Application Research On Image Deraining And Desnowing Based On Manifold Learning

Video or image taken in bad weather(rain,snow,sputum,etc.)seriously affects its research use by people.The image degradation caused by rain and snow weather makes the visual system unable to provide reliable feature extraction,object recognition and target detection,tracking and other computer vision algorithm processing results.Therefore,it is necessary to use image rain and snow technology as a preprocessing part of the image,and improve the accuracy of computer vision algorithms.Video or image,affecting by rain and snow,has the characteristics of poor visual performance and key information coverage.All in all,the research on image deraining and desnowing algorithm has great practical significance.The research content of this paper mainly aims at rain and snow streaks removal of video sequence images,for improving the visual effect of video image.The image deraining and desnowing algorithm has been widely applied and intensive research at home and abroad,it achieved good rain and snow streaks removal effect,but there is still the problem of losing the inherent information inside image.Rain or snow of video or image usually involves the transformation of data from high-dimensional nonlinearity to low-dimensional linearity.Rain and snow streaks can be marked by low-dimensional embedded manifolds in high-dimensional observation space,which preserves the structural information of rain or snow streaks.In this paper,some classic image deraining and desnowing algorithms are deeply studied and compared with.According to the knowledge above,two methods of image deraining and desnowing are proposed: image deraining and desnowing method based on improved BPDN algorithm,and image deraining and desnowing method based on local manifold structure.1)The work of this paper is based on the improvement of Kim algorithm.The Kim algorithm only considers the global information of rain and snow images,but lacks the description of the local information of rain and snow images.Firstly,this article considers the rain or snow of video or image as an image denoising problem.Secondly,a local feature constraint(manifold learning derived branch)is introduced in the refined rain or snow map model of Kim algorithm to extract more effective rain and snow streaks.Thirdly,the refined rain or snow map model with local feature constraints is deduced into a standard boundary-constrained quadratic programming problem by formula,and solved by BPDN algorithm.Finally,the refined rain or snow map and the low rank matrix complementation technique based on the EM algorithm are used to remove rain and snow streaks in image.The experimental results show that the refined rain and snow map of this paper extracts rain and snow streaks the more effective than the Kim algorithm,and can achieve the better removal effect.2)The image deraining and desnowing method based on the improved BPDN algorithm is proposed in this paper,it has a good effect of deraining and desnowing,but takes a lot more time.For solving the problem,this paper chooses a better method by constructing edge weight.Firstly,the edge weight matrix of the image pixel point neighbor graph is solved by using the LDMM algorithm based on the manifold learning theory.Secondly,the local manifold structure item,which is obtained by the use of the Laplace feature map(LE algorithm),is introduced into the refined rain or snow map model.Thirdly,the gradient descent method is applied to solving the optimization model,and naturally to obtain a more refined rain or snow map.Finally,the refined rain or snow map and the low rank matrix complementation technique based on the EM algorithm are used to remove rain and snow streaks in image.The experimental results show that the proposed algorithm greatly reduces the time cost,and achieves the better effect of image deraining and desnowing.The algorithm of this paper preferably preserves the inherent information of image after removing rain and snow streaks.