Enhancing Tool Wear Segmentation with LoRA-SAM and Point Prompts

Automated optical measurement of cutting tools offers a rapid and direct approach to monitoring tool wear. While the Segment Anything Model (SAM) exhibits strong zero-shot generalization capabilities in typical scenarios due to its powerful image encoder and prompt engineering, it underperforms in tool wear segmentation because of its lack of domain-specific knowledge. In this study, an image dataset containing various tools with different prompts was initially established, and SAM was fine-tuned using Low-Rank Adaptation (LoRA) and Mixture-of-Expert (MoE) LoRA to incorporate domain-specific knowledge of tool wear. The results consistently show that both LoRA-SAM and MoE-LoRA-SAM significantly outperform the original SAM in tool wear segmentation with over 55% improvement. Additionally, using a combination of three prompt points is sufficient for accurate tool wear segmentation. This finding highlights the potential applicability of SAM in industrial scenarios.