The Research Of Patch Prior Based Denoising Algorithm

Image denoising is one of the basic problems in image processing.The existence of noise causes many impor tant information to be covered,and seriously interfers with the application value of the image.Recently,many image denoising algorithms have been able to reach a good recover effect when the noise intensity is relatively weak.However,when the noise intensity is strong,the information inside the image is easily disturbed by noise,and then the denoising algorithm could not work well.Since the information can be used only by the prior of the image,the results of the denoising algorithm descents rapidly.Patch prior based image denoising algorithms can capture the statistical char-acteristics of external natural image,and guide internal image denoising.Such al-gorithms can effectively improve the effect of recovered image under high noise si-tuation.In this paper,Gaussian mixture model was used to learn the texture st-ructure of natural image patches,and a low-rank approximation and Wiener filt-ering algorithm based on image patch prior were proposed.The proposed method divided the image into a number of overlapped patches and clustered them for collaborative filtering by using the prior structures of external image patch and internal image self-similarity.By grouping nonlocal similar patches,low-rank a-pproximation was used as collaborative filtering to recover the texture structur-es.When the number of similar patches was small,Wiener filtering with patch prior was adopted to preserve texture features.The experimental results indicat-ed that the proposed method was more suitable for the images with fewer sim-ilar patches like boundary and corner etc.,and showed very competitive perfor-mance with state-of-the-art denoising method in terms of Peak Signal to Noise Ratio and visual quality.Gaussian noise,salt and pepper noise,Poisson noise,are relatively common digital image noises.Existing algorithms usually denoise image only for a single noise level of Gaussian noise.Moreover,many algorithms are based on the assumption that the level of noise variance is known.In this paper,convolutional neural network is used to learn the statistical characteristics of noisy image in order to estimate the type of noise and then a blind denoising algorithm is proposed based on the estimation of the noise type.This algorithm consists of two parts:1)Build multiple training sets with different type noises,and train different denoising models by learning the residual image of noisy and clean images;2)The overlapped patches from multiple different residual images are used as the input of the neural network to estimate the noise type.The denoising results could be obtained by aggregating image patches which are recovered by the denoising model according to the type of noises.The proposed algorithm not only effectively restore the Guassian noise with the single noise variance,but also can remove salt and pepper noise,Poisson noise effectively.