Study On Denoising Method Of Scanning Electron Microscopy Images Based On Deep Learning

Scanning Electron Microscopy(SEM)has the advantages of large depth of field and high resolution,and is widely used in many scientific fields as a powerful tool for observing the microstructure of matter.However,during the SEM imaging process,noise in the equipment circuit can contaminate the image and cause degradation of image quality,which in turn affects the researchers’ accurate analysis of the microstructure of matter.Researchers can reduce the noise by extending the scanning time,but the thermal effects caused by long scanning times of high-energy electron beams can damage samples,especially biological samples.Therefore,the denoising study of noisy SEM images due to low scan time can help improve the image quality and is of great practical importance for revealing the microstructure of matterTraditional denoising methods mostly rely on a priori knowledge of noise types,however,the noise types in SEM images are quite complex,and traditional algorithms often cause blurring of texture features when used directly for SEM image denoising.In recent years,deep learning algorithms have been widely used in the image processing field because of their powerful feature extraction ability.Starting from image feature expression,this thesis proposes two denoising methods based on deep learning for the rich texture features and non-local self-similarity of texture blocks in SEM microscopic images.The specific work is as follows:1.For the feature of rich texture of SEM images,this thesis proposes a multi-task denoising network based on texture inpainting.Inspired by the multi-task learning with shared network parameters,we design a deep learning network containing two tasks,texture inpainting and denoising,and construct corresponding datasets for both tasks.The proposed network first performs the texture inpainting task,and then performs the denoising task,and enhances the ability of the network to extract texture features by repairing the missing texture of the image,thereby assisting the denoising task to retain more texture features.To further enrich the detailed features of the denoised image,we use residual dense blocks to enhance the network’s ability to extract image texture features,and introduce a channel attention mechanism to adaptively adjust the channel weights to restore more image detail features.To balance the two tasks to achieve both noise removal and image texture feature protection,we propose a trade-off loss function,which consists of a texture inpainting loss function and a denoising loss function with different weights,and by optimizing the weights we can obtain the best balance of noise removal and texture inpainting.Finally,a comparison experiment of various denoising algorithms is conducted in the self-built SEM dataset,and the results show that the proposed algorithm achieves better results in both subjective visual performance and objective evaluation indexes.2.For the non-local self-similarity of texture blocks in SEM images,a multi-scale denoising network based on Transformer and CNN is proposed in this thesis.First,considering that the multiscale representation helps the network to capture the features of different sizes in the image,we use the feature pyramid network to perform the multiscale decomposition operation on the input image to obtain the multiscale images.Then,the Transformer module is used to calculate the correlation coefficient between each texture block,so as to determine the similarity of the texture block in the image.Next,given the importance of local information for recovering the image texture,a convolutional neural network(CNN)is used to capture the local information of individual texture block.Finally,we take full advantage of the Transformer and CNN to build a denoising network with the Transformer as the global texture feature extraction unit and the CNN as the local texture feature extraction unit.When recovering a noise-contaminated texture block,the global texture feature extraction unit can use other similar texture blocks to restore the contaminated texture block;when recovering a noise-contaminated pixel in the texture block,the local texture feature extraction unit can use its surrounding pixel information to recover the contaminated pixel,thus realizing the denoising process of SEM images.The comparative experimental results show that the proposed denoising algorithm achieves better results on the self-built SEM image dataset.