计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (10): 204-210.DOI: 10.3778/j.issn.1002-8331.2003-0096

• 图形图像处理 • 上一篇    下一篇

增强语义信息与多通道特征融合的裂缝检测

顾书豪,李小霞,王学渊,张颖,陈菁菁   

  1. 1.西南科技大学 信息工程学院,四川 绵阳 621010
    2.特殊环境机器人技术四川省重点实验室,四川 绵阳 621010
  • 出版日期:2021-05-15 发布日期:2021-05-10

Crack Detection Based on Enhanced Semantic Information and Multi-channel Feature Fusion

GU Shuhao, LI Xiaoxia, WANG Xueyuan, ZHANG Ying, CHEN Jingjing   

  1. 1.School of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
    2.Key Laboratory of Special Environmental Robotics in Sichuan Province, Mianyang, Sichuan 621010, China
  • Online:2021-05-15 Published:2021-05-10

摘要:

路面裂缝检测是用以判断道路安全与否的关键技术,由于裂缝的背景复杂多样,传统的裂缝检测算法难以准确检测裂缝。提出了一种增强语义信息与多通道特征融合的裂缝自动检测算法。网络整体为编码器-解码器结构,在编码器部分引入扩张卷积模块,扩大特征图有效感受野,整合图像上下文信息,增强特征语义表达能力,提高像素分类精度。在解码器部分搭建了一个基于注意力机制的多通道特征融合模块,利用高层全局注意力信息指导高层语义特征与低层细节特征的逐级融合,有利于恢复图像细节信息,进一步提升对裂缝的像素级检测精度。实验结果表明,在CRACK500公开数据集上训练的模型在测试集上取得72.5%的平均交并比(Intersection over Union,IoU)和96.8%的F1score,该模型直接用于CrackForest数据集测试,平均IoU和F1score分别提升2.0个百分点和1.1个百分点,表明模型具有很好的泛化性能,可用于复杂道路场景下的裂缝检测与质量评估。

关键词: 裂缝检测, 扩张卷积, 有效感受野, 注意力机制, 特征融合

Abstract:

Pavement crack detection is a critical technology used to judge the safety of roads. Due to the complex and diverse background of cracks, it is difficult for traditional crack detection algorithms to accurately detect cracks. This paper proposes an automatic crack detection algorithm that enhances the fusion of semantic information and multi-channel features. The entire network is an encoder-decoder structure. A dilated convolution module is introduced in the encoder part to expand the effective receptive field of the feature map, integrate the image context information, enhance the semantic expression ability of features, and improve the accuracy of pixel classification. In the decoder part, a multi-channel feature fusion module based on attention mechanism is built, which uses high-level global attention information to guide the step-by-step fusion of high-level semantic features and low-level detail features. It is helpful to restore the detailed information of the image and further improve the pixel-level detection accuracy of cracks. The experiment results show that the model trained on the CRACK500 public data set achieves an average Intersection over Union(IoU) of 72.5% and an F1score of 96.8% on the test set. This model is directly used on the CrackForest data set for test, and the average IoU and F1score are improved by 2.0 percentage points and 1.1 percentage points respectively, indicating that the model has good generalization performance and can be used for crack detection and quality evaluation in complex road scenarios.

Key words: crack detection, dilated convolution, effective receptive field, attention mechanism, feature fusion