计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (22): 215-222.DOI: 10.3778/j.issn.1002-8331.2007-0223

• 图形图像处理 • 上一篇    下一篇

基于改进U-Net网络的隧道裂缝分割算法研究

常惠,饶志强,赵玉林,李益晨   

  1. 1.北京联合大学 北京市信息服务工程重点实验室,北京 100101
    2.北京联合大学 城市轨道交通与物流学院,北京 100101
  • 出版日期:2021-11-15 发布日期:2021-11-16

Research on Tunnel Crack Segmentation Algorithm Based on Improved U-Net Network

CHANG Hui, RAO Zhiqiang, ZHAO Yulin, LI Yichen   

  1. 1.Beijing Key Laboratory of Information Service Engineering, Beijing Union University, Beijing 100101, China
    2.Urban Rail Transit and Logistics College, Beijing Union University, Beijing 100101, China
  • Online:2021-11-15 Published:2021-11-16

摘要:

针对复杂背景下隧道的细小裂缝图像特征难以提取以及裂缝像素类别不平衡等问题,提出了一种改进U-Net网络的隧道裂缝分割算法。将U-Net模型的编码器和解码器与残差模块相结合,使得网络参数共享,并避免出现深层网络梯度消失的问题;在此结构基础上引入挤压和激励(Squeeze and Excitation,SE)模块来提升重要特征,抑制无用特征,加强对裂缝边缘和形状等特征的权重分配;采用组合损失函数来处理裂缝像素正负样本不平衡的问题,进一步获得更加精细的分割结果。在公共隧道裂缝数据集和自制数据集上设计对比实验来验证改进模型的有效性。结果表明:该算法对裂缝的分割精度均优于其他方法,F1-Score分别达到了76.36%和75.46%,并且运行速度也有明显的提升,可以很好地满足实际工程的应用需求。

关键词: 隧道裂缝分割, U-Net网络, 残差模块, SE模块, 组合损失函数

Abstract:

Aiming at the problems of the difficult extraction of small crack image features of tunnels under complex background and the imbalance of crack pixel categories, an improved U-Net network tunnel crack segmentation algorithm is proposed. Firstly the encoder and decoder of the U-Net model are combined with the residual module to share network parameters and avoid the problem of the disappearance of deep network gradients. Secondly, based on this structure, the Squeeze and Excitation(SE) module is introduced to enhance important features, suppress useless features, and strengthen the weight distribution of crack edges, patterns and shapes. Finally, the combined loss function is used to deal with the problem of the imbalance of the crack pixel categories, and further finer segmentation results are obtained. A comparative experiment is designed on the crack data set of the public tunnel and the self-made data set to verify the effectiveness of the improved model. The results show that the algorithm is superior to other methods in segmentation accuracy. F1-Score has reached 76.36% and 75.46% respectively. The running speed has also been significantly improved, which can well meet the actual engineering application requirements.

Key words: tunnel crack segmentation, U-Net network, residual module, SE module, combined loss function