Multi-task learning for joint classification and segmentation of metallographic microstructures

Abstract

Accurate analysis of metallographic microstructures is crucial for evaluating material properties. We propose a multi-task learning framework based on a modified U-Net that jointly performs semantic segmentation and image-level classification. The model features a shared encoder and task-specific decoders, with an adaptive loss weighting mechanism for balanced optimization. Experiments on a manually annotated dataset of 200$\times$ and 500$\times$ optical images of pure iron and hypereutectoid steel show that our approach outperforms traditional and single-task baselines in terms of accuracy, boundary precision, and convergence efficiency in joint-task training scenarios. All code and data used in this study are publicly available at https://doi.org/10.5281/zenodo.15702261.