Research On Denoising Method Of Sem Image Based On Convolutional Neural Network

Scanning electron microscopy(SEM)is often used for microstructure research in many fields such as biology,medicine,materials,chemistry with its ultra-high resolution.However,SEM images are affected by the noise of the electron beam and the noise in the imaging process,resulting in quality degradation,although operators can reduce the noise by extending the scanning time,increasing the beam intensity and other parameters,the high temperature caused by long scanning and high beam strength can damage the sample.It is necessary to make a balance between parameter adjustment and noise intensity.But in fact,the adjustment of parameters is highly subjective,and the visual fatigue caused by long time shooting increases the uncertainty of image quality,so SEM images will inevitably be introduced into noise,which will cover up the details of images and is not conducive to the subsequent analysis of the micromorphology of the sample.Therefore,it is necessary to study the denoising of SEM images containing noise in order to improve the efficiency of image collection and image quality,which is helpful to promote the accuracy of material microstructure analysis and has high practical significance.Aiming at the problem of SEM image noise,traditional denoising algorithms will lead to the blurring of the edge and texture details of the target image,which will lose many image features,and then affect the microscopic analysis of the sample.Convolutional neural network has powerful feature extraction and generalization capabilities.In this thesis,it is applied to SEM image denoising,which can fully extract the detailed features of the microscopic image and improve the denoising effect of the image.The main research contents of this thesis are as follows:1.Aiming at the problem that traditional denoising algorithms tend to cause SEM image edge blur and detail texture loss,this thesis proposes an enhanced U-net++denoising network based on edge information.First of all,the edge information is extracted by a feature extraction network composed of guided filtering and residual dense blocks,which is used for feature fusion to promote the U-net++ network to generate denoised images with clear edges.In addition,in view of the problem that the conventional L2 loss function is easy to cause the blurring of denoising image,this thesis uses the L2 loss function,the total variation loss function,as well as the perceptual loss function to jointly optimize the network.The L2 loss function is used to restore the structural information of the image,the total variation loss function is used to smooth the flat area of the image,and the perceptual loss function is used to restore the texture details of the image.The experimental results show that the average PSNR of the proposed denoising network is 29.30 db and the average SSIM is 0.76 tested in the SEM microscopic image data set,which achieves a good denoising effect.2.Aiming at the weak robustness of traditional denoising models,this thesis proposes a blind denoising network for SEM images based on multi-scale and attention mechanism.The network uses atrous convolution to construct multi-scale modules to enhance the ability of the network to extract detailed information at different levels.Then the attention mechanism is used to adjust the channel weight of feature map adaptively.Finally,residual learning is introduced to make the network output noise items.According to the SEM image noise density histogram,it can be seen that the SEM noise distribution is closer to the Gaussian distribution,and then the random level of Gaussian noise and the real noise are alternately trained to improve the robustness of the network.The experimental results show that the proposed denoising network achieves good results in both subjective effects and objective indicators.