融合高效注意力的多尺度输电线路部件检测  

Multi-Scale Transmission Line Component Detection Incorporating Efficient Attention

作　　者：陈思雨 付章杰[1,2] CHEN Siyu;FU Zhangjie(Engineering Research Center of Digital Forensics,Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China;The State Key Laboratory of Integrated Services Networks,Xidian University,Xi’an 710126,China)

机构地区：[1]南京信息工程大学数字取证教育部工程研究中心,南京210044 [2]综合业务网理论及关键技术国家重点实验室,西安710126

出　　处：《计算机工程与应用》2024年第1期327-336,共10页Computer Engineering and Applications

基　　金：国家重点研发计划(2021YFB2700900);国家自然科学基金(62172232);江苏省杰出青年基金(BK20200039)。

摘　　要：针对在高分辨率输电线路图像中,不同种类部件尺度跨越大,难以被均衡检测的问题,提出一种融合高效注意力的多尺度输电线路部件检测算法。在YOLO v5目标检测算法中,设计添加高效注意力模块ECBAM提高算法特征提取能力。根据输电线路部件的特征分布统计,使用滑动窗口对高分辨率输电线路图像进行切片,并对切片前后的图像分别使用改进后的YOLO v5算法训练模型。将两个模型的检测结果进行集成,得到多尺度输电线路部件检测结果。在公开的PLAD架空输电线路图像数据集上,该模型的检测性能远超现有目标检测模型,Precision可达83.2%,Recall可达92.8%,相比数据集原作者提出的模型,mAP值提升了1.6个百分点,达到了90.8%,且能检测出未在原始数据集上标注出的隐蔽目标,验证了在高分辨率图像中检测多尺度输电线路部件的有效性。For the problem that different types of components span large scales and are difficult to be detected in a balanced manner in high-resolution transmission line images,a multi-scale transmission line component detection algo-rithm incorporating efficient attention is proposed.Firstly,an efficient attention module ECBAM is designed and added to the YOLO v5 object detection algorithm to improve the feature extraction capability of the algorithm.Secondly,according to the feature distribution statistics of transmission line components,the high-resolution transmission line images are sliced by sliding window,and the improved YOLO v5 algorithm is used separately to train the models for the images before and after slicing.Finally,the detection results of the two models are integrated to obtain the detection results of multi-scale transmission line components.On the publicly available PLAD overhead transmission line image dataset,the detection performance of the proposed model far exceeds existing object detection models,with Precision up to 83.2%and Recall up to 92.8%.Compared with the model proposed by the original authors of the dataset,the mAP value improves by 1.6 percentage points to 90.8%and it can detect hidden objects that are not labeled on the original dataset,which verifies the effectiveness of detecting multi-scale transmission line components in high-resolution images.

关 键 词：输电线路 多尺度目标检测 滑窗切片 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]