Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (4): 1062-1068.DOI: 10.11772/j.issn.1001-9081.2019081510

• Artificial intelligence • Previous Articles     Next Articles

Third-party construction target detection in aerial images of pipeline inspection based on improved YOLOv2 and transfer learning

CHEN Guihui1, YI Xin1, LI Zhongbing1, QIAN Jiren1,2, CHEN Wu1,2   

  1. 1. College of Electrical Engineering and Information, Southwest Petroleum University, Chengdu Sichuan 610500, China;
    2. Technology Innovation Center, Zhejiang ZheNeng Natural Gas Operation Company Limited, Hangzhou Zhejiang 310000, China
  • Received:2019-09-05 Revised:2019-11-11 Online:2020-04-10 Published:2019-11-18
  • Supported by:
    This work is partially supported by the Nanchong Science and Technology Strategic Cooperation Project of City and University (18SXHZ0041).

基于改进YOLOv2和迁移学习的管道巡检航拍图像第三方施工目标检测

谌贵辉1, 易欣1, 李忠兵1, 钱济人1,2, 陈伍1,2   

  1. 1. 西南石油大学 电气信息学院, 成都 610500;
    2. 浙江浙能天然气运行有限公司 科创中心, 杭州 310000
  • 通讯作者: 易欣
  • 作者简介:谌贵辉(1971-),男,四川三台人,教授,硕士,主要研究方向:智能仪表、图像处理;易欣(1994-),男,四川广安人,硕士研究生,主要研究方向:图像处理、目标检测;李忠兵(1985-),男,湖北武汉人,讲师,博士,主要研究方向:精密仪器、现代信号处理。
  • 基金资助:
    南充市市校科技战略合作项目(18SXHZ0041)。

Abstract: Aiming at the few datasets and low detection rate when the traditional target detection algorithm applying to third-party construction target detection and illegally occupied building detection in the aerial images of drone,an aerial image target detection algorithm based on Aerial-YOLOv2 and transfer learning was proposed. Firstly,the trained network combining with data enhancement and transfer learning strategy was used to expand the dataset size,and K-means clustering analysis was used to obtain the number and size of anchor blocks that meet the characteristics of the proposed dataset. Secondly,the adaptive contrast enhancement was used to pre-process the image. Finally,the improved convolution module was proposed to replace the convolution block in YOLOv2,and the feature fusion multi-scale prediction method was combined for target detection. The comparison experiments of different algorithms and training strategies on the aerial images of drone were carried out. Results show that the accuracy and recall rate of the Aerial-YOLOv2 algorithm combined with various training strategies can respectively reach 95% and 91%,and the detection time per image is 14 ms. It can be seen that the algorithm is suitable for the intelligent detection of third-party construction targets and illegally occupied buildings in the aerial images of drone.

Key words: pipeline inspection, deep learning, aerial image, small-object detection, data enhancement, transfer learning

摘要: 针对传统目标检测算法应用在无人机航拍图像上第三方施工目标检测和违章占压建筑检测的数据集少、检测率低等问题,提出基于Aerial-YOLOv2和迁移学习的航拍图像目标检测算法。首先,利用结合数据增强的迁移学习策略训练的网络来扩大数据集规模,并利用K均值聚类分析得到符合所提数据集特点的锚点框数量和尺寸;其次,通过自适应对比度增强的方法对图像进行预处理;最后,提出改进卷积模块替代YOLOv2中的卷积块并结合特征融合的多尺度预测方式进行目标检测。用不同的算法和训练策略在无人机航拍图像上进行对比实验,实验结果表明,Aerial-YOLOv2算法结合多种训练策略后,其准确率、召回率分别能达到95%、91%,每张图像检测时间为14 ms。由此可知,该算法适用于无人机航拍图像第三方施工目标及违章占压建筑的智能检测。

关键词: 管道巡检, 深度学习, 航拍图像, 小目标检测, 数据增强, 迁移学习

CLC Number: