Research On SAR Image Denoising Algorithm Based On Non-local Similarity

Synthetic aperture radar(SAR)is a coherent imaging system that can generate high resolution remote sensing images all day.Compared with visible imaging system,SAR can still work under harsh environment such as low light and fog.It also has the advantages of multi-band,multi-polarization and strong penetrating power.Therefore,SAR has become an irreplaceable observation tool in many fields such as environmental monitoring,town planning and disaster assessment.Coherent imaging system causes speckle noise in images,which seriously affects the subsequent processing of the image.Therefore,the research of SAR image denoising is of great significance for image understanding and interpretation.Non-local means(NLM)denoising uses non-local similarity to denoise,which achieves good denoising effect.However,NLM will contain dissimilar patches and increase unnecessary calculation when matching patches,and the calculation of weights is easily affected by patch size.This paper improves the shortcomings of NLM,and combines non-subsampled Shearlet transform(NSST)and weighted nuclear norm minimization(WNNM)to propose three SAR image denoising algorithms based on non-local similarity.The main research work of this paper is as follows:(1)SAR image denoising based on similarity validation and patch ordering in NSST domainIn order to improve the traditional non-local transform domain SAR image denoising algorithm without considering the shortcomings of patch relationship,a new SAR image denoising algorithm in NSST domain is proposed based on similarity validation and patch ordering.Firstly,the density distribution of the distance between similar patches of SAR image in NSST domain is constructed.Secondly,the patches with lower similarity are removed by using the similarity between patches.Finally,the SAR image is denoised by combining patch ordering and optimal one-dimensional filtering.The experimental results show that compared with other classical denoising algorithms,this algorithm can better preserve the image edge and texture information,and improve the visual effect of the image.(2)Speckle suppression based on sparse representation with non-local priorIn order to overcome the problem that coherent noise suppression algorithm based on sparse representation is prone to over-smoothing,a sparse domain coherent noise suppression algorithm based on non-local prior knowledge is proposed.Firstly,the image is sparsely represented by NSST.Then,a sparse representation denoising model based on non-local prior is established by using the non-local prior of the image as the restriction condition.Finally,the denoising model is solved by alternating iteration algorithm to reconstruct the image suppressed by coherent noise.The experimental results show that the algorithm can not only significantly remove speckle noise,but also better retain the texture information of the image.(3)Speckle suppression based on weighted nuclear norm minimization and grey theoryIn order to improve the effect of reference block size and search window size on similar block collection and computational efficiency when using Euclidean distance in traditional image block methods,we use grey theory to improve the weight calculation of traditional non-local method.Firstly,noisy image is logarithmically transformed.Secondly,the image is subjected to local block matching using the grey correlation theory to obtain a set of similar blocks of the reference block,and a matrix with low rank is constructed.Then,the noise variance of the image is estimated using wavelet transform.Finally,the weighted nuclear norm minimization theory is used to denoise.The experimental results show that this algorithm not only effectively improves the visual effect of the denoised image,but also better preserves the local structure of the image.