Scholarlink

Conventional methods for assessing the condition of urban road surfaces are predominantly manual, demanding significant time and effort, and often leading to inconsistent or inaccurate results due to human error. To address these limitations, the present study evaluates two state-of-the-art deep learning-based object detection models—Faster R-CNN and Detection Transformer (DETR)—for identifying and classifying surface-level road damage. Both models were fine-tuned using transfer learning and tested on widely recognized datasets, RDD2018 and RDD2022. The performance of Faster R-CNN was assessed with two different backbone networks: VGG16 and ResNet50. On RDD2018, it achieved 69.3% and 73.3% accuracy, respectively, while on RDD2022, it yielded 58.9% and 61.3%. DETR, paired with a ResNet50 backbone, surpassed these configurations, achieving 78.4% accuracy on RDD2018 and 63.5% on RDD2022. Despite its superior detection capability, DETR's higher computational overhead and slower convergence make Faster R-CNN a more suitable option for real-time applications where computational resources are limited. Overall, this study demonstrates the potential of deep learning to significantly improve pavement condition monitoring, while also offering insights into the practical balance between model accuracy and resource efficiency.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Faster R-CNN vs. Detection Transformer: Transfer Learning Approaches for Road Damage Detection and Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • R. Rakshitha,
  • S. Srinath,
  • N. Vinay Kumar,
  • S. Rashmi,
  • B. V. Poornima

摘要

Conventional methods for assessing the condition of urban road surfaces are predominantly manual, demanding significant time and effort, and often leading to inconsistent or inaccurate results due to human error. To address these limitations, the present study evaluates two state-of-the-art deep learning-based object detection models—Faster R-CNN and Detection Transformer (DETR)—for identifying and classifying surface-level road damage. Both models were fine-tuned using transfer learning and tested on widely recognized datasets, RDD2018 and RDD2022. The performance of Faster R-CNN was assessed with two different backbone networks: VGG16 and ResNet50. On RDD2018, it achieved 69.3% and 73.3% accuracy, respectively, while on RDD2022, it yielded 58.9% and 61.3%. DETR, paired with a ResNet50 backbone, surpassed these configurations, achieving 78.4% accuracy on RDD2018 and 63.5% on RDD2022. Despite its superior detection capability, DETR's higher computational overhead and slower convergence make Faster R-CNN a more suitable option for real-time applications where computational resources are limited. Overall, this study demonstrates the potential of deep learning to significantly improve pavement condition monitoring, while also offering insights into the practical balance between model accuracy and resource efficiency.