Automatic Image Recognition of Austenite Grain Boundary in Martensitic Metallography based on Image Style Transfer

Abstract

For most steel materials, the conventional corrosion method can only observe the martensite structure after transformation. There are some problems in measuring austenite grain size, such as complex operation, difficult to ensure the corrosion quality and so on. Therefore, we use machine learning to identify the original austenite grain boundary according to the martensite structure of conventional corrosion. In this paper, image style transfer is realized by iterative method based on generating model, and austenite grain boundary recognition during martensitic transformation is realized by means of pre training network model vgg19. Firstly, the pre trained deep network vgg19 is used to extract the style and content features of martensite metallographic images. Then, the loss function of style and content is defined, and the gradient descent method is used to iterate step by step to optimize the total loss. Finally, the austenite image with clear grain boundary is obtained by texture segmentation.