基于像素强度顺序变换和UNetFormer的裂缝分割模型
DOI:
作者:
作者单位:

1.甘肃省经济研究院;2.兰州交通大学

作者简介:

通讯作者:

中图分类号:

TN98

基金项目:

国家自然科学基金项目(面上项目,重点项目,重大项目)


Pavement crack segmentation model based on pixel intensity order transform and UNetFormer
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    路面裂缝作为道路的最普遍病害之一,及时准确地识别和定位裂缝对道路的养护与持续健康运行具有重要意义。然而路面裂缝检测易受路面光照、路面阴影以及路面环境等复杂因素的影响,导致路面裂缝分割精度低并容易出现断裂等问题。为实现路面裂缝图像快速、准确的语义分割,本文提出了基于像素强度顺序变换和UNetFormer的路面裂缝分割模型。首先采用像素强度顺序变换算法对裂缝图片进行预处理,根据各像素与其附近像素之间的强度顺序,将图像沿着对角线四个方向转换为对比度更高的四通道图像,保留裂缝曲线结构固有特征的同时有效增强了裂缝与背景像素对比度;然后基于UNet和Transformer网络的结构特征,通过构建UNetFormer分割模型完成对路面裂缝的高精度分割,其中设计并引用了全局-局部注意力机制以充分捕获路面裂缝特征信息。最后,使用CFD、Crack200和Crack500三个开源数据集进行实例验证,实验结果表明,本文所提的裂缝分割模型F1-score分别达到83.4%、82.6%和81.9%,模型参数仅为UNet网络模型的37.7%,相较于现有的裂缝分割模型具有更高的分割精度以及更强的泛化能力。

    Abstract:

    As one of the most common diseases of roads, the timely and accurate identification and localization of cracks is of great significance to the maintenance and continuous healthy operation of roads. However, the detection of pavement cracks is easily affected by complex factors such as road illumination, road shadows, and road environment, which leads to low segmentation accuracy of pavement cracks and prone to fracture and other problems. In order to realize fast and accurate semantic segmentation of pavement crack images, this paper proposes a pavement crack segmentation model based on pixel intensity order transform (PIOT) and UNetFormer. Firstly, the PIOT algorithm is used to preprocess the crack images, and according to the intensity order between each pixel and its neighboring pixels, the image is converted into a four-channel image with higher contrast along the four directions of the diagonal, which retains the intrinsic features of the crack curve structure and effectively enhances the contrast between the cracks and the background pixels. Then, based on the structural characteristics of UNet and Transformer networks, the high-precision segmentation of pavement cracks is accomplished by constructing the UNetFormer segmentation model, in which the global-local attention mechanism is designed and invoked to fully capture the pavement crack feature information. Finally, three open-source datasets, CFD, Crack200 and Crack500, are used for example validation, and the experimental results show that the F1-score of the crack segmentation model proposed in this paper reaches 83.4%, 82.6%, and 81.9%, respectively, and the model parameter is only 37.7% of that of the UNet network model, which provides higher segmentation accuracy compared to the existing crack segmentation model and stronger generalization ability than the existing crack segmentation models.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-04-23
  • 最后修改日期:2024-07-06
  • 录用日期:2024-07-08
  • 在线发布日期:
  • 出版日期: