Study On No-Reference Quality Assessment Of Scanning Electron Microscopy Images

Scanning electron microscopy(SEM)is a commonly used instrument for observing the microscopic world.In practical applications,it is necessary to obtain high-definition SEM images by constantly adjusting imaging parameters.This process is time-consuming and laborious,and the effect is different due to the fatigue degree of human eyes.Therefore,an automatic objective quality evaluation algorithm is needed to guide the selection of imaging parameters for SEM.In this thesis,the subjective quality evaluation of SEM images is studied.Two no-reference quality metrics are proposed for the evaluation of blur distortions and contrast distortions in SEM images.The main research work is as follows:First,the research of image quality evaluation aims to design image quality metrics closely related to human subjective scoring.In order to study the quality evaluation of SEM images,this thesis first establishes a database of SEM images.A total of 650 SEM images with different types of distortions are collected from the Advanced Analysis and Computation Center in China University of Mining and Technology.There are 50 different groups of image contents,and each group includes 13 images.The distortions of these images are generated in the shooting process.Then,an interactive platform is established with MATLAB,which allowed 30 subjects to conduct subjective experiments on SEM images under normal conditions.And 30 subjects learned how to use it.Finally,the experimental data obtained from 30 subjects is processed.The outliers are eliminated according to the confidence interval,and the remaining valid values are averaged to obtain mean opinion score(MOS)of each image.The MOS value has important guiding significance for studying the no-reference quality evaluation of SEM images.Second,since the SEM images are grayscale images,and their background and the surface of the displayed objects are dim,which results in low channel values in the dark channel of the SEM images.In addition,the dark channel prior is very sensitive to the image changes caused during the blurring process.Blurred images have less dark channel information,and different blurred images have obvious differences in dark channel maps.Inspired by the above facts,a method based on dark channel prior for evaluating the sharpness of SEM images is proposed.Firstly,the dark channel map of the SEM blurred image is extracted.Secondly,the edge of the dark channel map is extracted because the edge expansion is usually used as the blur feature.Then the edge is preserved and denoised by the edge-preserving filter based on weighted least square.Finally,the pooling of the maximum gradient and the average gradient is taken as the final sharpness score.The experimental results conducted on the SEM blurred image database show that the proposed algorithm can accurately evaluate the quality of SEM images and maintain high consistency with the subjective results.Third,the human visual system has the multi-scale characteristics in perceiving visual scenes.In this thesis,a no-reference quality evaluation model of contrast distortions of SEM images is proposed based on the multi-scale representation.Firstly,the singular values of contrast distorted SEM images are calculated on four scale-space images,and the similarity of singular values are calculated as a group of features.Then,the one-dimensional entropy and two-dimensional entropy of the contrast distorted SEM images are extracted to reflect the aggregation characteristics and spatial features of the gray distribution.In addition,combined with the features of Log-Gabor filter response,a total of 10 effective features are extracted.The extracted features and the MOS values are combined to train the regression model using support vector machine(SVM).Finally,the regression model is used to predict SEM contrast distortion image quality score.The experimental results show that the proposed method can effectively evaluate the quality of contrast-distorted SEM images,and it is superior to the existing mainstream full-reference and no-reference quality evaluation methods.