Intelligent Corrosion Sensing and Detection for Aerospace Ground Equipment: A YOLOv11-Based Framework with Shallow Attention and Bidirectional Feature Fusion

Abstract

Aerospace ground facilities operating in harsh "three-high" marine environments face rapid corrosion that threatens structural safety. Automated visual inspection is fundamentally hindered by two intertwined challenges: extreme scarcity of annotated real samples, and the intrinsic complexity of corrosion targets themselves—micro-pitting and fine cracks occupy only a handful of pixels with signals easily overwhelmed by complex backgrounds, while mature corrosion regions exhibit extreme geometric irregularities that defeat conventional detection and regression methods. To tackle these challenges, this paper proposes CorrSense, a YOLOv11-based framework with three synergistic innovations. First, we establish a dual-path data ecosystem combining real acquisition with diffusion-driven augmentation that synthesizes visually realistic and structurally consistent corrosion samples to expand training distribution. Second, we develop a saliency-based feature enhancement strategy using parameter-free SimAM attention in shallow layers, which amplifies micro-corrosion responses while suppressing background activations with zero parameter overhead. Third, we formulate a dual-pronged geometric calibration strategy: a customized BiFPN with learnable weighted fusion for dynamic cross-scale feature orchestration, coupled with a morphology-adaptive CIoU loss that modulates aspect ratio constraints according to target shape. Extensive experiments demonstrate that CorrSense consistently outperforms state-of-the-art detectors including YOLOv8, YOLOv9, and RT-DETR on challenging samples, particularly on irregularly shaped corrosion and micro-pitting targets where competing methods typically struggle, validating its effectiveness as a promising solution for intelligent corrosion inspection in coastal launch sites.