Research On Low-SNR Image Denoising

Image denoising is an important research area in digital image processing technology.Because of various factors,images are often polluted by noise and cause some problems such as low SNR(signal-to-noise ratio)and image blur or the loss of image details,which will influence the image processing or information extraction afterward.In order to remove the noise in the image effectively and improve the low-SNR in image,this paper in view of the basic theory of image denoising and emphasizing on the low-SNR image,has done the research works in image denoising,and has proposed two algorithms as well: actual scene image denoising with convolutional neural network and a method of image denoising with nonlocal weighted least squares estimation.The details summarized as follows:Convolution neural network is widely used in the field of image processing because of its powerful data processing capabilities.This paper is on the basis of elementary structure of convolution neural network and propose a denoising model based on actual scene image.In network model training process considering the masking effect of Human visual system(HVS),namely when human observing images,usually effected by the background illumination and the texture complexity,thus using a large number of smooth actual scene images and analog line images and color block images to build a data set as a constraint of training,so as to realize actual image denoising.Also combine with simulated annealing algorithm for training on the basis of the original training algorithm,to minimize the error function and optimize training model.Experimental results show that for low-SNR images,not only getting rid of the noise greatly but also retain the image details effectively,with improving the SNR value and achieving a good denoising effect.In the algorithm of image denoising with nonlocal weighted least square estimation,in order to estimate the original image from the noisy image,through minimizing the weighted error sum of squares of pixels in image patch to realize denoising.Besides,in order to avoid the borderline blurred problem caused by the local neighborhood pixel information in least square estimation,we establish nonlocal similar patches set according to nonlocal similarity principle so that build intermediate variable and the correlation coefficient that used in estimating.In the operation process of estimation,we calculate the value of the weighted matrix according to the covariance matrix which represents the characterization of correlation of pixels in the image block.Again by referencing the principle of nonnegative matrix decomposition solve the objective function.Eventually we get the final denoised image in combination with all the parameters that solved before.Experimental results demonstrate that the proposed method can improve the SNR value ofimage and effectively achieve denoising performance both in gray images and indoor shooting images of low-SNR.