Confidence-Aware Mean Teacher for semi-supervised metallographic image semantic segmentation

Abstract

Semantic segmentation of metallographic images using deep learning has become an emerging topic in materials science. However, existing methods generally rely on plenty of labeled images for supervised learning but pay insufficient attention to the problems of class imbalance and limited samples in datasets, which usually result in degraded model performance. Therefore, to accurately segment microstructures in metallographic images, we propose a new semi-supervised semantic segmentation model called Confidence-Aware Mean Teacher Network (CA-MT). By generating pseudo-labels, unlabeled images can also participate in training. According to the learning state of the model at different stages, the training process is divided into three phases: exploration, growth, and stabilization, and different strategies are applied to generate adaptive pseudo-labels, respectively. This can improve the quality of the training set and enable the model to learn more adequate semantic information for each class. Meanwhile, data augmentation is an effective way to improve the generalization ability of the model in a sample-limited situation. Based on this, we propose an adaptive threshold adjustment strategy (CCAT) depending on the class confidence to deal with the class imbalance problem and an adaptive CutMix augmentation method (CDAC) that utilizes the global confidence of unlabeled images as a guide to mitigate the disturbance caused by insufficient samples. Experiments show that CA-MT outperforms five state-of-the-art semi-supervised segmentation models under different partitioning protocols for two metallographic image datasets, MetalDAM and UHCS. The source code of the present work is available in https://github.com/Cccoral/CA-MT.