计算机应用 ›› 2021, Vol. 41 ›› Issue (8): 2366-2372.DOI: 10.11772/j.issn.1001-9081.2020101603

所属专题: 前沿与综合应用

• 前沿与综合应用 • 上一篇    下一篇

基于改进单点多盒检测器的大坝缺陷目标检测方法

陈静1, 毛莺池1, 陈豪2, 王龙宝1, 王子成3   

  1. 1. 河海大学 计算机与信息学院, 南京 211100;
    2. 华能澜沧江水电股份有限公司, 昆明 650214;
    3. 中国电建集团昆明勘测设计研究院有限公司, 昆明 650051
  • 收稿日期:2020-10-15 修回日期:2021-01-15 出版日期:2021-08-10 发布日期:2021-01-27
  • 通讯作者: 毛莺池
  • 作者简介:陈静(1998-),女,安徽庐江人,硕士研究生,CCF会员,主要研究方向:数据融合、分布式计算;毛莺池(1976-),女,上海人,教授,博士,CCF高级会员,主要研究方向:分布式计算、数据挖掘;陈豪(1982-),男,云南大理人,高级工程师,博士研究生,主要研究方向:大坝安全监测;王龙宝(1977-),男,江苏盐城人,副教授,博士,CCF会员,主要研究方向:大数据、云计算、管理信息化;王子成(1990-),男,湖北监利人,工程师,硕士,主要研究方向:安全监测预警。
  • 基金资助:
    国家重点研发计划项目(2018YFC0407105);华能集团总部科技项目(HNKJ19-H12);国网新源科技项目(SGTYHT/19-JS-217)。

Dam defect object detection method based on improved single shot multibox detector

CHEN Jing1, MAO Yingchi1, CHEN Hao2, WANG Longbao1, WANG Zicheng3   

  1. 1. College of Computer and Information, Hohai University, Nanjing Jiangsu 211100, China;
    2. Huaneng Lancang River Hydropower Incorporated Incorporation, Kunming Yunnan 650214, China;
    3. Power China Kunming Engineering Corporation Limited, Kunming Yunnan 650051, China
  • Received:2020-10-15 Revised:2021-01-15 Online:2021-08-10 Published:2021-01-27
  • Supported by:
    This work is partially supported by the National Key Research and Development Program of China (2018YFC0407105), the Science and Technology Project of China Huaneng Group Headquarter (HNKJ19-H12), the Science and Technology Project of State Grid XINYUAN Company (SGTYHT/19-JS-217).

摘要: 为提升大坝安全运维的效率,大坝缺陷目标检测模型有助于辅助巡检人员进行缺陷检测。大坝缺陷几何形状多变,而采用传统卷积方式进行特征提取的单点多盒检测器(SSD)模型无法适应缺陷的几何变换。针对上述问题,提出可变形卷积单步多框检测器(DFSSD)模型。首先将原始SSD的主干网络VGG16中的标准卷积替换为可变形卷积,用于处理缺陷的几何变换,并且通过学习卷积偏移量来提升模型的空间信息建模能力;其次针对不同特征的尺寸,改进先验框比例,从而提高模型对条形特征的检测精度与模型的泛化能力;最后为解决训练集正负样本不均衡的问题,采用改进的非极大值抑制(NMS)算法来优化学习效果。实验结果表明:DFSSD模型较基准模型SSD在大坝缺陷图像上的平均检测精度提升了5.98%。相较于基于区域的更快卷积神经网络(Faster R-CNN)和SSD模型,DFSSD模型在大坝缺陷目标检测精度提升上有较好的效果。

关键词: 目标检测, 工程缺陷, 可变形卷积, 单点多盒检测器, 非极大值抑制

Abstract: In order to improve the efficiency of dam safety operation and maintenance, the dam defect object detection models can help to assist inspectors in defect detection. There is variability of the geometric shapes of dam defects, and the Single Shot MultiBox Detector (SSD) model using traditional convolution methods for feature extraction cannot adapt to the geometric transformation of defects. Focusing on the above problem, a DeFormable convolution Single Shot multi-box Detector (DFSSD) was proposed. Firstly, in the backbone network of the original SSD:Visual Geometry Group (VGG16), the standard convolution was replaced by the deformable convolution, which was used to deal with the geometric transformation of defects, and the model's spatial information modeling ability was increased by learning the convolution offset. Secondly, according to the sizes of different features, the ratio of the prior bounding box was improved to prompt the detection accuracy of the model to the bar feature and the model's generalization ability. Finally, in order to solve the problem of unbalanced positive and negative samples in the training set, an improved Non-Maximum Suppression (NMS) algorithm was adopted to optimize the learning effect. Experimental results show that the average detection accuracy of DFSSD is improved by 5.98% compared to the benchmark model SSD on dam defect images. By comparing with Faster Region-based Convolutional Neural Network (Faster R-CNN) and SSD models, it can be seen that DFSSD model has a better effect in improving the detection accuracy of dam defect objects.

Key words: object detection, engineering defect, deformable convolution, Single Shot MultiBox Detector (SSD), Non-Maximum Suppression (NMS)

中图分类号: