基于改进型TridentNet输电设备本体图像缺陷检测方法  

Image Defect Detection Method of Power Transmission Equipment Based on Improved TridentNet

作　　者：徐凡 廖逍 卢大玮 白景坡 XU Fan;LIAO Xiao;LU Dawei;BAI Jingpo(State Grid Information&Telecommunication Group Co.,Ltd.,Beijing 102211,China)

机构地区：[1]国网信息通信产业集团有限公司,北京102211

出　　处：《电力信息与通信技术》2022年第12期39-46,共8页Electric Power Information and Communication Technology

基　　金：国网信息通信产业集团有限公司科技项目资助“输电线路智能视频巡检设备及系统研发”(5268002XXX2W)。

摘　　要：输电线路本体图像缺陷检测能够及时有效地发现设备缺陷,预防设备故障的发生。目前的输电图像缺陷检测方法对不同尺度的缺陷检测能力有待提升,需做进一步算法研究。基于输电线路巡检场景,文章提出一种针对TridentNet网络的改进方法。优化基于尺度的训练策略,通过调整筛选范围提升模型训练效果,引入骨干网络中的新维度——势,以提升特征提取效率;设计权重归一化处理环节加入网络的特征提取部分,实现训练数据batch较小时的模型收敛加速与精度提高,从而提升同一图像中多尺度缺陷检测的准确率与召回率。实验结果表明了TridentNet网络的改进方法的有效性与准确性。Transmission line image defect detection can find equipment defects timely and effectively,and some measures can be taken to prevent equipment failures.The existing transmission image defect detection methods are not enough to deal with different scale defect detection tasks,and further algorithm research is needed.Based on the power transmission scene,this paper proposes an improved method for TridentNet network.Firstly,scale-aware training scheme is optimized.The effect of model training is improved by adjusting the screening range,new dimensions cardinality in the backbone network is introduced to improve feature extraction efficiency.Secondly,the weight normalization process scheme is redesigned for feature extraction part of the network to achieve model convergence acceleration and accuracy improvement when the training data batch is small.Thereby the accuracy and recall rate of multi-scale defect recognition in the same image is improvd.Finally,an experiment is carried and results show that the proposed method is effective and accurate.

关 键 词：神经网络 缺陷检测 输电设备 TridentNet 权重归一化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]