Multi-view Feature And Convolutional Neutral Network Based Prior Information For Image Restoration

Image restoration is one of the most classic problems in computer vision processing.In the past half century,researchers have been studying image restoration in various applications due to its importance.However,it is always inevitable to be affected by noise pollution and other factors in image processing.The essence of image processing is to preserve the fine structure and texture information while removing noise as much as possible.It is well known that the image restoration is an underdetermined and inverse problem.It is one of the most important to expoit image prior information in critical technologies.The multi-feature image prior information is widely used in the image processing,and it is also a key point in the image restoration.Based on the image restoration algorithm theory and convolutional neural network model,this paper studies the traditional mathematical model and deep learning network model on various applications:(1)For image completion,we first formulate several multi-view groups by convoluting the target image with different multi-filters groups.Then,each multi-view group can be a regarded as an image sequence,whose similarity-grouped cube set is termed as a low-rank tensor.Next,we obtain the recovery results from the low-rank tensors subjecting to different multi-view groups,followed by an aggregation step that simply combines the recovery outcomes.Experimental results demonstrate that the proposed algorithms can faithfully recover image and outperform the current state-of-the-art approaches in terms of visual inspection,the quantitative peak signal-to-noise ratio(PSNR),and structural similarity(SSIM).(2)For magnetic resonance image denoising,this work captures the pixel-level and feature-level distribution information by means of supervised network learning.A progressive network learning strategy is proposed,via fitting the Rician distribution at pixel-level and feature-level.The first and second sub-networks play the role of crude and refinement estimation,respectively.Due to the nonlinear property of Rician noise,we reasonably arrage the ResNet and BN layer,which consider that BN layers can accelerate network learning and boost accuracy by normalizing the weights and parameters.Meanwhile,BN may get rid of range flexibility from networks by normal-izing the features.By discussing and analyzing the variant network,we demonstrate the rubost of network.Finally,experimental results demonstrated that the proposed network obtained improved PSNR values and preserved more edges and structures than the conventional denoising methods.(3)For nature image restoration,we first select the denoising autoenchoders network as prior information.At the network training stage,the multi-channels prior information is obtained via the relevance among R,G,B channel of color image training sets.According to the aggregation principle,multi-models weighted based network is developed for grayscale IR.After formulating the mathematical model,we adopt the alternative optimization and proximal gradient method to tackle the non-convex grayscale IR minimization.And at the iterative IR stage for grayscale image,auxiliary variables technique is applied to embed the three-channel prior into the single-channel intermediate solution.In summary,this paper focus on exploiting image prior information and proposes three image restoration algorithm/convolutional netural models in image completion,magnetic resonance image denoising in medical imaging and image restoration application.And the proposed achieved better performance,and it effectively balances computational time complexity and restoration performance to obtain a good implementability.