Research On Image Denoising Technology Based On Convolutional Neural Network

Image denoising is still an active and challenging research topic in the field of computer vision,and also has been widely used in many high-tech industries such as video perception and intelligent analysis,remote sensing imaging and aerospace image analysis.Among many denoising methods,convolutional neural network(CNN)has made a huge breakthrough in the field of image denoising with its powerful feature learning capabilities.However,CNN-based image denoising technology still has the following two crucial problems to be solved: First,the essence of denoising is to retain the information of the image itself as completely as possible while removing the noise.The image will still inevitably lose the image details.Second,the CNN-based image denoising technology performs well on additive Gaussian white noise(AWGN)images,but does not perform well on real noise images,even worse than traditional methods.Regarding the problem that image detail information will inevitably be lost after denoising,first,reserching the image denoising technology based on residual learning,exploring the difference between noisy images and clean image pairs,ie residual images,and finding the residual image contains not only most of the AWGN,but also rich image detail information,which leads to a serious loss of detail in the denoised image.Then,a loss minimization problem is proposed;finally,a mathematical model and an end-to-end detail retaining convolutional neural network(DRCNN)are built by a analysis of the minimization problem.DRCNN has two functional modules: a generation module(GM)and a detail retaining module(DRM),where the GM module separates noise from the noise image to generate an intermediate feature map(IFM);the DRM is used to learn a large number of image detail features which lossing in the IFM,so that the output image of DRCNN can retain the image detail information.Unlike most CNN-based denoising technologies,DRCNN not only focuses on image denoising,but also focuses on the integrity of high-frequency image content.At the same time,DRCNN requires less parameters and storage space.In addition,DRCNN can also be adapted to different image restoration tasks,such as blind image denoising,single image super-resolution(SISR),blind deblurring,and image removing watermark.The experiment results show that,the peak signal to noise ratio(PSNR)/dB and structural similarity index method(SSIM)values of DRCNN on standard public datasets BSD68 and Set12 with noise levels of 15,25,and 50 can reach 31.74 dB / 0.8975,29.26 dB / 0.8467,26.29 dB / 0.7323 and32.88 dB / 0.9401,30.56 dB / 0.8948,27.29 dB / 0.8190,which is superior to the compared classic and novel methods.For the problem that CNN-based denoising method performs well on AWGN images,but does not perform well on real noise images,first,exploring the difference between the gray histogram distribution of AWGN and real noise images,and finding the pixel distribution of the real noise image is significantly more complicated;second,explorating deeply of the attention mechanism,multi-path network,and feature fusion methods,which are more conducive to learning rich pixel features;finally,establishing an end-to-end blind denoising network EDFNet based on the attention mechanism,which can effectively learn complex pixel distribution features.EDFNet consists of a weighted feature extraction module(E),a multipath residual dense module(D),and a multiscale feature fusion module(F).E adaptively adjusts the importance of the channel to extract more discriminative underlying pixel features.Unlike DenseNet,which is widely used in advanced computer vision tasks,D includes densely connected layers,local feature fusion(LFF),and local residuals learning(LRL)and can fully extract hierarchical spatial features.It is very suitable for image restoration tasks.Different from simple feature fusion methods like pixel addition and channel splicing,F uses spatial attention mechanism to adaptively weight and fuse features of different scales to highlight the spatial and channel specificity of multi-scale features.EDFNet's PSNR / SSIM value on the real image denoising test set Darmstadt noise dataset(DND)can reach 36.10dB/ 0.9019,which is an increase of 3.67 dB / 0.1119 over the compared blind denoising method,and an average increase of 1.79 dB / 0.0536 over the non-blind denoising method.At the same time,it shows outstanding visual effects and has greater practical application value.