Diffusion Based Data Synthesis for Enhanced Tool Wear Segmentation

Accurate tool wear segmentation is essential for automated machining. However, deep learning models often require large labeled datasets that are rarely available in industrial environments. Limited image quality, diverse tool geometries, and costly annotation constrain data volume and weaken model robustness. This study proposes a diffusion based data synthesis framework to enhance wear segmentation under small sample conditions. A DreamBooth fine-tuned Stable Diffusion Inpainting model generates realistic wear patterns while preserving tool geometry, guided by structured prompts and a patch conditioned masking strategy. The synthetic data are used to retrain a U-Net segmentation network, achieving about 6% improvement in Dice coefficient on unseen tool types. Results demonstrate that combining real and diffusion-generated data effectively mitigates data scarcity and improves generalization, offering a practical approach for reliable tool wear monitoring in industrial applications.