Image Restoration Based On Properties Enhancement And Mixed Constraints

Digital images are important carriers to transmit information,but they are easy to receive noise,occlusion or pixel loss and other forms of interference in the process of acquisition or transmission.Due to limitation of equipment or time,the original clear image cannot be acquired again,applying image processing algorithms to interference repair of acquired low-quality image and restore potential high-quality images has become an issue of great concern to researchers.Image inpainting technology usually uses information such as low-quality images and the reasons for their formation to restore or reconstruct clear images that eliminate noise,also known as prior knowledge of low-quality images,remove occlusion or lose pixel completions,thus improving image quality and enabling them to be applied to subsequent scenes,such as reognation,classification,semantic understanding and others.For the degradation process of the small sample or single sample image data,the essential properties mining and enhancement of the disturbed image are beneficial to specific data reconstruction tasks.For the restoration task of the single image,mining and enhancing the essential properties embodied in the image data,combined with the prior regularity constraints such as the degradation process of the image,have important research significance for exploring the restoration task of small sample image data.The work done in this paper is as follows:1.Propose an iterative reweighing group sparse constraint reconstruction algorithm for small sample images.Inspired by learning algorithms,aiming at the problem that the existing regression reconstruction algorithm cannot distinguish the importance of samples and cannot remove invalid features in samples,set different weights to separate abnormal samples,and set distance weights and feature weights for reconstruction errors and sample features respectively to solve the optimization coefficient problem.At the same time,propose an efficient solution algorithm to iteratively and adaptively update the weight vector.Convex functions are extended to nonconvex approximation functions,and sparse solutions are approximated with stricter l_2,pnorm.In order to verify the effectiveness of the algorithm,the public face image data sets with sample labels are used to reconstruct and identify the disturbed images.Experiments show that the algorithm has a higher reconstruction recognition effect in images affected by noise interference,color block occlusion or their mixture,which lays a foundation for the subsequent image restoration of the single sample.2.Aiming at the problem of noise elimination and fuzzy restoration of single-sample single-channel images,an image group training with non-convex constraint denoising and deblurring algorithm is proposed.For the case of constructing dictionary set without training samples,the smoothness of images specifically the local smoothness and non-local similarity are used to acquire image blocks with an overlapping sliding window on a single image.And an image block search algorithm with variance constraint between groups is used to avoid the influence of shadows on image blocks with similar structures when Euclidean distance is used to calculate similarity.After the overcomplete dictionary with approximate full rank is constructed,the dictionary is trained in the transposed domain to further reduce the rank of matrices formed by similar image blocks,so as to improve the coefficients'sparsity of inter-group.In addition,the nonconvex l_p-norm constraint is adopted in the reconstruction process to further improve the training efficiency while ensuring the strong sparseness of the reconstruction coefficients.Moreover,in order to adapt to different levels of noise and ensure the robustness of the proposed algorithm,an adaptive soft threshold for image block matching is set up in the dictionary training process.Experimental results show that the algorithm can effectively compare the structural feature information of image blocks and improve the denoising and deblurring performance of a single image.3.For single-sample multi-channel images,a sparsifying transform learning with weighted singular values minimization is proposed for image restoration.To enhance the performance of low rank in image inpainting process,the traditional algorithm is further combined with the transposed domain learning algorithm based on low rank relaxation optimization.In common algorithms,multi-form kernel norm and its evolution form are used instead of low rank to ensure the convexity of the model and find the global optimal solution.This type of algorithm can be broadly called image inpainting algorithm based on the image domain,but its main defect is that it requires a large number of iterations to solve the convergence results.At the same time,as a fast algorithm,the efficiency of the transposed domain learning algorithm is greatly improved compared with the image domain algorithm.In order to pursue the efficiency and restoration effect of the algorithm at the same time,an image restoration algorithm based on sparse transposition and weighted singular value minimization is proposed by taking advantage of the sparsity,smoothness,low rank and transposition of the image.The algorithm combines the high efficiency of the transpose domain with the excellent repair results of the image domain in the same framework.The experimental results show that the PNSR and SSIM values of the proposed algorithm have been greatly improved in image restoration quality,and the robustness of the proposed algorithm is further shown as the noise level increases.In terms of computational efficiency,compared with the traditional image domain algorithm,the time complexity of the proposed algorithm is also greatly reduced,and the convergence condition of the algorithm can be achieved with only a small number of iterations.4.For the problem consistent repair of multi-channel images,a multi-channel repair algorithm for missing elements in color images is proposed.Most of the existing algorithms adopt the operation mode of image segmentation to construct dictionaries and reconstruct them sparsely.Although combining image transform learning algorithm can greatly improve its efficiency,this type of algorithm is still a time-consuming problem on multi-color channels.By exploring the singular value and gradient distribution of RGB channels and the expansion matrix of natural image samples,a matrix repair technology combining multiple channels is developed according to the low rank and local smoothness of images.The fast color image inpainting application can be realized by adopting the local smoothing?piecewise smoothing?characteristic of the image and imposing low rank constraints on it.Furthermore,a convex approximate differential operator is proposed to solve the non-convex non-smooth model.Experimental results show that compared with the tensor patch algorithm or other forms of multi-channel matrix patch algorithm,the proposed algorithm in this chapter has improved patch effect and image visual quality for color images.