Tiny Rust Detection Algorithm Fusing Multi-scale High-Resolution Features and Attention Mechanism

Early corrosion defects on metal surfaces are characterized by their extremely small size, blurred textures, and strong background interference (e.g., oil stains and light spots). Traditional object detection algorithms often lose critical features after multiple downsampling operations, leading to high missed detection rates. To address this challenge, this paper proposes an improved YOLOv11-based algorithm for tiny rust detection. Based on the YOLOv11s architecture, we construct a P2 micro-object detection layer (4 × downsampling) to preserve high-resolution texture details. Simultaneously, the Convolutional Block Attention Module (CBAM) is integrated to suppress background noise introduced by the P2 layer, ensuring effective feature selection. Ablation studies demonstrate that introducing the P2 layer alone leads to significant false positives and a drop in recall (to 53.6%), whereas the synergistic combination of P2 and CBAM achieves a qualitative breakthrough. Experimental results show that the proposed model achieves an mAP@50 of 62.2% and improves the Recall to 63.2% (a 2.3% increase over the baseline), effectively mitigating the risk of missed detections in industrial inspection. Furthermore, comparative experiments reveal that structural feature enhancement is more robust than simply increasing input resolution (e.g., 1024px) or model depth under limited sample conditions.