精准识别和定位飞机目标是航空安全和信息化战争胜利的关键，针对传统飞机目标识别抗干扰性差，对遮挡、光照、尺度敏感难应对复杂场景需求的问题，提出了 一种基于改进YOLOv5的飞机目标检测算法。通过IOU-NWD Similarity Metric for Bounding Boxes策略解决了 IOU机制对飞机小目标的标签分配歧义问题;使用GFPN based on NLnet模块完成了"跨层"与"跨尺度"的自适应融合，更加丰富和具有代表性的特征信息;使用soft-NMS方法解决了在目标密集区域中飞机小目标存在的漏检问题。在飞机数据集上进行实验，结果表明:与原始 YOLOv5 相比，改进后的模型在 Precision、Recall、mAP0。5、mAP0。5:0。95 分别提高了 1。9％、10。4％、3。6％和5。8％。该算法通过针对性的网络调整和模块迁移来提高模型对小型和遮挡的飞机目标的检测效果，并通过实验验证了该算法的优越性，实验结果表明，AIR-YOLO在检测精度和鲁棒性方面优于YOLOv5，解决了原始YOLOv5算法的小飞机目标误检的问题。
YOLO-Plane:An aircraft detection algorithm based on improved YOLOv5
Accurate identification and positioning of aircraft targets is the key to the victory of aviation safety and information war.In view of the problems that traditional aircraft target identification has poor anti-interference perform-ance and is sensitive to occlusion,illumination and scale and difficult to cope with complex scene requirements,an aircraft target detection algorithm based on improved YOLOv5 is proposed.IOU-NWD Similarity Metric for Bounding Boxes solves the ambiguity of label assignment for aircraft small targets by IOU mechanism.Using GFPN based on NL-net module,the"cross-layer"and"cross-scale"adaptive fusion is completed,and more abundant and representative characteristic information is obtained.soft-NMS method is used to solve the problem of missing detection of small air-craft targets in crowded target areas.The experimental results show that compared with the original YOLOv5,the Pre-cision,Recall,mAP0.5 and MAP0.5:0.95 of the improved model are increased by 1.9％,10.4％,3.6％and 5.8％,respectively.Through targeted network adjustment and module migration,the algorithm can improve the detec-tion effect of the model on small and blocked aircraft targets.The superiority of the algorithm is verified by experi-ments.The experimental results show that AIR-YOLO is superior to YOLOv5 in terms of detection accuracy and ro-bustness,which solves the problem of false detection of small aircraft targets in the original YOLOv5 algorithm.
NETL
NSTL
万方数据
