基于改进型Cascade Mask R-CNN的架空输电线路多目标识别算法的研究  

Research on Multi-Target Recognition Algorithm of Overhead Transmission Line Based on Improved Cascade Mask R-CNN

作　　者：张彦宇 尚嘉诚 丁祎琦 鲁亮 王强 丁亚雄 ZHANG Yanyu;SHANG Jiacheng;DING Yiqi;LU Liang;WANG Qiang;DING Yaxiong(State Grid Ningdong Power Supply Company,Yinchuan 750001,Ningxia,China)

机构地区：[1]国网宁东供电公司,宁夏银川750001

出　　处：《电力大数据》2022年第11期9-19,共11页Power Systems and Big Data

摘　　要：近几年来,国网架空输电线路巡检人员依托于巡检无人机、巡检机器人等智能巡检设备,获取了海量的图像和视频信息,巡检图片判断的工作量越来越大,同时因个人技能、经验等方面的差异,可能导致基于巡检图片对于设备缺陷和隐患分析不满足要求。为了对获取的信息进行快速而准确的分析和处理,本文借助公司在电力线路巡检方面积累的丰富图像,将Cascade Mask R-CNN智能识别算法在原有基础上进行改进,使用了ResNet101+FPN的方式来提升算法对于小目标的识别能力,并且将IoU Loss替换为DIoU Loss作为边界框的损失,使网络朝着预测框与真实框重叠度较高的方向去优化,减小正负样本不平衡带来的误差,增强了对远景目标的检测;同时搭建了深度学习TensorFlow实验平台,并在该平台基础上设计并开发了巡检图像目标检测实验平台交互界面,实现了对巡检图像目标检测的可视化操作。In recent years, using unmanned aerial vehicle and the inspection robot, inspectors have acqired vast amounts of image and video information. The workload in checking picture is more and more big, at the same time, the analysis of the defects and hidden danger the equipment based on inspection pictures cannot be fufilled because of the difference of personal skills, experience, etc. In order to quickly and accurately analyze and process the acquired information, this paper improves the Cascade Mask R-CNN intelligent recognition algorithm based on the rich image database accumulated in power line inspection, and uses ResNet101+FPN to improve the algorithm’s recognition ability for small targets. In addition, IoU loss is replaced by DIoU loss as the loss of the boundary frame, so that the network is optimized in the direction of high overlap between the predicted frame and the real frame, reducing the error caused by the imbalance of positive and negative samples and enhancing the detection of the long-range target. At the same time, the deep learning tensorflow experimental platform is built, and based on it, the interactive interface of inspection image target detection experimental platform is designed and developed to realize the visual operation of inspection image target detection.

关 键 词：智能巡检 Cascade Mask R-CNN 残差网络 小目标 交互界面 

分 类 号：TM752[电气工程—电力系统及自动化]