复杂场景下灯杆和电力立杆病害细粒度检测
Fine-grained detection of damages on streetlights and power poles in complex scenarios
戴激光 1王劲翔 1吴玉洁 1王妍雯 1黄泽超 2胡彦玲2
作者信息
- 1. 辽宁工程技术大学测绘与地理科学学院,辽宁阜新 123000
- 2. 辽宁省交通建设管理有限责任公司,沈阳 110005
- 折叠
摘要
针对路灯和电力立杆病害的局部特征微弱、类内差异大、模型参数量大等问题,该文提出了一种面向路灯和电力立杆病害细粒度检测的Pole-YOLO算法.根据立杆表现为长条形的管状结构,引入动态蛇形卷积,解决模型难以关注局部微弱的特征,缓解识别中断的情况;针对病害的类内多样的特征变化,将可变形卷积融入C2f模块,以适应任意尺寸病害特征,增强特征提取能力;并使用WIOUv3损失函数,提高对小 目标病害的学习能力.最后,使用分布移位卷积降低模型参数量.在自制的立杆病害数据集上进行对比实验.结果表明Pole-YOL O方法的精确率、召回率、F1-Score和平均精度mAP@0.5分别达到了 84.2％、79.8％、81.9％和83.7％,可以满足城市场景的立杆病害的实时检测.
Abstract
In response to challenges such as faint local features,significant intra-class variations,and a high model parameter count specific to pole damages,a Pole-YOLO algorithm designed for fine-grained detection of defects on streetlight and utility poles was proposed in this paper.Given that poles exhibit elongated cylindrical structures,the introduction of dynamic snake convolution addressed the challenge of model inattention towards faint local features,thereby alleviating recognition interruptions.For the diverse intra-class variations of damages,deformable convolutions were integrated into the C2f module to adapt to arbitrary-sized damage features,enhancing the capability of feature extraction.Furthermore,the WIOUv3 loss function was employed to enhance the learning capability for small target damages.Finally,the employment of distribution shift convolutions reduced the model parameter count.Comparative experiments conducted on a custom-made datasct for pole damages demonstrate that the Pole-YOLO method achieved precision,recall,F1-Score,and mean average precision mAP@0.5 of 84.2％,79.8％,81.9％,and 83.7％,respectively.This indicated its suitability for real-time detection of pole damages in urban scenarios.
关键词
立杆病害检测/目标检测/细粒度分类/街景影像Key words
upright pole disease detection/object detection/fine-grained classification/street scene imagery引用本文复制引用
出版年
2024
NETL
NSTL
万方数据