CorDet-3H: A Physics-Aware Data Augmentation and Multi-Scale Dynamic Perception Framework for Corrosion Detection in Aerospace “Three-High” Environments

The integrity of ground-based infrastructure at coastal spaceports is critically threatened by accelerated corrosion under synergistic high temperature, humidity, and salt spray conditions. Automated visual inspection is urgently needed to replace hazardous and subjective manual methods, but it faces significant challenges: a severe scarcity of domain-specific labeled data, and the intrinsic difficulty of detecting corrosion defects that exhibit extreme multi-scale variation, blurred boundaries, and low contrast against complex backgrounds. To address these challenges, this paper proposes CorDet-3H, a novel detection framework that synergistically integrates a physics-aware data augmentation pipeline with a multi-scale dynamic perception network. First, we construct the first dedicated Aerospace Corrosion-$\text{3 H} (\text{A C - 3 H})$dataset. Beyond systematic real-world collection, we employ a diffusion model fine-tuned via Low-Rank Adaptation (LoRA) to generate high-fidelity synthetic corrosion samples through latent space manipulation, ensuring physical consistency via Fréchet Inception Distance (FID) filtering. Second, we design an enhanced YOLOv11-based detector featuring two key architectural innovations: (i) parameter-free SimAM attention modules integrated into shallow layers to amplify the salient features of micro-pits against noisy backgrounds, and (ii) a Bidirectional Feature Pyramid Network (BiFPN) with learnable fusion weights to enable adaptive, cross-scale feature interaction for handling both minute and large-scale defects. Third, we adopt the Complete-IoU (CIoU) loss for bounding box regression, which is particularly effective for the precise localization of slender cracks and diffuse corrosion patches due to its explicit penalties for center-point distance and aspect ratio discrepancy. Comprehensive experiments on the AC-3H dataset demonstrate the effectiveness of each component. Our final model achieves a mean Average Precision (mAP @ 0.5) of 92.3 %, with marked improvements in small-target recall and localization accuracy for boundary-ambiguous targets. The proposed framework provides a robust and reliable solution for the intelligent corrosion inspection of critical aerospace infrastructure.