测绘通报 ›› 2022, Vol. 0 ›› Issue (3): 83-89.doi: 10.13474/j.cnki.11-2246.2022.0082

• 学术研究 • 上一篇    下一篇

改进的HRNet应用于路面裂缝分割与检测

张伯树1,2,3, 张志华1,2,3, 张洋1,2,3   

  1. 1. 兰州交通大学测绘与地理信息学院, 甘肃 兰州 730070;
    2. 地理国情监测技术应用国家地方联合工程研究中心, 甘肃 兰州 730070;
    3. 甘肃省地理国情监测工程实验室, 甘肃 兰州 730070
  • 收稿日期:2021-04-01 修回日期:2022-01-25 出版日期:2022-03-25 发布日期:2022-04-01
  • 通讯作者: 张志华。E-mail:43447077@qq.com
  • 作者简介:张伯树(1997-),男,硕士生,研究方向为计算机视觉、智能图像识别。E-mail:1113993479@qq.com
  • 基金资助:
    国家自然科学基金(41861059);兰州交通大学优秀平台(201806)

Improved HRNet applied to segmentation and detection of pavement cracks

ZHANG Boshu1,2,3, ZHANG Zhihua1,2,3, ZHANG Yang1,2,3   

  1. 1. Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China;
    2. National-Local Joint Engineering Research Center of Technologies and Applications for National Geographic State Monitoring, Lanzhou 730070, China;
    3. Gansu Provincial Engineering Laboratory for National Geographic State Monitoring, Lanzhou 730070, China
  • Received:2021-04-01 Revised:2022-01-25 Online:2022-03-25 Published:2022-04-01

摘要: 针对利用传统卷积神经网络进行路面裂缝分割时存在准确率低、信息丢失及边缘模糊的问题,本文提出了基于改进HRNet模型的路面裂缝分割算法。模型在原始HRNet的基础上进行改进,主干网络部分采用DUC模块代替双线性插值上采样;下采样改为passthrough layer代替原始卷积;在模型解码部分,进行逐级上采样的同时引入SE-Block,对不同特征层的融合重新标定权重。通过与原始HRNet及传统卷积神经网络U-Net对比可知,本文算法在公共数据与自制数据集上的分割精度表现优秀,F1分值分别达到了91.31%和78.69%,可以很好地满足实际工程的需求。

关键词: 路面裂缝;HRNet;DUC;passthrough layer;SE-Block;图像分割

Abstract: Aiming at the problems of low accuracy,loss of information and blurred edges in the traditional convolutional neural network for pavement crack segmentation,a pavement crack segmentation algorithm based on the improved HRNet model is proposed.The model is improved on the basis of the original HRNet,the backbone network part uses DUC module instead of bilinear interpolation;downsampling is changed to passthrough layer to replace the original convolution,SE-Block is introduced while performing step-by-step upsampling to re-calibrate the fusion of different feature layers.Comparing with the original HRNet and the other traditional convolutional neural networks U-Net,it can be concluded that the segmentation accuracy of this algorithm is the best on public data and self-made data sets,with F1 score reaching 91.31% and 78.69% respectively,proving that the algorithm can be very good to meet the needs of actual engineering.

Key words: road cracks;HRNet;DUC;passthrough layer;SE-Block;image segmentation

中图分类号: