Image Super-Resolution Based On Image Gradient And Generative Adversarial Networks

Image Super-Resolution(SR)refers to recovering a given low-resolution image into a corresponding high-resolution image through a specific algorithm.High resolution images has the ability to provide more details.Hence,they often play a key role in many applications.To obtain high-resolution images,a straightforward method is to use high-resolution image sensors.Nevertheless,due to the complex manufacturing processes and expensive costs of the sensors,it is difficult to deploy in real-world occasions.Therefore,it is of great practical significance to obtain high-resolution images through super-resolution technology using existing equipment.However,the existing image super-resolution algorithms still have the following problems:(1)The reconstructed high-resolution image cannot well recover the high-frequency(structure)information in the super-resolved images,resulting in poor visual effects;(2)The existing reconstruction models do not consider detailed information and hence deteriorate the authenticity of images.Besides,it is very non-trivial to directly reconstruct large-scale resolution images(such as 4x super-resolution or 8x super-resolution).In order to solve the challenges of super-resolution technology,we propose the corresponding improvement solutions for each problem.Specifically,for the problem(1),we propose a novel loss function based on image gradient to ensure that the super-resolution model pays more attention to the structural information in the image.For the problem(2),we introduce a generative adversarial network(GAN)to generate the details in the high-resolution images in order to improve the visual fidelity of the images.In addition,to ensure the performance of super-resolution model,we build a progressive reconstruction scheme,which gradually improves the image resolution scale during training and inference.Extensive experiments demonstrate that the proposed methods in this paper effectively improve the performance of the super-resolution model.