BEKDNet: A boundary-enhanced Mamba network with cross-image relational distillation for aerospace thermal control coating defect segmentation

Thermal control coatings, applied as continuous functional coatings onto spacecraft exterior surfaces, are fabricated via roll-to-roll deposition, yet automated in-line detection of surface defects formed during this manufacturing process poses significant challenges due to specular reflections on metallic oxide surfaces, boundary ambiguity caused by overlapping defects, and the lack of domain-specific benchmark datasets. This paper proposes BEKDNet, a boundary-enhanced knowledge distillation segmentation network for thermal control coating defect detection. The framework employs a state space model-based encoder integrated with a defect-aware module, which synergistically combines global context modeling and multi-scale local convolutions to effectively capture multi-scale defects ranging from sub-millimeter voids to centimeter-level scratches while maintaining linear computational complexity. To address the challenge of low-contrast boundary discrimination, an edge guidance module incorporating Laplacian edge extraction and a boundary refinement head with auxiliary boundary supervision are designed. Furthermore, a cross-image relational knowledge distillation framework with DINOv2 as the teacher network is introduced to transfer illumination-invariant semantic representations under limited annotation conditions. Additionally, TCC-Defect, the first benchmark dataset for aerospace thermal control coating defect segmentation comprising three representative defect categories, is established. Experimental results demonstrate that BEKDNet achieves mIoU scores of 88.17%, 85.34%, and 83.56% on TCC-Defect, NEU-Seg, and KSDD2 datasets, respectively, surpassing state-of-the-art methods by 2.12%, 3.04%, and 0.47%. The proposed method provides a reliable solution for automated quality inspection of aerospace thermal control coatings and has been deployed in a digital twin real-time monitoring system at 60 FPS, demonstrating industrial applicability.