Material structure segmentation method based on graph attention

Abstract

With the development and integration of multiple disciplines, the integration of computer vision and materials science has greatly changed the original materials research methods. Existing methods can effectively segment the image of a specific scene, but there is no general method to segment and analyze the image of material accurately. To solve the problems of complex texture, blurred boundary and low contrast in material image, we propose a method that relies on multidimensional feature fusion to train the network more effectively with limited and available annotation samples. The architecture consists of an encoder, a graph attention module, a multi-scale feature fusion module and a decoder. We show that such a network can be trained end-to-end from the image. In electron microscope image, the segmentation results are superior to many previous advanced methods. Using this method, we can accurately identify multiple structures in material images, which provides important insights for multiphase segmentation of material images and searching for new mechanisms of structural transformation in material science.