Open-set image segmentation based on an improved Segment Anything model

Image segmentation is crucial in computer vision, but traditional approaches are limited by dataset and label diversity. Segment Anything introduces open-set segmentation capabilities but struggles with clause-level prompts and precise mask edges. This paper presents SAM-TP, a new algorithm combining Grounding DINO for clause-level prompt handling and Segment Anything for segmentation. By integrating non-maximum suppression and connected component denoising, we refine segmentation accuracy. Additionally, incorporating ViT-Matte enhances edge precision through a trimap-based structure. Experiments on the COCO2017 dataset show that SAM-TP improves Mean Pixel Accuracy by 6.04%, achieving 24.74%, demonstrating its enhanced foreground-background segmentation ability.