Image Restoration In Complex Secnes Combining Model-based Optimization And Deep Learning

Image restoration in complex scenes is to restore the image that is photographed under poor conditions,such as bad weather and low light,which is of great significance for enhancing Intelli Sense in complex scenes.Existing image restoration algorithms can be divided into model-based optimization and deep-learning methods: Model-based optimization methods fully model the image degradation and knowledge characteristics,but the manual designed priors are always inaccuracy.Deep-learning methods do the opposite.Due to the problems of the complex degradation characteristics and multiple coupling degradation in real complex scenes,it is difficult to achieve satisfactory results by using either the above method alone.In this paper,we make full use of the complementary advantages of model-based optimization and deep learning,and proposes two combination methods including designing networks inspired by optimization and deep network instead of manually designed prior for image deraining and noisy image deblurring,respectively.The shape of rain streaks are complex and diverse,and it is difficult to distinguish the similar textures between rain streaks and images.To address the problems,we propose a progressive guidance network inspired by the image decomposition model for image deraining.We extract the principal direction of rain streaks as a discriminative feature,and develop the progressive guidance network inspired by the decomposition model,which jointly model the rain layer and the image layer.The network consists of three mutually guided stages,which successively estimates the univariate direction,sparse rain layer and dense image layer in a simple-to-complex and coarse-to-fine way,and decorrelate the rain layer and background image.The qualitative and quantitative experiments show that the progressive guidance network inspired by the decomposition model is superior to other comparison algorithms.To solve the problem of insufficient expression of image features in the progressive guidance network,in this paper,we develop the residual-aware block inspired by iterative regularization,and improve the progressive guidance network by residual-aware block to achieve better performance in image deraining.Since the iterative regularization retrieves residual information in iterative optimization,we construct the residual-aware block inspired by iterative regularization to adaptively perceive the information lost in convolution,which enhances image modeling and expression capabilities.Extensive comparison experiments have proved that the residual-aware block can effectively suppress the rain streaks and preserve the image structure.To address the coupling problem of noisy and blurry image,in this paper,we develop deep blind denoising prior,and propose the iterative framework based on optimization and deep blind denoising network for noisy image deblurring.Taking advantage of half-quadratic strategy,we split the coupling problem into two iterative sub-problems including deconvolution and denoising.Then we propose a two-stage blind denoising network inspired by optimization,which is incorporated into the iterative framework to solve the denoising sub-problem.The deep blind denoising network adaptively learns the prior and balance parameter,avoiding tedious manual parameter adjustment and under-denoising or over-smoothing effects caused by the inaccurate balance parameters.Extensive comparative experiments have verified the superiority of the iterative framework combining optimization and deep blind denoising network.