Research On Background Refinement Image Restoration Method Based On Dynamic Deblurring Prior

With the application of computer vision,how to restore the missing areas of the image with high quality and efficiency has become a hot field of increasing concern.This image repair problem faces important challenges:The reasons for blur images and videos through natural and man-made are very com-plex.In real life,when the light is dim or foggy,rainy,windy,the natural image or frame may become blurry.For man-made reasons,different choices of aperture size and focal length may lead to Gaussian blur.Human operation errors,camera jitter and complex scenes of moving objects can lead to all kinds of blurs.Therefore,the cause of natural vagueness is passive,while the reason of artificial vagueness is active.Aerial images or daily images encounter bad weather(rain,fog,clouds)in the long-distance multi-angle shooting process,resulting in blurred images.Artificial camouflage military targets or civilian facilities encounter occlusions in the shooting process,and specific targets are hidden.With the gradual development of deep learning in the field of computer vision,many studies use a single deep learning neural network model to repair low-quality images.These methods only focus on the structural features of objects and ignore important at-tributes such as background and texture color,which often have multiple artifacts or color distortion,resulting in poor image restoration.In order to solve the above problems,this paper proposes a background refinement and feature fusion network with dynamic blur removal leader,which takes dynamic blur removal as a priori,while extracting and removing blur features and image background texture structure features.Combined with the multi-path refinement framework of the multi-path context attention module,the image details are restored deeply from coarse to fine.In view of these three challenges,the solutions and contributions of this paper are as follows:(1)This paper puts forward the background refinement theory and demonstrates its significance and role in image restoration in detail.This paper puts forward a theory about the relationship of images between panoramic and local,foreground and background.By initially removing blur from the background,or by regularizing and refining according to certain rules,and then filling the real part of the internal block,the structural similarity of image restoration and the visual effect after restoration will be significantly improved.(2)The dynamic selection of kernels is developed to remove the image background blur.The specific blur kernel is dynamically selected according to the blur type of different regions of the image,and the multi-path context attention module is used to repair the image a priori in order to obtain the image with sharp edges,which is called background refinement.In addition,lightweight and residual network parameters are used to reduce the model parameters.(3)With the help of the attention mechanism of background edge reconstruction,it is beneficial to the repair of image structure.In view of the internal loss of the image,in order to improve the correlation between intelligent filling and repair of blocks of different scales,this paper uses multi-path similar filling algorithm to improve the texture details of blocks,which is called foreground filling.(4)This paper designs a multi-path refinement network which can be used for rough image repair and background refinement.The result of the foreground filling of the back-ground refined core is integrated by the multi-path fusion module,and the final high-quality image is obtained.(5)This paper designs the joint training function to optimize the training conver-gence speed and model reasoning accuracy in the process of image restoration.In this paper,the joint training mode is adopted to integrate the image background refinement and foreground filling into a whole through the cascade network structure,which can effectively reduce the model memory and improve the testing speed.