Hypergraph-Enhanced Lightweight Detection Architecture for Foreign Object Debris Identification on Airport Runways

Abstract

Foreign Object Debris (FOD) on airport runways poses a persistent and costly safety hazard to aviation operations worldwide. In this work, we developed a lightweight, real-time detection framework grounded in hypergraph-based adaptive correlation modeling for accurate FOD identification across 31 debris categories. The architecture integrates three core mechanisms. First, a hypergraph-based correlation enhancement module constructs learnable hyperedges from multi-scale backbone features to capture high-order inter-pixel relationships through a two-stage message passing scheme with linear computational complexity. Second, a full-pipeline aggregation-and-distribution strategy routes correlation-enriched features to backbone–neck junctions, internal neck layers, and neck–head interfaces through learnable gated tunnels. Third, depthwise separable convolution blocks replace standard convolutions throughout the network to maintain large receptive fields at a fraction of the parameter cost. The framework was fine-tuned from general-purpose object detection weights on the FOD-A benchmark (33,793 images, 31 classes) and trained for 100 epochs at 640×640 resolution. The resulting model, containing roughly 2.5 million parameters and 6.4 GFLOPs, achieved 99.37% mAP at IoU 0.50 and 93.91% mAP at IoU 0.50:0.95 on the validation set, with precision and recall exceeding 99.6% and 99.5%, respectively. These results surpass previously reported FOD detection benchmarks while operating within the computational budget required for realtime runway monitoring. The source code is publicly available at https://github.com/encoder43/Hypergraph-Enhanced-Light weight-Detection-for-Runway-Foreign-Object-Debris, and the complete training results and model weights can be accessed at https://drive.google.com/drive/folders/1-Ayv3sgXCgPJub2K9m 2-4QygXVdSWOOW?usp=drive link.