计算机与现代化

• 模式识别 • 上一篇    下一篇

一种采用级联RPN的多尺度特征#br# 融合电表箱锈斑检测算法

  

  1. (国网浙江省电力有限公司信息通信分公司,浙江杭州310012)
  • 收稿日期:2019-05-27 出版日期:2020-02-13 发布日期:2020-02-13
  • 作者简介:王文(1985-),男,浙江衢州人,工程师,本科,研究方向:信息化项目管理,E-mail: wangwen@zj.sgcc.com.cn; 周晨轶(1993-),男,浙江杭州人,助理工程师,硕士,研究方向:信息化项目管理,E-mail: zhouchenyi@zj.sgcc.com.cn; 徐亦白(1992-),男,山东枣庄人,助理工程师,硕士,研究方向:信息化项目管理,E-mail: xuyibai@zj.sgcc.com.cn; 卢杉(1992-),男,浙江丽水人,助理工程师,硕士,研究方向:信息化项目管理,E-mail: lushan@zj.sgcc.com.cn; 周梦兰(1993-),女,浙江绍兴人,助理工程师,本科,研究方向:信息化项目建设,E-mail: zhoumenglan@zj.sgcc.com.cn。
  • 基金资助:
    国网浙江省电力有限公司科技项目(5211XT17000C)

A Multi-scale Feature Fusion Meter Box Rust Spot Detection Algorithm Using Cascaded RPN

  1. (Information and Communication Branch, State Grid Zhejiang Electric Power Co. Ltd., Hangzhou 310012, China)
  • Received:2019-05-27 Online:2020-02-13 Published:2020-02-13

摘要: 配电柜锈蚀会导致的后果有接触不良,严重的甚至会导致火灾、部分电气控制设备爆炸。为此,本文提出一种基于神经网络的多尺度电表锈斑检测方法。首先,基于大量的锈斑数据,训练识别锈斑的卷积神经网络(CNN)模型;其次,利用训练出的CNN模型,对电表表箱的位置进行检测,同时实现对电表表面锈斑的实时识别。算法融合了通过级联RPN网络获得多尺度的特征映射,充分利用低层特征的位置信息和高层特征的强语义信息来增强检测效果。针对采集的电表锈斑数据集,电表检测达到94.9%的精确度,优于采用YOLOv2达到的91.1%的精确度,锈斑分类精度达到94.5%。锈斑识别的识别率、实时性和稳定性可以较好地满足实际应用的需要。

关键词: 多尺度特征, 级联RPN, 锈斑检测

Abstract: The consequences of corrosion of the power distribution cabinet may result in poor contact, which may even lead to fire and explosion of some electrical control equipment. To this end, this paper proposes a neural network based multi-scale meter rust spot detection method. Firstly, based on a large number of rust spot data, a convolutional neural network (CNN) model for identifying rust spots is trained. Secondly, the trained CNN model is used to detect the position of the meter box, and real-time recognition of the rust spot on the surface of the meter is realized. The algorithm combines the multi-scale feature mapping through the cascaded RPN network, and makes full use of the location information of the low-level features and the strong semantic information of the high-level features to enhance the detection effect. For the collected rust spot dataset of the meter, the meter detection reaches 94.9% precision, which is better than 91.1% precision achieved by YOLOv2, and the rust spot classification precision reaches 94.5%. The recognition rate, robustness, real-time and stability of rust spot recognition can better meet the needs of practical applications.

Key words:  multi-scale features, cascaded RPN, rust spot detection

中图分类号: