测绘通报 ›› 2022, Vol. 0 ›› Issue (11): 74-78.doi: 10.13474/j.cnki.11-2246.2022.0328

• 学术研究 • 上一篇    下一篇

深度学习支持下的地图图片典型地理目标检测

王铮1,2, 符校3, 杜凯旋4, 刘纪平2, 车向红2   

  1. 1. 浙江省测绘科学技术研究院, 浙江 杭州 311100;
    2. 中国测绘科学研究院, 北京 100036;
    3. 广东省测绘产品质量监督检验中心, 广东 广州 510075;
    4. 武汉大学资源与环境科学学院, 湖北 武汉 430079
  • 收稿日期:2021-12-14 修回日期:2022-09-21 发布日期:2022-12-08
  • 通讯作者: 符校,E-mail:176001689@qq.com
  • 作者简介:王铮(1992-),男,硕士,主要研究方向为地理信息系统。E-mail:303448372@qq.com
  • 基金资助:
    国家自然科学基金(41901379);浙江省自然资源厅2022年科技项目(2022-52);广东省自然资源厅测绘产品质量监督检验管理系统项目(DG-2020-10-0535);国家重点研发计划(2017YFB050360103)

Detection of typical geographic object in maps based on deep learning

WANG Zheng1,2, FU Xiao3, DU Kaixuan4, LIU Jiping2, CHE Xianghong2   

  1. 1. Zhejiang Academy of Surveying and Mapping, Hangzhou 311100, China;
    2. Chinese Academy of Surveying and Mapping, Beijing 100036, China;
    3. Guangdong Surveying and Mapping Product Quality Supervision and Inspection Center, Guangzhou 510075, China;
    4. Faculty of Resources and Environment, Wuhan University, Wuhan 430079, China
  • Received:2021-12-14 Revised:2022-09-21 Published:2022-12-08

摘要: 针对地图图片中典型地理目标识别问题,本文首先介绍了两种基于深度学习的目标检测方法(YOLO网络和采用focal loss替换交叉熵损失函数的RetinaNet网络),然后将地图图片分别输入两种神经网络模型中进行训练和测试,最后对目标检测结果进行对比分析。结果表明,RetinaNet网络模型对地图图片进行目标检测的准确率有明显提高,且运行速度依然可达秒级。该地理目标检测方法的高准确度与高效性可在地图审查时节约大量人力、时间成本,为地图内容智能理解及互联网地图监管提供了新的技术参考。

关键词: 地理目标检测, 深度学习, 卷积神经网络, YOLO网络, RetinaNet网络

Abstract: Aiming at the problem of typical geographic target recognition in maps, this paper introduces two object detection methods based on deep learning: YOLO and RetinaNet which replaces cross entropy loss function with focal loss. Images are input into two neural network models for training and testing, and object detection results are compared and analyzed. The results show that the RetinaNet model has significantly improved the accuracy of object detection on maps, and the running speed is still up to seconds. The high accuracy and efficiency of the geographic object detection method can save a lot of manpower and time costs during map review, providing a new technical reference for intelligent understanding of map content and Internet map supervision.

Key words: geographic object detection, deep learning, convolutional neural network, YOLO, RetinaNet

中图分类号: