Skip to main navigation Skip to search Skip to main content

Cross-Light Contrastive Learning of Pavement Crack Segmentation in Low-Light Conditions for Autonomous Road Inspection Vehicles

  • Ping Cao
  • , Chunjie Zhang*
  • , Jun Huang
  • , Yutong Wang
  • , Lili Fan
  • , Yao Zhao
  • , Xiaolong Zheng
  • *Corresponding author for this work
  • Beijing Jiaotong University
  • Macau University of Science and Technology
  • University of Chinese Academy of Sciences
  • Beijing Institute of Technology

Research output: Contribution to journal › Article › peer-review

Abstract

Regular visual monitoring of pavement defects through autonomous road inspection vehicles is critical to road infrastructure and traffic safety. Since autonomous road inspection vehicles need to operate reliably 24/7, automatic road crack segmentation under low-light conditions is particularly important for detecting road defects. However, most existing segmentation methods are designed for crack images captured under normal illumination conditions, resulting in a drastic degradation in performance when applied to low-light crack images. Low-light environments present challenges for crack segmentation due to insufficient photons and undesirable noise. To enhance crack segmentation in low-light conditions for autonomous road inspection vehicles, we propose a Cross-Light Contrastive Network(CLCNet) that incorporates normal-light crack images as auxiliary information. CLCNet comprises both light-level contrastive learning and segmentation-level contrastive learning components. Light-level contrastive learning treats images under different lighting conditions as distinct domains, aiming to learn illumination-robust features through style matrix analysis. Segmentation-level contrastive learning, on the other hand, ensures that embeddings from the same semantic class are more similar than those from different classes by comparing pixel-level embeddings. Our CLCNet can be implemented based on mainstream crack segmentation architectures and trained end-to-end using conventional crack segmentation datasets such as Crack500 and DeepCrack. We achieved superior performance over recent segmentation networks in both low-light and normal-light conditions.

Original languageEnglish
JournalIEEE Transactions on Intelligent Vehicles
DOIs
Publication statusAccepted/In press - 2026
Externally publishedYes

Keywords

  • Semantic segmentation
  • autonomous road inspection vehicles
  • contrastive learning
  • crack detection
  • low-quality visual recognition

Fingerprint

Dive into the research topics of 'Cross-Light Contrastive Learning of Pavement Crack Segmentation in Low-Light Conditions for Autonomous Road Inspection Vehicles'. Together they form a unique fingerprint.

Cite this