A Semi-Supervised Method for Grain Boundary Segmentation: Teacher–Student Knowledge Distillation and Pseudo-Label Repair

Abstract

Grain boundary segmentation is crucial for the quantitative analysis of grain structures and material optimization. However, challenges persist due to noise interference, high labeling costs, and low detection Accuracy. Therefore, we propose a semi-supervised method called Semi-SRUnet, which is based on teacher–student knowledge distillation and pseudo-label repair to achieve grain boundary detection for a small number of labels. Specifically, the method introduces SCConv (Spatial and Channel Reconstruction Convolution) and boundary regression to improve the U-Net (a convolutional neural network architecture) as a teacher network. These innovations aim to reduce spatial and channel redundancy, expand the receptive field, and effectively capture contextual information from images, thereby improving feature extraction robustness and boundary precision in noisy environments. Additionally, we designed a pseudo-label repair algorithm to enhance the Accuracy of pseudo-labels generated by the teacher network and used knowledge distillation to train a lightweight student network. The experimental results demonstrate that Semi-SRUnet achieves 88.86% mean Intersection over Union (mIoU), 96.64% mean Recall (mRecall), 91.5% mean Precision (mPrecision), and 98.77% Accuracy, surpassing state-of-the-art models and offering a novel approach for reliable grain boundary segmentation and analysis.