摘要
Abstract
The detection of small-scale targets,characterized by limited available features and unclear tex-tures,has perennially posed a challenge in the field of object detection.To address issues related to false positives and false negatives unmanned aerial vehicle(UAV)targets,we propose an improved UAV small-target detection algorithm,termed LASD-YOLOv5.This algorithm introduces a polarized self-attention mechanism to more accurately extract minute features,incorporates a weighted bidirectional fea-ture pyramid network to replace the path aggregation network,thus enhancing the utilization of low-level features.Furthermore,it decouples the detection head to expedite model convergence.Additionally,to tackle the scarcity of small targets and incomplete scene coverage in existing UAV small-target datasets,we contribute a multi-scene,low-speed,small UAV target dataset(LASD-D).The experimental results demonstrate that our proposed algorithm achieves an average precision of 98.29%in LASD-D,surpass-ing the baseline network by 2.87%.Notably,it outperforms mainstream algorithms such as YOLOv7、YOLOv8 and QueryDet,effectively meeting the demands of UAV small-target detection applications.关键词
小目标检测/无人机/注意力机制/特征融合/YOLOv5Key words
small object dtection/UAV/attention mechanism/feature fusion/YOLOv5分类
信息技术与安全科学
