Analysis And Development Of Fidelity And Regularization Learning Based Image Restoration Techniques

Abstract

Image restoration, which seeks to restore a damaged observation's underlying clean image, is a basic challenge in low level vision. The majority of extant non-blind restoration approaches are predicated on the knowledge of a specific deterioration model. Due to the fact that the deterioration process can only be partly understood or correctly represented, photos may not be recovered completely. Two notable instances of such tasks are the elimination of rain streaks and picture deconvolution using inaccuracy blur kernels. Although an input picture may be divided into a scene layer and a rain streak layer for the purpose of rain streak removal, there is no clear formulation for modeling rain streaks and their composition with the scene layer. Due to the estimate error generated by the blur kernel in blind deconvolution, the following non-blind deconvolution step does not adequately recover the latent picture. In this article, we offer a principled approach for picture restoration using a partly known or erroneous deterioration model inside the maximum a posteriori framework.