基于改进YOLOv3的接触网设备目标检测方法

Object Detection Method of Overhead Lines Equipment Based on Improved YOLOv3

为了改善铁路接触网设备检测的漏检率和误检率高的问题,提出一种改进YOLOv3的接触网设备目标检测方法。在Darknet-53与预测特征层之间加入一块空间金字塔模块,增大感受野;引入图像增强方法,增加了数据集的鲁棒性和多样性;用GIoU损失函数代替原来的IoU损失函数,提升目标检测精度。实验结果表明,将改进模型应用在铁路接触网设备中,平均检测精度(mAP)达到89.19%,相比原来的YOLOv3模型提高了9.38%,且目标检测速度变化不大,说明该目标检测模型可满足接触网设备的检测要求。

In order to improve the high rate of missed detection and error detection of catenary equipment,propose an improved target detec⁃tion method of YOLOv3 catenary equipment.A spatial pyramid module is added between Darknet-53 and the predicted feature layer to in⁃crease the receptive field.The method of image enhancement is introduced to increase the robustness and diversity of data set.GIoU loss func⁃tion is used to replace the original IoU loss function to improve the target detection accuracy.Experimental results show that when the improved model is applied to catenary equipment,its average detection accuracy(mAP)reaches 89.19%,which is 9.38%higher than the original YO⁃LOv3 model,and the target detection speed has little change,indicating that the proposed target detection model meets the detection effect of catenary equipment.