巡检机器人电力设备文字识别技术研究  

Research on Text Recognition Technology for Inspection Robot Power Equipment

作　　者：刘钊 蔡笋 王增彬 尹泓澈 姚剑[2] 杨英仪 李文胜 梅鹏 LIU Zhao;CAI Sun;WANG Zengbin;YIN Hongche;YAO Jian;YANG Yingyi;LI Wensheng;MEI Peng(Electric Power Research Institute of Guangdong Power Grid Co.,Ltd.,Guangzhou,Guangdong 510080,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan,Hubei 430079,China;China Southern Power Grid Technology Co.,Ltd.,Guangzhou,Guangdong 510180,China)

机构地区：[1]广东电网有限责任公司电力科学研究院,广东广州510080 [2]武汉大学遥感信息工程学院,湖北武汉430079 [3]南方电网电力科技股份有限公司,广东广州510180

出　　处：《自动化应用》2025年第4期1-5,共5页Automation Application

基　　金：南方电网公司科技项目“南方电网公司电力机器人联合实验室开放基金”(GDDKY2022KF06)。

摘　　要：变电站仪表文字提取与识别对于感知电力设备运行状态至关重要。人工巡检存在视觉疲劳的问题,并且在恶劣环境下存在安全隐患,机器人是替代人工巡检的理想方案。然而,定点停车巡检效率较低,为了提高机器人巡检效率,可采用不停车巡检策略。针对不停巡检时,由于背景无关物体干扰以及非正射观测视角造成的文字识别准确率下降的问题,提出了一种基于目标影像正射纠正的文字识别方法。首先通过目标检测算法排除无关物体干扰,然后通过估计目标6DoF位姿对场景影像中的目标区域进行正射纠正,从而提高OCR算法的文字识别准确率。实验结果表明,经过正射纠正的目标影像,文字识别准确率提升了6%以上,有效地保障了机器人不停车巡检时文字识别的准确性。Substation instrumentation text extraction and recognition is crucial for sensing the operating status of power equipment.Manual inspection has problems such as visual fatigue in harsh environments.Robot is an ideal solution to replace manual inspection.However,the efficiency of fixed-point parking inspection is low.In order to improve the robot inspection efficiency,the non-stop inspection strategy can be adopted.A text recognition method based on ortho-correction of the target image is proposed by studying the problem of decreasing text recognition accuracy due to the interference of irrelevant objects in the background and the non-projective observation perspective during the non-stop inspection.Firstly,the method excludes irrelevant object interference by the target detection algorithm,and then orthorectifies the target region in the scene image by estimating the target's 6DoF position,so as to improve the text recognition accuracy of the OCR algorithm.The experimental results show that the text recognition accuracy of the ortho-corrected target image is improved by more than 6%,which effectively guarantees the accuracy of text recognition when the robot inspects without stopping.

关 键 词：电力作业 深度学习 目标检测 位姿估计 正射纠正 文字提取 

分 类 号：TH865[机械工程—仪器科学与技术]