CL-PSDD: Contrastive Learning for Adaptive Generalized Pavement Surface Distress Detection

The accurate detection of pavement surface distress is crucial for timely maintenance and prevention of traffic accidents. Traditional deep-learning techniques, while effective, rely heavily on supervised learning, necessitating extensive and costly annotated datasets. This paper presents CL-PSDD(Contrastive Learning for Adaptive Generalized Pavement Surface Distress Detection), a novel unsupervised contrastive learning approach that overcomes these limitations by capturing robust feature representations of pavement surface distress across various conditions without manual labeling. We introduce a synergistic methodology that combines SwAV for effective unsupervised learning and YOLO for efficient object detection, enabling the model to learn intricate feature representations of pavement surface distress. Furthermore, we propose RE-RPN-K(Road Enhanced Region Proposal Network in K), a specialized region proposal network that distills precise road surface information, and integrates a transformer-based DETR-Head, eliminating the need for conventional NMS(Non-Maximum Suppression) post-processing. Our experiments on the GRDDC and IRRDD datasets demonstrate that CL-PSDD achieves competitive mean Average Precision (mAP) scores of 76% and 60%, respectively, rivaling the performance of supervised methods and excelling in generalization. This study pioneers the application of unsupervised learning to extensive pavement surface distress detection, establishing a significant milestone and providing a valuable reference for future research in this vital field.

• Record URL:
  • https://doi.org/10.1109/TITS.2024.3525193
• Availability:
  • Find a library where document is available. Order URL: http://worldcat.org/oclc/41297384
• Supplemental Notes:
  • Copyright © 2025, IEEE. The contents of this paper reflect the views of the author[s] and do not necessarily reflect the official views or policies of the Transportation Research Board or the National Academy of Sciences.
• Authors:
  • Dong, Ruchan
  • Xia, Jinwei
  • Zhao, Jin
  • Hong, Lei
• Publication Date: 2025-4

Language

• English

Media Info

• Media Type: Web
• Features: References;
• Pagination: pp 5211-5224
• Serial:
  • IEEE Transactions on Intelligent Transportation Systems
  • Volume: 26
  • Issue Number: 4
  • Publisher: Institute of Electrical and Electronics Engineers (IEEE)
  • ISSN: 1524-9050
  • Serial URL: http://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6979

Subject/Index Terms

• TRT Terms: Detection and identification systems; Machine learning; Object detection; Pavement distress; Pavement surface course
• Subject Areas: Data and Information Technology; Highways; Maintenance and Preservation;

Filing Info

• Accession Number: 01975723
• Record Type: Publication
• Files: TRIS
• Created Date: Jan 6 2026 9:16AM