Research On Structured Model For Image Restoration

With the development of digital media technology and the popularization of mobile intelligent devices,humans have entered the information age in an all-round way,and image has therefore become the main way for humans and machines to draw information.However,due to the hardware limitations of imaging equipment and the interference of external environment,the final acquired image often suffers from serious quality degradation,which makes it difficult to meet the needs of subsequent applications such as medical diagnosis,target recognition and detection,etc.Therefore,it is of great significance to study restoration algorithm to improve image quality and to provide convenience for subsequent processing.Image restoration is an ill-posed inverse problem,and the Maximum a posterior(MAP)method guided by image statistical prior model can effectively alleviate its ill-posed.Therefore,this paper takes the image restoration method based on MAP as the starting point,and focuses on the following two aspects of research work on the problems of the current image statistical prior model:(1)An image restoration method based on adaptive Gaussian mixture model is proposedIn order to overcome the weak expression ability and the inability to accurately represent image structure information of the fixed Gaussian mixture model(GMM)in traditional spatial model,we propose an image restoration method based on adaptive Gaussian mixture model(AGMM).Firstly,patch matching method based on Euclidean distance is used to find several similar image patches for the image patch to be processed.Then,based on nonlocal self-similarity of image,all similar patches are regarded as the independently identically distributed samples,and the parameters of Gaussian component in GMM of the patch to be processed are estimated.By estimating the parameters of components of all the patch to be processed through above process,the AGMM matching the structure information of the image to be processed can be implicitly determined.Finally,the learned AGMM is applied to image restoration,and Alternating direction method of multipliers(ADMM)algorithm is used to solve the obtained optimization problem to restore the potential high-quality image.The experimental results show that the proposed model and restoration method can protect image structure better than traditional spatial model.(2)An image restoration method based on two-stage orthogonal transform and Laplacian distribution is proposedThe AGMM model proposed in the above work does not perform well for images with rich textures,contours and other structures,while sparse domain model can simplify model and better protect texture,contour and other structural information.Therefore,this paper further proposes an image restoration method based on two-stage orthogonal transform and Laplacian distribution.First,PCA(Principal component analysis)transform is performed on all image patches to remove the correlation within image patch,and first-stage sparse coefficients are obtained.Then,PCA transform is performed on similarity coefficient group to remove the correlation between similar patches,and the final sparse coefficients are obtained.After that,the Laplacian distribution with variable parameters is used to statistically model coefficients,and the parameters of the Laplacian distributions are adaptively estimated from the coefficients to be processed,so that the model and image structure are more matched.In addition,since the vectorization of image patch will destroy the intrinsic structure information,2D Singular value decomposition(2D SVD)and PCA transform are further exploited to remove the correlation within image patch and the correlation between similar image patches respectively,and the Laplacian distribution with adaptive parameters is then used to model coefficients.Finally,the proposed sparse domain models are applied to the image restoration framework based on the maximum posteriori probability and solved by ADMM algorithm,so as to obtain the restored images.Our experimental results prove the effectiveness of the proposed model and restoration method in terms of objective and perceptual quality.