SAM2–Watershed 하이브리드 입자 형태 분할 및 정량 평가

초록

Particle size and morphology analysis using scanning electron microscopy (SEM) images is essential for quantitatively evaluating material microstructures. However, accurate particle segmentation in SEM images is challenging due to weak boundaries, local intensity variations, surface texture, noise, and particle contact. In this study, we propose a SAM2–Watershed hybrid method based on residual correction for particle instance segmentation in microscopy images. First, initial particle masks were generated using SAM2 within the Segmenteverygrain framework, which enables automatic segmentation based on generated point prompts. Although SAM2 provides precise object boundaries, it may miss small or low-contrast particles. To recover these missed regions, Watershed segmentation was independently applied to the same SEM image, and regions detected by Watershed but not by SAM2 were defined as residual candidate regions. These regions were re-segmented using the intensity and boundary information of the original image, and candidate objects were filtered by area, overlap, and morphological features. The selected residual particles were then added to the SAM2 result to produce the final segmentation. The proposed method was evaluated against manually generated ground truth at both pixel and instance levels. The results demonstrate that the proposed SAM2– Watershed hybrid approach effectively improves particle detection and morphological quantification in microscopy image analysis.

키워드