Research On Blind Image Restoration Based On Sparsity In Transform Domain And Multiple Prior Knowledge

In real life, the degradation of images obtained by the imaging equipment will appearinevitably. It makes the quality of images decline which prevents the follow-up imageprocessing and application. Blind image restoration technology utilizes prior knowledge ofthe degraded image, the imaging system and the original image to build the model.According to the model, the original image is reconstructed.This paper performs the researches on the blind image restoration by studying theprior knowledge and sparsity in different domain, mainly including the following threeaspects.Firstly, researches on effects of using sparsity in gradient domain and in waveletdomain on images restoration are done. According to this, an algorithm of blinddeconvolution based on fusion of regularizations in gradient domain and wavelet domainis introduced. And according to local characteristics of the threshold in gradient domain,the method to select the threshold adaptively based on local features is introduced. Thethreshold is small where is the image edge and big at the flat areas, aiming at keepinformation of image edges. Experimental results show that better restored images can beobtained.Secondly, using the sparsity of dictionary can restore the image texture better. Butonly one dictionary can not be the optimal to represent a whole image. So the orthogonalclassified dictionary is proposed which is learned by PCA. It contains many subdictionaries. And make some improvements to deal with the block effect because of usingdictionary. Blind image restoration based on the classified dictionary is proposed.Experimental results show that the image texture part is restored better.Finally, because the result can not converge to the best one using single regularization,so an algorithm based on using multiple regularizations is introduced. Although theartifacts are removed by using traditional regularizations, the details of images are blurred.A sharpen regularization is proposed to deal with it. Furthermore, an auxiliary variable isintroduced to decompose the question of using multiple regularizations to two simple questions, which reduces the complexity of the algorithm. Experimental results show thatthe restored images are clearer and visual effects are improved.