전역 상 분율 정보를 활용한 금속 미세조직 분할

다국어 초록 (Multilingual Abstract)

Automated segmentation of metallographic images is essential for quantifying material phases that determine mechanical properties. These images, however, pose challenges due to overlapping boundaries, similar textures, and complex microstructural characteristics requiring precise phase discrimination. This study proposes a Phase Learning Module (PLM) that improves segmentation accuracy by incorporating global phase-ratio information, a form of expert domain knowledge. The PLM consists of three components: a Phase Ratio Encoder that converts ratio data into neural-compatible representations, Feature Adaptation that aligns this information with spatial context, and Feature Fusion that integrates original image features with ratio-enhanced features. Experiments with U-Net, U-Net++, nnU-Net, and UNet3+ show consistent performance gains, with UNet3+ achieving a Dice score of 90.21% on the MetalDAM dataset.