《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3353-3362.DOI: 10.11772/j.issn.1001-9081.2024101486

• 前沿与综合应用 • 上一篇    

渐进式上下文交互和注意力机制的混凝土路面裂缝检测网络

尹学辉1(), 傅林琳1, 周尚波2   

  1. 1.重庆邮电大学 软件工程学院,重庆 400065
    2.重庆大学 计算机学院,重庆 401331
  • 收稿日期:2024-10-24 修回日期:2025-01-17 接受日期:2025-01-22 发布日期:2025-03-14 出版日期:2025-10-10
  • 通讯作者: 尹学辉
  • 作者简介:尹学辉(1986—),男,四川广安人,副教授,博士,CCF会员,主要研究方向:计算机视觉、机器学习、智能软件工程 Email:yinxh@cqupt.edu.cn
    傅林琳(2000—),女,重庆人,硕士研究生,主要研究方向:图像分割、深度学习
    周尚波(1963—),男,广西宁明人,教授,博士生导师,博士,CCF会员,主要研究方向:视频信号处理、人工神经网络。
  • 基金资助:
    国家自然科学基金资助项目(62176034)

Concrete pavement crack detection network with progressive context interaction and attention mechanism

Xuehui YIN1(), Linlin FU1, Shangbo ZHOU2   

  1. 1.School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China
    2.College of Computer Science,Chongqing University,Chongqing 401331,China
  • Received:2024-10-24 Revised:2025-01-17 Accepted:2025-01-22 Online:2025-03-14 Published:2025-10-10
  • Contact: Xuehui YIN
  • About author:YIN Xuehui, born in 1986, Ph. D., associate professor. His research interests include computer vision, machine learning, intelligent software engineering.
    FU Linlin, born in 2000, M. S. candidate. Her research interests include image segmentation, deep learning.
    ZHOU Shangbo, born in 1963, Ph. D., professor. His research interests include video signal processing, artificial neural networks.
  • Supported by:
    National Natural Science Foundation of China(62176034)

摘要:

为保障道路质量与安全,自动化裂缝检测在混凝土路面维护中至关重要。针对现有的基于深度学习的裂缝检测方法因过度下采样导致裂缝像素信息丢失的问题,提出一种基于渐进式上下文交互和注意力机制的混凝土裂缝检测网络。首先,以优化后的UNet++为主干,采用非对称卷积块增强特征提取能力;其次,引入渐进式上下文交互机制(PCIM)以高效地捕捉与融合相邻特征图的多尺度特征;再次,在特征增强阶段,用注意力组合(AC)方式提高特征表达能力;最后,在特征融合阶段,使用多语义注意力动态融合模块(MADFM)增强细节恢复和保留效果。在3个公开数据集上的测试结果表明,相较于DeepCrack、CrackFormer、PAF-Net(Progressive and Adaptive Feature Fusion Network),所提网络的性能更优。在DeepCrack测试集上,所提网络的F-score分别提高了1.33、5.07和3.93个百分点;在Crack500测试集上,分别提升了3.04、4.35和0.82个百分点;在CFD测试集上,分别提升了3.03、6.00和4.73个百分点。这些结果充分验证了所提网络在裂缝检测准确性方面实现了提升,同时在不同测试集上均展现出了出色的鲁棒性能。

关键词: UNet++, 裂缝检测, 非对称卷积, 多尺度特征融合, 注意力机制

Abstract:

To ensure road quality and safety, automated crack detection is crucial for the maintenance of concrete pavement. To address the issue of pixel information loss caused by excessive down-sampling in the existing deep learning-based crack detection methods, a concrete crack detection network based on progressive context interaction and attention mechanisms was proposed. Firstly, with an optimized UNet++ as the backbone, asymmetric convolution blocks were applied to enhance feature extraction ability. Secondly, Progressive Context Interaction Mechanism (PCIM) was introduced to capture and fuse multi-scale features of adjacent feature maps efficiently. Thirdly, in the feature enhancement phase, the Attention Combination (AC) approach was used to improve feature representation capability. Finally, in the feature fusion phase, a Multi-Semantic Attention Dynamic Fusion Module (MADFM) was utilized to enhance detail recovery and retention effects. Test results on three public datasets show that compared to DeepCrack, CrackFormer, and PAF-Net (Progressive and Adaptive feature Fusion Network), the proposed network achieves superior performance. Specifically, the proposed network has the F-score improved by 1.33, 5.07, and 3.93 percentage points, respectively, on the DeepCrack test set; enhanced by 3.04, 4.35, and 0.82 percentage points, respectively, on the Crack500 test set; and increased by 3.03, 6.00, and 4.73 percentage points, respectively, on the CFD test set. These results verify fully that the proposed network achieves enhanced accuracy in crack detection and has excellent robust performance on different test sets.

Key words: UNet++, crack detection, asymmetric convolution, multi-scale feature fusion, attention mechanism

中图分类号: