A pseudo-labeling based weakly supervised segmentation method for few-shot texture images

Abstract

Automatic segmentation of key region shapes from material microstructure images is one of the primary steps to count material phase data and mine material properties. With the rapid development of deep learning networks, the image segmentation task has achieved great success. However, deep learning-based material microstructure image segmentation still faces the problems of sample scarcity, annotation difficulty and model generalizability. These problems lead to the poor performance of existing deep learning networks on material microstructure image segmentation. To address these problems, we propose a weakly supervised pseudo-labeling texture semantic segmentation (PTS) network based on scribble annotations. The PTS network uses the auxiliary branch to generate pseudo labels for additional supervision and to optimize the scribble annotation segmentation results. This achieves an improvement in segmentation accuracy and generalization ability. We conduct experiments for the PTS network on small-sample titanium alloy images, ceramic images and carbon steel images. The segmentation results of the PTS network achieves 14.4%∼19.8% improvement in mIoU (mean intersection over Union) and 9.5%∼18% improvement in mDice (mean Dice) compared to the existing deep learning networks. Extensive experiments have validated that the PTS network can well overcome the problems of sample scarcity, annotation difficulty and model generalizability in material microstructure image segmentation.