摘要
航拍图像识别广泛应用于各类军用、民用领域,因其距离远、成像模糊、目标相互遮挡等特点使得目标检测准确度一直不高,针对这一问题,提出了一种基于YOLOv5模型的改进方法。通过引入数据增强和标签平滑方法、改进损失函数为DIoU和增加针对小目标的网络处理层来提高目标检测效果。实验结果表明,在相同训练条件下,改进后的YOLOv5算法对大多数种类的目标检测准确率都有所提升,平均精确率提高了17%,平均召回率提高了2%,mAP@0.5达到了70.4%,比原始模型提升了6.1%。
Aerial image recognition is widely used in various military and civilian fields.Because of its long distance,blurred imaging,and mutual occlusion of targets,the accuracy of target detection has been low.To solve this problem,an improved method based on the YOLOv5 model is proposed..By introducing data enhancement and label smoothing methods,improving the loss function to DIoU,and adding a network processing layer for small targets,the target detection effect is improved.Experimental results show that under the same training conditions,the improved YOLOv5 algorithm has improved the accuracy of most types of target detection.The average accuracy rate has increased by 17%,the average recall rate has increased by 2%,and mAP@0.5 has reached 70.4%,an increase of 6.1%over the original model.
作者
张麒麟
林清平
肖蕾
Zhang Qilin;Lin Qingping;Xiao Lei(Air Force Early Warning Academy,Hubei Wuhan 430000)
出处
《长江信息通信》
2021年第3期73-76,共4页
Changjiang Information & Communications