Refining Microstructural Analysis in Metallurgy: Semantic Segmentation Enhanced by Post-Processing

Abstract

This study investigates post-processing strategies to improve semantic segmentation models for microstructural characterization in metallurgy. Microstructural analysis is critical for understanding material properties and relies largely on expert knowledge. Deep learning algorithms have the potential for automation, but they encounter limitations such as small datasets and poor image quality. A watershed segmentation method is proposed after semantic segmentation, with a focus on the hypoeutectic Al-Si alloy. Post-processing methods were evaluated by automatically measuring the number of contours and average object area (features that could be fundamental in many microstructural images) within the study. After evaluating Euclidean, City Block, and Chebyshev distance transforms, the results demonstrate that while semantic segmentation provides high pixel accuracy, watershed segmentation enhances precision, particularly with City Block distance. Post-processed models have closer alignment with ground truth contours, indicating better segmentation. This emphasizes the importance of post-processing in improving semantic segmentation for metallurgical microstructural analysis, which allows for faster and more exact material assessments.