Center for Transformative Infrastructure Preservation and Sustainability

Research Reports
Report Details

Title:
An End-to-End Deep Learning System for Pavement Distress Detection, Severity Estimation, and Condition Reporting
Authors:
University:
Publication Date:
Oct 2026
Report #:
CTIPS-26-010
Project #:

Abstract

Regular assessment of pavement condition is critical for safe and functional roadways, yet traditional inspections are subjective, labor-intensive, and difficult to scale. Recent AI-based approaches have automated distress detection, but they produce only numerical outputs without the plain-language explanations needed to communicate findings to engineers, agencies, and the public. To address this gap, this project introduces PaveCap, a multimodal framework that assesses pavement condition through both numerical prediction and natural language description. The framework consists of two components: (1) a single-shot pavement condition index (PCI) estimation network that combines YOLOv8, the segment anything model (SAM), and a four-layer convolutional neural network to predict PCI values on a 0 to 100 scale; (2) a dense captioning network that uses a YOLOv8 backbone, a transformer architecture, and a convolutional feed-forward module to generate structured descriptions of pavement severity and distress conditions. To train and evaluate the framework, we developed a multimodal dataset of 7,158 pavement images with bounding box annotations, expert-authored textual descriptions, and PCI values aligned with ASTM D6433 guidelines. The PCI estimation network achieved a mean absolute error of 16.21, and the dense captioning network generated accurate descriptions with BLEU-1, GLEU, and METEOR scores of 0.3799, 0.3234, and 0.4253, outperforming five alternative backbones. Overall, PaveCap advances pavement condition assessment from numerical prediction to interpretable, human-centered AI, providing transportation agencies with a scalable tool that both quantifies pavement condition and explains it in natural language to support more transparent maintenance decisions.

How to Cite

Aboah, Armstrong, Denver Tolliver, Blessing Agyei Kyem, Eugene Denteh, and Joshua Kofi Asamoah. An End-to-End Deep Learning System for Pavement Distress Detection, Severity Estimation, and Condition Reporting, CTIPS-26-010. North Dakota State University - Upper Great Plains Transportation Institute, Fargo: Center for Transformative Infrastructure Preservation and Sustainability, 2026.