| With the development of image processing,deep learning and computer vision,high-resolution digital images are gaining more and more attention and are widely used in many fields such as biomedicine,public safety and intelligent transportation.Due to the limitations of storage space,transmission rate and equipment imaging capability,image quality degradation is unavoidable,therefore,it is necessary to use image super-resolution reconstruction technology to obtain high-resolution images.The image super-resolution reconstruction technology not only meets human requirements for high quality images,but also does not require replacement of imaging equipment and does not add additional economic costs,and has received widespread attention in the industry.In this thesis,we will study the image super-resolution reconstruction algorithm based on Generate Adversarial Networks(GAN),and the main contents includes:1.Conduct research on degraded models.The classical degradation model only includes a fixed degradation process,which is difficult to simulate the complex degradation process in the real world.In order to synthesize more practical low resolution images,a higher-order degradation process is introduced.At the same time,a sinc filter is added to simulate the ringing and artifacts commonly seen in reality,so as to better simulate the complex degradation situation in the real world.2.When performing super-resolution reconstruction on damaged low resolution images,there are problems such as boundary distortion in the damaged area.Therefore,Edge-ESRGAN based on edge information reconstruction is proposed,which is mainly composed of an edge generation module and a super-resolution reconstruction module.The edge generation module reconstructs the edge structure of high-resolution images based on known low-resolution images and corresponding edge images,and generates high-resolution image edge image.The super-resolution reconstruction module utilizes high-resolution edge image as prior guidance conditions,zooms in while filling in texture details to generate the final super-resolution image.The Edge ESRGAN method proposed in this article improves the structure of damaged LR images,enriches the texture details of low resolution images,and improves the quality of high-resolution images.3.The Real-ESRGAN model has good reconstruction performance for undamaged low resolution images,but only obtains the overall image information,without considering the local details of the image,resulting in insufficient acquisition of local details.The DU3-Real-ESRGAN super-resolution reconstruction algorithm is proposed to address this issue.Replacing the U-Net structure of the Real ESRGAN discriminator with the UNet3+ structure improves computational efficiency while capturing more image details at full scale.Introducing a dual discriminator structure,where one discriminator learns details from a complete synthesized image as input,paying attention to the edge features of the image.Another discriminator takes down sampled synthesized images as input and is responsible for mastering the global view.The DU3 Real ESRGAN network proposed in this article performs better in processing complex images,removing noise and artifacts while restoring richer detail information and distinct image edges,resulting in better visual super-resolution reconstructed images.4.Implement Web development of super-resolution reconstruction algorithms based on front-end and back-end separation technology,improve development efficiency,and ensure data security.The front-end is responsible for rendering and interacting with pages,while the back-end focuses on handling business logic.The front-end completes data interaction tasks with the back-end through network requests. |
