Image Denoising Based On Similar Image Patch Clustering And Sparse Coding

Image is an important tool for people getting information.Image can be polluted by different noise in the process of acquiring,storage and transport,leading to the quality degradation.So,in image processing area,image denoising becomes a critical step.Image denoising is the premise work of image edge detection,feature extraction,segmentation and pattern recognition.Synthetic aperture radar(SAR)is one of the most important breakthroughs in remote sensing technology.All day,all-weather imaging capabilities make it highly anticipated from the beginning of its research.It has become the main means of earth observation This technology greatly improves the efficiency of observation,but how to reduce the speckle of SAR image efficiently and accurately is still an urgent problem to be solved.In recent years,more and more algorithms are combined clustering algorithm with the relationship between non local similarity of images,such as: patch based near-optimal image denoising algorithm(PNO).In addition,the sparse representation theory and dictionary learning has become a research hotspot in the field of image denoising,the theoretical basis is that by selecting the appropriate dictionary,using sparse decomposition algorithm to sparsly decompose image blocks in initialized dictionary of image,and then use the dictionary learning and training method to update the dictionary and the sparse coefficients.Finally get the sparse representation of the image and image reconstruction.In this paper,through the study of these two algorithms,and based on this discussion and improvement,the following three image denoising methods are proposed:1.Proposed a SAR image speckle reduction method based on NSCT and the patch based near-optimal algorithm.The patch based near-optimal algorithm through clustering according to the geometric structure of the image,use information relation between image blocks in each cluste,such as the mean and covariance,and then achieve the effect of denoising.Considering the presence of speckle noise in SAR images,it may affect the recognition of similar image blocks.So we use NSCT to get the main detail information(high frequency coefficients)in the image,so we can better identify similar blocks and classify more accurate,which can better retain the details in the process of the image denoising,smooth homogeneous region.2.Proposed an dictionary learning image denoising method based on steering kernel regression and clustering ensemble.The feature vector of each image block is obtained by using the steering kernel regression weight(SKRW),and then the feature sets is classified by clustering ensemble based on center matching.In order to retain more texture details in denoising,we use principal component analysis to get the PCA dictionary of image patchs in each cluster which can be used as initial dictionary,finally through using improved dictionary learning to denoise image and reconstruct image,and then get the denoised image.3.Proposed a new method of SAR image speckle reduction based on joint sparse coding under Gaussian scale mixture model and texture enhancement.Joint sparse coding under Gaussian scale mixture model can transform each sparse coefficient into a Gauss distribution,through describing all the sparse coefficients of the similarity blocks which have similar prior distribution(such as the same nonzero mean,scale variables)can effectively utilize the dependencies between local and non local sparsity coefficient,which can make the effect of denoising better.In addition,utilizing the improved gradient histogram preservation(S-GHP)algorithm estimate the gradient histogram of SAR image,and then this gradient histogram is used as a reference,we use the gradient histogram of current image to constrain the image in each iterative process,which can effectively keep more detail information in class in the process of denoising.