计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (21): 167-176.DOI: 10.3778/j.issn.1002-8331.1903-0441

• 图形图像处理 • 上一篇    下一篇

改进Mask R-CNN在航拍灾害检测的应用研究

李梁,董旭彬,赵清华   

  1. 1.太原师范学院 管理系,山西 晋中 030619
    2.太原理工大学 信息与计算机学院&新型传感器和智能控制教育部(山西)重点实验室 微纳系统研究中心,太原 030600
  • 出版日期:2019-11-01 发布日期:2019-10-30

Application Research of Improved Mask R-CNN in Aerial Disaster Detection

LI Liang, DONG Xubin, ZHAO Qinghua   

  1. 1.Department of Management, Taiyuan Normal University, Jinzhong, Shanxi 030619, China
    2.Micro-Nano System Research Center, College of Information and Computer & Key Lab of Advanced Transducers and Intelligent Control System(Ministry of Education), Taiyuan University of Technology, Taiyuan 030600, China
  • Online:2019-11-01 Published:2019-10-30

摘要: 目标检测在众多领域具有极高的理论意义与应用价值,更稳定、更精确的目标检测方法是目前灾害检测领域研究的热点与难点。将基于深度学习的目标检测方法应用于灾害检测中,提出基于改进Mask R-CNN的航拍灾害检测方法。针对检测中出现的准确率较低,采用改进特征金字塔的结构,充分利用特征映射图的信息,提高各尺寸灾害目标的检测精度;并引入在线困难样本挖掘机制,解决正负样本不均衡的问题,减少误检率和漏检率。同时采用多部件结合的方法剔除误检目标。为验证该方法的有效性,在Tensorflow深度学习框架上,选取不同高度的森林火灾、滑坡、泥石流、地震航拍图像进行验证实验。结果表明,该方法能实现对不同类型的灾害进行快速而又准确的检测,同时对基于其他应用背景的目标识别研究也具有一定的参考意义。

关键词: 深度学习, 灾害检测, 特征金字塔, 在线困难样本挖掘, 多部件结合

Abstract: Target detection has extremely high theoretical significance and application value in many fields. More stable and more accurate target detection methods have the hotspots and difficulties in the field of disaster detection. The target detection method based on deep learning is applied to disaster detection, and an aerial disaster detection method based on improved Mask R-CNN is proposed. In view of the low accuracy rate in the detection, the structure of the improved feature pyramid is used, and the information of the feature map is fully utilized to improve the detection accuracy of disaster targets of various sizes. And the online difficult sample excavator system is introduced to solve the problem of imbalance between positive and negative samples. At the same time, the multi-component combination method is adopted to eliminate the false detection target. In order to verify method effectiveness, forest fires, landslides, debris flows, and seismic aerial images of different heights are selected for verification experiments on the Tensorflow deep learning framework. Experimental results show that the proposed method can achieve fast and accurate detection of different types of disasters, which also has certain reference significance for target recognition research based on other application backgrounds.

Key words: deep learning, disaster detection, feature pyramid, online difficult sample mining, multi-part combination