An Improved Aircraft Landing Runway Feature Detection MethodBased on YOLOv5

To address the challenges associated with traditional visual runway number identification during aircraft landing, particularly under adverse weather conditions at decision heights where manual visual confirmation becomes error-prone and poses safety risks, this study proposes an enhanced YOLOv5s-based detection framework for aircraft landing runway markings. The methodology integrates four key technical enhancements: 1) Convolutional Block Attention Module (CBAM) for adaptive feature refinement, 2) Soft Non-Maximum Suppression (Soft-NMS) to mitigate occlusion-induced false negatives, 3) Alpha-CIOU loss function to optimize bounding box regression gradients, and 4) Bidirectional Feature Pyramid Network (BiFPN) for multi-scale feature fusion. Complementary external dataset augmentation techniques are employed to expand feature representation diversity, effectively reducing both missed detection and false positive rates. During training, the Alpha-CIOU module demonstrates superior gradient optimization efficiency compared to conventional IOU variants. Experimental validation confirms that the proposed YOLOv5s-GLW architecture achieves performance improvements of 2.06% in precision (p%), 1.05% in recall (R%), and 2.92% in mean average precision (mAP%), establishing an effective solution for robust runway marking detection in complex meteorological scenarios.