测绘通报 ›› 2022, Vol. 0 ›› Issue (5): 74-78,88.doi: 10.13474/j.cnki.11-2246.2022.0144

• 学术研究 • 上一篇    下一篇

复杂场景下农村道路裂缝分割方法

张晋赫1,2, 秦育罗1,2,3, 张在岩1,2,4, 宋伟东1,2, 朱洪波1,2   

  1. 1. 辽宁工程技术大学测绘与地理科学学院, 辽宁 阜新 123000;
    2. 交通时空大数据研究中心, 辽宁 阜新 123000;
    3. 宿迁学院建筑工程学院, 江苏 宿迁 223800;
    4. 黑龙江科技大学矿业工程学院, 黑龙江 哈尔滨 150022
  • 收稿日期:2021-03-22 修回日期:2021-05-19 发布日期:2022-06-08
  • 通讯作者: 宋伟东。E-mail:lntu_swd@163.com
  • 作者简介:张晋赫(1997-),男,硕士生,主要研究方向为路面病害自动检测。E-mail:46046118@qq.com
  • 基金资助:
    国家自然科学基金(42071343);宿迁市指导性科技计划项目(Z2020138)

Rural road crack segmentation method in complex scene

ZHANG Jinhe1,2, QIN Yuluo1,2,3, ZHANG Zaiyan1,2,4, SONG Weidong1,2, ZHU Hongbo1,2   

  1. 1. School of Geomatics, Liaoning Technical University, Fuxin 123000, China;
    2. Institute of Spatiotemporal Transportation Data, Fuxin 123000, China;
    3. School of Civil Engineering and Architecture, Suqian College, Suqian 223800, China;
    4. School of Mining Engineering, Heilongjiang University of Science and Technology, Harbin 150022, China
  • Received:2021-03-22 Revised:2021-05-19 Published:2022-06-08

摘要: 针对农村道路裂缝识别中存在训练样本数量少、场景单一、提取结果不准确等问题,本文首先依托辽宁省多年份实测道路图像数据,构建具有多种类、多场景的路面裂缝数据集(PCDs),以ResNet50为编码器、SegNet为解码器,构建路面裂缝图像识别模型Res-SegNet,通过增大卷积核的大小获取更丰富的裂缝信息,使用Focal Loss损失函数,令模型更专注困难样本。然后采用分块预测方法提升裂缝在图片中的占比,使图片预测更加精细。最后通过网络模型和预测方法进行对比试验。结果表明,使用Res-SegNet识别PCDs的测试集,在不同的场景中F值为0.691,使用Res-SegNet结合分块预测识别PCDs的测试集,在不同的场景中F值达0.753。

关键词: 裂缝识别, 深度学习, 数据集, Res-SegNet模型, 分块预测

Abstract: Aiming at the problems of small number of training samples, single scene and inaccurate extraction results in rural road crack identification,based on the measured road image data of Liaoning province for many years,this paper constructs a pavement crack datasets (PCDs) with multiple types and scenes,using ResNet50 network as encoder and SegNet as decoder,a pavement crack image recognition network Res-SegNet is constructed,by increasing the size of the convolution kernel to obtain more abundant crack information,the Focal Loss function is used to makes the model more focused on difficult samples.The block prediction method is used to improve the proportion of fractures in the image and make the image prediction more precise.The network model and prediction method are compared:The test set of PCDs is identified by Res-SegNet in different scenes, the average F value is 0.691,Res-SegNet combined with block prediction is used to identify the test set of PCDs,and the average F value is 0.753 in different scenarios.

Key words: crack detection, deep learning, dataset, Res-SegNet model, block prediction

中图分类号: