摘 要: 针对人工检测大型工程车辆频繁进出施工现场容易出现漏检或误检的问题,文章提出一种改进DETR(基于Transformer的端到端目标检测网络)模型对输电线路工程车辆进行检测识别。首先在原始的DETR主干网络中,引入空洞卷积法获取更多深层次的特征,扩大感受野;其次加入特征金字塔网络(FPN),融合不同尺度的特征,增强特征的健壮性;最后将损失函数GIOU改为CIOU,使模型在训练的过程中达到更快和更好的收敛效果。实验结果显示,改进后的DETR模型在自制数据集中AP50(IOU阈值取0.5)和AP50-95(IOU阈值取0.5~0.95)分别达到了92.1%和61.3%,说明该改进模型在识别输电线路工程车辆场景中具有较高的应用价值。 |
关键词: 空洞卷积;特征金字塔网络;DETR;损失函数 |
中图分类号: TP311
文献标识码: A
|
|
Research on Transmission Line Engineering Vehicle Detection Based on Improved DETR Mode |
ZHANG Linlong1, HU Xuxiao1, HU Kezhen2
|
(1.College of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China; 2.School of Computing University of Sydney, Sydney NSW2006, Australia)
19357388782@163.com; huxuxiao@zju.edu.cn; 13588266234@163.com
|
Abstract: In view of the problem of missing or false detection in manual detection of large engineering vehicles frequently entering and leaving the construction site, this paper proposes an improved DETR (End-to-End Object Detection with Transformers) model to detect and identify transmission line engineering vehicles. Firstly,in the original DETR backbone network, the dilated convolution method is introduced to obtain more deep features and enlarge the receptive field. Then,the Feature Pyramid Network (FPN) is added to fuse the features of different scales to enhance the robustness of the features. Finally, the loss function GIOU is changed to CIOU to make the model converge faster and better in the training process. The experimental results show that AP50 (IOU threshold of 0.5) and AP50-95 (IOU threshold of 0.5~0.95) of the improved DETR model reach 92.1% and 61.3% respectively in the self-made dataset, which indicates that the improved model has high value of application in identification of transmission line engineering vehicles. |
Keywords: dilated convolution; FPN; DETR; loss function |