Weakly-Supervised Hair SEM Microscope Image Segmentation Using A Priori Structure Information

The paper introduces a novel weakly-supervised pipeline for segmenting hair in SEM images, requiring only image-level annotations. It integrates the Radon transform, Sobel operator, and a Boundary Discrimination Module to estimate boundaries, achieving over 30% improvement in mean Hausdorff Distance and over 50% in standard deviation compared to Unet and SAM. Evaluated on 429 images, the method reduces annotation costs while maintaining accuracy. However, refinement modules may introduce artefacts, affecting performance. The study also proposes a new quality estimation metric and makes the SEM hair dataset open-source for further research.