测绘通报 ›› 2021, Vol. 0 ›› Issue (10): 132-135,167.doi: 10.13474/j.cnki.11-2246.2021.320

• 技术交流 • 上一篇    下一篇

基于改进卷积神经网络的城市地下排水管道视频智能识别

彭述刚, 王大成, 谭军辉, 敖旋峰, 俞佳颖   

  1. 广东绘宇智能勘测科技有限公司, 广东 广州 510665
  • 收稿日期:2020-09-11 出版日期:2021-10-25 发布日期:2021-11-13
  • 通讯作者: 王大成。E-mail:948609547@qq.com
  • 作者简介:彭述刚(1989-),男,硕士,从事遥感与测绘研究。E-mail:1933811175@qq.com
  • 基金资助:
    2019年珠海市促进实体经济高质量发展专项资金

A video intelligent detection method for urban underground drainage pipe based on improved convolutional neural network

PENG Shugang, WANG Dacheng, TAN Junhui, AO Xuanfeng, YU Jiaying   

  1. Guangdong Huiyu Intelligent Survey Technology Co., Ltd., Guangzhou 510665, China
  • Received:2020-09-11 Online:2021-10-25 Published:2021-11-13

摘要: 城市地下排水管道是市政工程中的一个重要基础设施,但随着管道的老化,不同类型的缺陷开始产生,如何及时有效地对其进行智能识别是当前市政管理面临的一个重要挑战。对此,本文利用当前最新的深度学习进行管道视频图像的智能识别,提出了基于改进卷积神经网络的IM-CNN算法。算法基于InceptionV3框架,针对需要识别的管道图片不同类别缺陷间具有类不平衡性的特点,设计了面向管道识别的改进特征融合策略及新的目标函数。基于地下排水管道数据集的试验表明,该算法的预测识别能力不仅优于传统的机器学习算法,而且强于已有直接利用卷积神经网络的算法。

关键词: 地下排水管道, 深度学习, 卷积神经网络, 多分类学习, 图像分类

Abstract: Urban underground drainage pipeline is an important infrastructure in municipal engineering, but with the aging of the pipeline, different types of defects begin to appear, it is an important challenge that how to identify the defects in a timely and effective manner for municipal management. In this regard, this paper proposes an IM-CNN algorithm based on improved convolutional neural network at the use of latest deep learning to recognize the pipeline video images intelligently. This algorithm uses the framework of InceptionV3. At the meantime, we design an improved feature fusion strategy and a new objective function for pipeline recognition for characteristics that defects in different categories of pipeline images has the unbalance of different categories. Experiments based on the underground drainage pipeline data set show that the predictive recognition ability of the proposed algorithm is not only superior to the traditional machine learning algorithm but also superior to the existing algorithms that directly utilize the convolutional neural network.

Key words: underground drainage pipeline, deep learning, CNN, multi-class learning, image classification

中图分类号: