燕山大学学报2025，Vol.49Issue(6)：515-524,10.DOI:10.3969/j.issn.1007-791X.2025.06.006
基于改进YOLOv7和ConvNeXt的吊弦状态判识方法
A method for detecting status of dropper based on improved YOLOv7 and ConvNeXt
摘要
Abstract
The complex background and slender shape of the dropper in railway overhead contact network images captured by UAVs often result in missed and erroneous detection of the dropper status in existing detection methods and algorithms.Therefore,To accurately identify droppers'defects in UAVs images,a method based on improved YOLOv7 and ConvNeXt is proposed for recognizing the state of dropper.Firstly,in the dropper positioning stage,the YOLOv7 method uses a feature separation module to suppress background information and highlight the target.The CBAM attention module is introduced during downsampling to enhance the features of the dropper targets,and a Wasserstein distance-based bounding box loss is integrated to improve target localization accuracy.This accurately locates the dropper areas in high-resolution images.The dropper state classification stage is responsible for the fine-grained identification of the dropper states.Images of the dropper areas are input at high resolution into the ConvNeXt-T network for fine-grained classification prediction to identify the dropper states.Experiments on the railway catenary dropper dataset show that the proposed method achieves an average precision accuracy of 97.74％,an improvement of 4.37％compared to the baseline method.The defect classification accuracy for dropper can reach 97.25％.The proposed algorithm can accurately detect abnormal states such as slack and breakage of droppers in images captured by UAVs,providing an important reference for UAV-based monitoring of catenary dropper states.关键词
接触网吊弦检测/状态识别/YOLOv7/注意力机制/损失函数Key words
contact network dropper detection/status recognition/YOLOv7/attention mechanism/loss function分类
信息技术与安全科学引用本文复制引用
王文斌,李瑞,宋宗莹,王勇,曾杉..基于改进YOLOv7和ConvNeXt的吊弦状态判识方法[J].燕山大学学报,2025,49(6):515-524,10.基金项目
中国神华能源股份有限公司科技项目(SHGF-21-01) （SHGF-21-01）
河北省科技计划项目(216Z1704G) （216Z1704G）
中国科学院战略性先导科技专项(XDB0740200) （XDB0740200）
