Research On Image Recovery Algorithm Based On Non-Local Similarity Model

Image restoration is aimed to recovery the original image with high fidelity as far as possible.Improving the image reconstruction performance is a hot research topic in the field of image processing.Image restoration is closely related to image acquisition,storage and transmission process,thus the effective image acquisition framework plays an important role in the post image restoration.Compressed sensing,a new theory of signal acquisition,brings revolutionary breakthrough in the image processing.Based on this theory,we can sample and compress signals with far below Nyquist sampling frequency synchronously and can recover signals only with a small number of random measurements perfectly,as long the signal is sparse enough in some spaces.In last few years,compressed sensing has attracted the extensive attention of scholars in the field of image processing.Image restoration,as one of the core issues of compressed sensing theory,has been a research focus in the field.At present,most of the compressed sensing image reconstruction algorithms build target optimization function by using the sparse property of the image signal in a feature space.They do not take the local features and structural properties of signal into account fully,which lead to constraints of the recovery performance and flexibility.For image signals,they have a lot of other properties besides sparse prior under the specific space,such as local features and structural properties.How to use these properties of image effectively and improve the image restoration performance is the research emphasis in this paper.Based on non-local self-similarity(NLSS)model,we researched the image restoration algorithm for compressed sensing.Considering the non-local self-similarity in images,we propose an image CS reconstruction method based on the image low-rank property by converting the CS recovery problem into a matrix rank minimization problem of aggregating similar image patches.The proposed algorithm builds optimization model under the constraint of minimal recovery errors and employs the weighed nuclear norm minimization(WNNM)method to solve the low-rank optimization problem.By taking advantage of them,the proposed method exploits the self-information and structured sparse characteristics of the image very well.Therefore,it provides a better protection of image structures and textures.Experiments on different test images under various sampling rates have shown the effectiveness of the proposed algorithm.Especially,for richly-textured images,the proposed method outperforms the art-of-the-state algorithms significantly under low sampling rates.Furthermore,the traditional image restoration algorithms based on the non-local similarity model always use a simple rectangular shape to complete the image samples extraction and the similarity block matching.It will destroy the structure of the image information,especially the structure characteristics around image edges.We propose a shape adaptive compressed sensing image restoration algorithm based on the image non-local similarity.We use super pixels algorithm to extract the sample blocks for the given image,and use the same shape of the extracted blocks for the similarity block matching.Since using the structure information of a given image effectively,it improves the boundary dependence of the sample blocks.Meanwhile the redundancy of pixels in the same block is higher,which leads to a lower rank for the matrix making up with similar blocks and more useful for CS image restoration.