Defect segmentation for aerospace alloy components (e.g., aero-engine blades) is critical for safety but suffers from data scarcity due to both the rarity of defects and the wide variations in image appearance caused by complex geometries. Various data augmentation methods have been developed, yet most remain limited to the distribution of existing data.
Here, we propose a physically based image generation method that creates arbitrary scratches governed solely by physical laws, aiming to improve defect segmentation accuracy with limited real images. We reveal that variations in image appearance due to complex geometry have a greater influence on segmentation performance than those due to texture. Because physically synthesized data provide broad coverage of geometric variance, they outperform data-driven methods, which are better at enriching textural variance. The low textural quality of physically synthesized data can be readily compensated with only a few additional real images—adding just 10 or 20 real images already yields significant improvement.
This work presents a systematic approach to physically synthesized data generation and its application for automatic visual inspection of aerospace alloys with complex geometries, and experimentally explains the underlying reasons for its effectiveness from multiple perspectives.
The work titled “Physically synthesized data for deep learning-based visual scratch inspection of aerospace alloys with complex geometries”, was published on Journal of Materials Informatics (published on Mar. 5, 2026).
DOI：10.20517/jmi.2025.74