Image Denoising Method Based On Attention Mechanism And Multi-scale Model

As an important visual information medium in modern society,digital image is widely used in daily life.However,due to the interference in the process of acquisition,recording and transmission,digital image inevitably has a lot of noise.The existence of these noises is not conducive to the processing and analysis of digital images.Therefore,the research of digital image denoising method has important practical significance.The current image denoising methods have the following two problems to be solved :(1)How to avoid the loss of image details in the process of image denoising;(2)Most image denoising methods have a good denoising effect on additive gaussian white noise,but the denoising effect on real noise needs to be improved.Aiming at the above problems,this paper introduces attention mechanism and multi-scale method to study image denoising.The research contents are as follows:First,in view of the problem of detail loss in the process of image denoising,this paper firstly analyzes the causes of image detail loss in the process of denoising,and proposes a mathematical model of detail retention.Based on this model,an image denoising network OGFNet with detail preserving module is proposed.The network consists of RSGBlock module and ORSNet module.Among them,RSGBlock module is used to learn residual image,ORSNet module is used to eliminate image details in residual image.The attention unit SE and the gated loop unit CVGRU are introduced in RSGBlock to adjust the importance of channels and obtain more global features.The ORSNet module is composed of three original resolution blocks(ORBs)in series to obtain high resolution features.The denoising performance of OGFNet at gaussian noise levels of 10,30,50 and70 was tested on CBSD68,Kodak24 and Urban100.The average PSNR of OGFNet on the CBSD68 at 10,30,50 and 70 gaussian noise levels were 36.55 d B,30.71 d B,28.36 d B and26.96 d B,respectively.The average PSNR of OGFNet on the Koda24 at 10,30,50 and 70 gaussian noise levels were 36.55 d B,30.71 d B,28.36 d B and 26.96 d B,respectively.The average PSNR of OGFNet on the Urban100 at 10,30,50 and 70 gaussian noise levels were36.55 d B,30.71 d B,28.36 d B and 26.96 d B,respectively.The results show that the denoising performance of ORSNet in the above three datasets is superior to that of comparison method,and the image details can be well preserved.Second,aiming at the real noise image denoising problem,this paper proposes an improved Transformer image denoising network RFNet.The network is composed of multi-scale denoising module and feature fusion module,which can realize end-to-end denoising of real noise in image.The multi-scale denoising module adopts pyramid structure,and more global features can be obtained through different degrees of down-sampling of images.The feature fusion module uses convolution kernels of different sizes in each channel to fuse the features obtained in the multi-scale denoising module,and finally outputs the denoising image.The results show that RFNET-5 has good noise reduction effect on Poly U,SIDD and CBSD68.The average PSNR value of RFNET-5 is 37.08 d B on the dataset of Poly U.The average PSNR value of RFNET-5 is 40.94 d B on the dataset of SIDD.The average PSNR values of RFNET-5 at 15,25 and 50 Gaussian noise levels are 34.91 d B,34.74 d B and28.16 d B on the dataset of CBSD68,respectively.The results show that RFNET-5 can achieve good denoising effect in both real and Gaussian noise denoising tasks.