Sparse Image Denoising Method Based On Example Learning

Due to the influence of atmospheric turbulence,sensor noise,space dust and object high speed moving,leading to obtained images from ground telescope appear more serious degradation.To reconstruct high quality image from degraded data,different methods have been proposed by scholar at home and abroad in image deblurring,image super-resolution and image denoising.According to various types of noise,the scholars put forward different methods of denoising,moreover,practical applications and ways have been achieved.To reduce the influence of denoise,sparse representation is combined with the character of spatial object image for our study.The topic of this paper comes from the national key laboratory project: "Research on spatial XXX target image fusion method".According to the study,we found that the presence of the target space satellites have the following characteristics:(1)the number of goals on orbit is limited.(2)the majority of the target satellite known the shape;(3)the different satellites have similar parts and shape features.Based on prior information imaging target,an image denoising method based on example learning sparse is carried out,similar section of these image noise in high-quality image library is taken full advantage to improve over-complete dictionary learning.Two image sparse denoising methods based on a sample of similar are proposed with sparse denoising model.1.Sparse denoising algorithm based on local similarity learning.SIFT key points around the n×n image blocks are taken as a learning sample for the algorithm.Firstly,the SIFT algorithm is used to extract the key points and the corresponding feature descriptor of each image in the image database,then,points and descriptor are used to construct the SIFT feature group;secondly,SIFT key points and the corresponding feature descriptor from noise image are extracted,meanwhile,finding similar matching in SIFT feature group;finally,all the obtained local image blocks are used as the sample input for the over-complete sparse dictionary learning.Because the proposed algorithm improves the correlation between the learning sample and the noise image,to a certain extent,it can enhance the local information represent ability of the over-complete sparse dictionary,it can get better denoising effect.2.Sparse denoising algorithm based on regional similarity learning.Firstly,SIFT and MSER algorithms are used to extract the SIFT feature descriptor and the MSER region of each image database to construct the SIFT-MSER feature cluster database;secondly,SIFT-MSER feature clusters are extracted,meanwhile,finding similar matching in SIFT-MSER feature cluster database;afterward,all MSER regions that matched SIFT-MSER feature cluster are used as the sample input for the over-complete sparse dictionary learning;finally,denoising image is get from the sparse decomposition and reconstruction of noisy images are obtained by using the over-complete sparse dictionary.The algorithm dictionary learning has both local and global information,and it can improve the PSNR of the image in the process of high frequency compensation.