Improved small foreign object debris detection network based on YOLOv5

Abstract

In response to the challenges of detecting foreign object debris (FOD) on airport runways, where the objects are small in size and have indistinct features leading to false detections and missed detections, significant improvements were made to the YOLOv5 algorithm. First, the original YOLOv5-n model was optimized by incorporating multi-scale fusion and detection enhancements. To improve detection speed and reduce parameters, the detection head for large objects was removed. Second, the C3 module in the backbone network was replaced with the C2f module, resulting in enhanced gradient flow and improved feature representation capabilities. Additionally, the spatial pyramid pooling-fast (SPPF) module in the backbone network was refined to expand the receptive field and enhance the model’s perception of dependencies between targets and backgrounds. Furthermore, the coordinate attention (CA) mechanism was introduced in the neck layer to further enhance the model's perception of small FOD items. Lastly, the SCYLLA-IoU (SIoU) loss function was introduced to further improve the speed and accuracy of bounding box regression. Moreover, the nearest neighbor interpolation upsampling method was substituted with the lightweight Content-Aware ReAssembly of FEatures (CARAFE) upsampling operator to better exploit global information. Experimental results on the Fod_Tiny dataset, which consists of small FOD items in airports, demonstrated a significant 5.4% improvement over the baseline algorithm. To validate the generalizability of the algorithm, experiments were conducted on the Mirco_COCO dataset, resulting in a notable 1.9% improvement compared to the baseline algorithm.