计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (13): 281-288.DOI: 10.3778/j.issn.1002-8331.2203-0531

• 工程与应用 • 上一篇    下一篇

多层次特征融合和注意力机制的道路裂缝模型

宋榕榕,王财勇,田启川,张琪   

  1. 1.北京建筑大学 电气与信息工程学院,北京 100044
    2.建筑大数据智能处理方法研究北京市重点实验室,北京 100044
    3.中国人民公安大学 信息网络安全学院,北京 100038
  • 出版日期:2023-07-01 发布日期:2023-07-01

Road Crack Model Based on Multi-Level Feature Fusion and Attention Mechanism

SONG Rongrong, WANG Caiyong, TIAN Qichuan, ZHANG Qi   

  1. 1.School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture, Beijing 100044, China
    2.Beijing Key Laboratory of Intelligent Processing for Building Big Data, Beijing 100044, China
    3.Schoo of Information and Cyber Security, People’s Public Security University of China, Beijing 100038, China
  • Online:2023-07-01 Published:2023-07-01

摘要: 针对现有裂缝检测方法易受各类噪声干扰从而导致误检及小尺度裂缝漏检的问题,提出一种基于多层次特征融合和注意力机制的道路裂缝检测模型,采用编解码结构DeepCrack网络作为基本架构。设计单尺度多层次特征融合模块应用于特征提取,通过多级网络表征增强裂缝的细节特征。同时在编解码端对称融合部位,引入改进后的三重注意力模块,从通道、高和宽3个维度进行跨维度交互,凸显裂缝特征和抑制噪声特征,并进行跨维度的特征融合,以获得更具互补性的裂缝特征。实验表明,在道路裂缝数据集CRKWH100上,模型在多个边缘评估指标上实现了当前最优,同时在Stone331石材裂缝数据集中也验证了该模型的泛化性。

关键词: 深度学习, 编解码网络, 裂缝检测, 特征融合, 注意力机制

Abstract: In order to solve the problem that existing crack detection methods are easily interfered by various noises, which leads to false detection and missed detection of small-scale cracks, this paper proposes a road crack detection model based on multi-level feature fusion and attention mechanism, and adopts the codec structure DeepCrack network as the basic framework. A single-scale multi-level feature fusion module is designed for feature extraction, and the detailed features of cracks are enhanced through multi-level network representation. At the same time, the improved triple attention module is introduced into the symmetric fusion part of the codec end, and cross-dimensional interaction is carried out from the channel, height and width dimensions to highlight crack features and suppress noise features, and cross-dimensional feature fusion is carried out to obtain more complementary crack features. Experiments show that the model achieves the current best in several edge evaluation indexes on the road crack data set CRKWH100, and the generalization of the model is also verified in the Stone331 stone crack data set.

Key words: deep learning, encoder-decoder network, crack detection, feature fusion, attention mechanism