Curvature-modeling intelligent detection of ablation morphology on CMC sharp leading edges

Ceramic Matrix Composites (CMCs) are essential in hypersonic vehicle thermal protection systems due to their lightweight and superior ablation resistance. However, sharp leading edges are subject to extreme aerodynamic heating and chemical ablation, resulting in irregular, multi-scale, and curvature-rich defects that pose severe safety risks. To address this challenge, we present SLEAM (Sharp Leading-Edge Ablation Morphology Dataset), the first dataset specifically designed for detecting ablation defects in CMC leading edges. SLEAM contains 1226 real-world ablation images under diverse service conditions, angles, and morphologies, annotated with both bounding boxes and polygons to establish a task-oriented benchmark. Based on this dataset, we propose YOLOv11-CMC, an enhanced detection framework incorporating gradient-operator-based curvature modeling. Specifically, (1) Curvature-Guided Attention (CGA) employs learnable gradient operators during feature extraction to highlight edge- and curvature-sensitive regions, while (2) Curvature-Aware Feature Fusion (CAFF) dynamically balances semantic context and geometric details in multi-scale feature aggregation. Experiments on SLEAM demonstrate that YOLOv11-CMC achieves a 2.1% improvement in mAP compared to baseline YOLOv11, particularly excelling in detecting small-scale and geometrically complex defects.