Research On Image Classification Of Circular Particles Scanning Electron Microscope Based On Particle Size Distribution

In the material science,the particle size distribution of the material particles is very important for the physical or chemical properties of materials,as well as particle type,particle shape and surface texture of the constituent materials.When analyzing the particle size distribution of different materials,the existing direct measurement method based on particle images is affected by the overlapping of particles,the accuracy is not high,and it is easy to cause large errors in the classification.However,the method based on image texture analysis generally adopts traditional machine learning methods to classify images,the feature extraction process and the classification process are independent of each other,the implementation process is more complex,and the accuracy of classification is also not high.In response to the above issues,the main work of this thesis is as follows:(1)For the problem of particle overlap,an improved elliptical overlapping particle segmentation algorithm is proposed.Compared to the Bounded Erosion-Fast Radial Symmetry(BE-FRS)algorithm,the improved algorithm achieves the purpose of automatic parameter selection by circle fitting at the concave points.In addition,through calculating FRS in the local edge after edge segmentation,the improved method has a better segmentation result on overlapping particles with various sizes.In this thesis,the proposed method is compared with the multiple existing overlapped particle segmentation algorithms through experiments,and the results show that the segmentation accuracy of the proposed method is improved by more than 2%.(2)The overlapped particle segmentation algorithm proposed in the previous part of this thesis was used to remove the overlap of particles,the histogram of particle size distribution of each image was calculated and classified using Support Vector Machine(SVM).In order to verify the method,the powder image dataset with eight different particle size distributions was used in this thesis.The test results show that the method can achieve a classification accuracy of 91.41%.Compared with the original algorithm,the classification accuracy is improved by 2.25%.(3)For the problem of poor accuracy in particle size classification based on direct measurement,the Inception-ResNet v2 model is used to achieve particle size classification based on image texture.The traditional machine learning methods have poor accuracy in classifying images with different particle size distributions,feature extraction and classification are divided into separate processes,the implementation is also complex.In this thesis,deep learning method is used to classify images with different particle sizes.The Inception-ResNet v2 model is trained using powder image dataset containing eight different particle size distributions,and the classification effect of the model is verified in the test set.Compared to the classification accuracy of 88.87% in the test set using the Bag of Visual Words(BoVW),the classification accuracy of Inception-ResNet v2 model can achieve 98.24%.The application of particle size distribution classification is studied based on particle size direct measurement method and image texture classification method.An improved overlapping particle segmentation algorithm is proposed,then deep learning methods are used to characterize and classify images with different particle size distribution,and good results are achieved.