Research On Single Image Super-resolution Method Based On Gradient Profile And Nonlocal Self-similarity Feature

In the process of digital acquisition and image acquisition,the quality of the images we acquired will be degraded due to various factors such as atmospheric disturbance and defocusing.In addition,the noise introduced in the imaging process will further exacerbate the degradation of the acquired image.With the significant improvement in computer technology and computer performance,people have higher requirements for image quality.However,due to the limitations of environment and hardware,the quality of the images we collected is not as good as expected.As a result,ideas have been developed to improve the resolution of images by improving hardware,software,or environmental methods.If the resolution of the image is improved by improving the hardware,the corresponding cost will be large.If we improve the resolution of the image by improving the environment,it is difficult to achieve in practice,and may cause greater losses.So software is usually used to improve the resolution of the image.The super-resolution algorithm of an image is to generate a corresponding high-resolution image from a known low-resolution image.This paper focuses on super-resolution algorithm of single image.Firstly,this paper lists the popular single image super-resolution algorithms in recent years,focusing on the interpolation-based methods,learning-based methods and reconstruction-based methods,and compares their advantages and disadvantages.The super-resolution reconstruction algorithm based on gradient profile has been a hot topic in recent years.The gradient profile takes into account the spatial layout of the image gradient,which effectively improves the resolution of the image.In addition,the self-similarity of image is widely used in super-resolution algorithms.Based on the research of related algorithms,this paper improves the super-resolution algorithm based on reconstruction:combining gradient profile prior and nonlocal self-similarity prior,a new image reconstruction framework is constructed.Firstly,this paper constructs a gradient diffusion function to improve the gradient direction near the edge.Based on the neighborhood gradient profiles,this paper proposes a gradient profile sharpness optimization function to make the estimated sharpness more accurate.In order to reduce noise in the process of reconstructing HR images,this paper proposes a new non-fixed search method to search for non-local similar blocks,and then constructs new image domain constraint based on non-local feature similarity.In order to suppress artifacts,this paper uses gradient profile prior as the gradient domain constraint.Gradient domain constraint and image domain constraint are alternately performed to constrain the iterative process more effectively,which better guarantees the stability of the algorithm.Finally,this paper designs a high-pass filter function to get the high-frequency portion of the HR image.Shock filtering is only performed on the high frequency portion of the HR image to further enhance the edge detail.This paper verifies the proposed method in terms of subjective perception and objective quantification.The experimental results show that the proposed method has higher PSNR,SSIM and IFC values while obtaining clearer images,which is better than the representative super-resolution algorithm in the past.Especially in the high frequency part of the edge,texture,etc.,the method of this paper can better retain the high frequency information.