map before activation for comparison,and improves the utilization of the feature map;Adopts a double loop network structure to improve the utilization of data samples.At last,MRA-GAN based super-resolutionimproves learning ability and perceived quality of network of the reconstructed image without increasing the quantity of network parameters and data set size.Experimental results show that the reconstructed image obtained by MRA-GAN algorithm has improved quality compared to SRGAN and MRAN.(3)Based on Qt application development framework and proposed super-resolution technology,a low-bandwidth ultra-high-resolution image transmission system is designed.The sender of the system transmits the down-sampling image file,and the super-resolution reconstruction model is deployed on the receiver to complete reconstruction process of the received low-resolution image.In case of poor network conditions,the transmission speed of the image will be significantly increased.