Several Problems In Image Restoration And Enhancement

Image restoration and enhancement is a representative problem in image processing,which has attracted a lot research attention for a long time.Undesired external conditions are very im-portant reasons being responsible for poor image quality,such as camera shaking,bad weather,unfavorable lighting conditions.These factors make the observed image appeared as blurred images,rainy or hazy images,high-dynamic-range images,noisy images,etc.It is valuable to reconstruct visually clear images from the observed degraded images.Image restoration and enhancement is usually the fundamental low-level processing step in many practical vision systems,which plays a critical role for subsequent vision decisions.It is also of great research value currently.However,most conventional image restoration and enhancement methods still have limitations in terms of complex degradation modelling,descrip-tors of natural image priors,artifacts suppression,etc.The recent deep learning based methods can avoid the hand-crafting process of optimization model and priors.Nevertheless,they usually require training on large datasets,which is computationally expensive.In addition,they work in a way of black box,leading that the insight of their working principle has not been revealed thoroughly.Hence,proposing effective algorithms for reconstructing high-quality clear images from observed degraded images still needs to be resolved with further research efforts.Image restoration and enhancement covers plenty of specific problems.This thesis mainly focuses on three of them:single image deblurring,joint-image-filtering enhancement,and single image de-raining.Three algorithms are proposed to handle with the three problems.The main contents and contributions of this thesis are as follows:(1)Blind image deblurring using elastic-net based rank prior.We propose a non-local im-age prior for blind image deblurring.Non-local self-similarity is an intrinsic property for natural images.We insist that the deblurred image should also satisfy this property.Based on this analysis,we design a novel image prior which exploits the low-rankness property of non-local similar patches with combinations to a strong convex term to enhance the convexity.The proposed prior statistically favors clear images over blurred images,which is beneficial to avoiding trivial solutions.Our deblurring method based on the proposed prior does not require any additional edge selection strategy,and can generate good re-sults.In addition,the prior can be extended to deal with non-uniform image deblurring problem.(2)Structure-Preserving Guided Filtering for Image Enhancement.We propose a guided-image-filtering enhancement algorithm which is able to avoid structure informa-tion loss problem.To address the structure loss limitation near edges of existed guided filtering method,we propose to employ another measurement to form the data term for improving its fidelity.The proposed algorithm is able to generate structure-preserving outputs,which can further avoid introducing unpleasing effects in some enhancement ap-plications,such as errors and artifacts.The proposed algorithm can be applied to many vision problems directly or indirectly,including detail enhancement and high dynamic range compression,flash/no-flash image restoration,image dehazing,and image matting.And it can achieves good performance in these applications.(3)Deep image de-raining network based on squeeze-and-excitation and non-local mean.We propose a deep network for single image de-raining which combines squeeze-and-excitation mechanism and non-local mean operations.Based on the analysis that spatial contextual information in large region is useful for extracting rain components,we propose to embed non-local mean operations in the network architecture for aggregating non-local features and improving the capture of spatial context information,which can further facilitate to extract rain components.We also adopt squeeze-and-excitation on each convolution layer in both encoder and decoder stage to enhance the representation ability of feature maps.Experimental results demonstrate the proposed method achieves good performance in single image de-raining.