Study On Quality Assessment Method Of Scanning Electron Microscopy Images Based On Deep Learning

Scanning Electron Microscopy(SEM)uses a highly focused electron beam to scan and image samples,and is widely used for the analysis of material microstructures in many fields due to its advantages of large depth of field,large field of view,and high resolution.However,subjective differences in operator evaluation and visual fatigue from prolonged shooting can lead to deviations in judgments of image quality,resulting in varying degrees of distortion in SEM images that may affect researchers’ judgments of the true structural morphology of the material.Therefore,effective SEM image quality assessment(IQA)methods can help operators obtain high-quality images and provide real and reliable information for studying material microstructures.Traditional natural image quality evaluation algorithms rely heavily on prior knowledge of noise types or artificially designed image features,but perform poorly on SEM images with complex noise types and rich texture details.In recent years,deep learning has been widely used in various visual tasks due to its ability to handle complex nonlinear problems and extract image features.Based on deep learning methods,this article proposes two no-reference SEM-IQA algorithms for studying image quality assessment of SEM images with rich texture details and sensitivity to distortion,as well as non-local self-similarity characteristics of image features.The specific research work is as follows:1.In view of the characteristics of SEM images,which have less edge information,rich texture details,and are extremely sensitive to distortion,and in combination with semantic distribution in latent semantic space,this thesis evaluates the quality of SEM images from both intuitive form and deep semantic aspects.Firstly,a neural network containing a Sparse Mask Module(SMM)is constructed,which generates spatial masks for locating spatial domain texture information and channel masks for locating channel redundancy information through training.Then,sparse convolution is used to skip redundant information,and pure intuitive texture features are selected in two scales: spatial domain and channel domain.Meanwhile,Information Growth Attention(IGA)is introduced into SMM,and the entropy increase of the current feature and past feature of the network is calculated by information theory,and attention to maximize feature map information is generated by progressive convolutional pooling,in order to extract deep semantic features.The experiment shows that the Texture and Semantic Image Quality Assessment(TSIQA)algorithm proposed in this thesis has better performance than mainstream algorithms in typical indicators,exhibits good robustness on datasets with various distortion types,and reduces computing complexity compared to baseline algorithms,while achieving improved running speed on CPU due to skipping redundant information.2.In view of the non-local self-similarity of texture blocks in SEM images,this thesis proposes a SEM image quality assessment method based on the combination of convolution and self-attention(CSIQA).A new Convformer module is introduced,which includes an Atrous Position Embedding(APE)module,a Multi-Head Relation Aggregator(MHRA)module,and a Feed-Forward Network(FFN),seamlessly integrating convolution and self-attention within the Transformer framework.Compared with traditional Transformers and CNNs,the Convformer block uses atrous convolution to embed positional information and broaden the receptive field.It also uses local relation aggregators and global relation aggregators to capture the affinity between tokens locally and globally,avoiding local redundant computation while calculating global dependencies,thereby achieving efficient and effective representation learning.Finally,multiple Convformer blocks are stacked in the network structure,with several convolutions interspersed between them,and mapped to quality scores through pooling and fully connected layers.The experiments show that this method is more capable of obtaining both local and global information of SEM images,and achieves more accurate quality score prediction compared to mainstream algorithms.This thesis includes 28 charts,13 tables and 80 references.