Research on Classification of Cement-based Electron Microscope Images Based on Improved Residual Network

In order to automatically identify and classify the cement electron microscope images accurately and quickly, this paper proposes that the residual network ResNet50 is used as the main frame, and the depth separable convolution and grouping convolution are introduced into the residual structure, which reduces the network parameters and computation, accelerates the model convergence and improves the classification accuracy. In addition, GAM module is embedded in the network to recalibrate the output features and extract more important feature information, which further improves the classification performance of the network. Finally, this method is compared with AlexNet, GoogleNet, VGG-16 and standard ResNet-50. Using the Pytorch training framework, the experimental results show that the accuracy of the test set obtained by ResNet-50 pre-training model is 94. 31%, which is higher than 88.12% of AlexNet, 91.32% of GoogloNet, 91.25% of VGG-16 and 91.55% of standard Resnet-50, effectively improving the recognition accuracy.

I. Introduction

Scanning electron microscope (SEM) is a complex electronic optical instrument, which scans the sample surface, collects, magnifies and images the obtained sample information, describes the microscopic characteristics of the sample surface, and can analyze the composition of micro-areas [1]. In recent years, electron microscope characterization has been widely used in petrochemical, pharmaceutical and forestry fields [2]. Although the surface morphology of the object can be directly observed according to the electron microscope images, it is difficult to identify whether the detected object meets the requirements with naked eyes because the scanning electron microscope images are gray images with little difference in content, and the recognition rate is low and the work repeatability is high. Therefore, intelligent classification of electron microscope images of detected objects has become a hot issue in research. Using artificial intelligence algorithm, the computer is directly used to classify the images formed under the electron microscope, or to judge whether the objects meet the requirements, which can greatly reduce the interference of artificial factors on the obtained results, improve the accuracy of classification and recognition of electron microscope images, and further improve the work efficiency.