Grain boundary detection and second phase segmentation based on multi-task learning and generative adversarial network

Abstract

Abstract The size, shape and distribution of microstructures (second phase particles, grains) play an important role in the mechanical properties of alloy products. So, it is important to detect grains and second phase particles precisely. In this paper, we use multi-task learning and generative adversarial network (GAN) to realize the segmentation of the second phase and the boundary detection of grains at the same time. Specifically, a richer convolutional features (RCF) architecture based on multi-task learning is designed for preliminary detection and segmentation. Then, a generative adversarial network is employed to fine tune the hidden grain boundaries that covered by the second phase. Finally, a quantitative analysis module is designed to extract quantitative indicators according to the results of the two deep networks. We achieve 96.65% (accuracy), 0.8325 (IoU), 0.7824 (AJI) in the segmentation task and 92.65% (precision), 91.90% (recall) in the boundary detection task, which reach the state-of-the-art meanwhile.