Research On Image Denoising Algorithm Based On Frequency Domain Enhancement And CNN

With the advent of the information age,the information people receive is no longer a single voice signal,but more image signals.It is the main way for humans to spread and communicate information.However,image information will be interfered by various noises in the process of transmission,and noise will cause the loss of image details and blur the image.Therefore,image denoising has always been a research hotspot.At present,a large number of image denoising algorithms have been proposed and achieved good results.However,it is still the key to image denoising problem to keep more edge details of the original image while removing noise.Aiming at this problem,this paper starts from preserving the edge details of the image,and combines frequency domain enhancement and convolutional neural network to study the problem of image denoising.The specific work content includes:(1)An image denoising algorithm based on NSCT domain and residual network is proposed.First,use the Non-local Means(NLM)denoising algorithm to pre-denoise the noisy image,and then perform Canny edge.detection on the result to obtain the edge matrix;at the same time,perform non-subsampled contourlet transform(NSCT)decomposition of the original noise image to obtain its high-frequency sub-bands;locate the edge pixels through the edge matrix position,amplify the corresponding edge coefficients in the high-frequency sub-bands,and then inversely transform the NSCT to obtain the edge-enhanced noise image;finally,the edge-enhanced noise image is input to the residual network training to obtain the final denoised image.Experiments show that this method achieves a good denoising effect,and the denoised result can retain more edge texture details.(2)An image denoising algorithm based on NSST domain and Res2Net network is proposed.First,use the mean filtering algorithm to pre-denoise the noisy image;perform non-subsampled shearlet transform(NSST)transformation on the pre-processed image to obtain a low-frequency sub-band image and multiple high-frequency sub-bands;Each high-frequency sub-band and low-frequency sub-band are respectively input into Res2Net for training and learning;Finally,the trained high-frequency sub-bands and low-frequency sub-band are reconstructed by NSST inverse transformation to obtain the final denoised result.Experiments show that this method can obtain clear visual effects and higher indicators.