Dataset for: Multiple-type distress detection in asphalt concrete pavement using infrared thermography and deep learning

Description

Abstract

Artificial intelligence, particularly Convolutional Neural Network (CNN), has emerged as a highly effective methodology for detecting pavement distresses. This study aimed to apply infrared thermography (IRT) and deep learning to multiple-type distress detection. The dataset encompassed five image types (visible images, infrared images, and fusion images with varying infrared ratios) along with five distress types (longitudinal cracking, transverse cracking, fatigue cracking, edge cracking, and potholes). Four CNN object detection models underwent training and evaluation on the dataset, employing transfer learning. Evaluation metrics encompassed accuracy, complexity (model and computation), and memory usage. Eigen-CAM was employed to interpret the performance of CNN models across diverse image and distress types. The study delved into the impact of image types on multiple-type distress detection and also explored the potential of infrared thermography for pavement distress detection, including multiple-type distress detection, crack severity classification, and crack segmentation. Results indicated that fusion images (25IRT) achieved the highest accuracy across all four CNN models, closely followed by visible images and fusion images (50IRT). YOLOv5 demonstrated the highest accuracy for all image types except fusion images (50IRT). Fatigue cracking consistently exhibited the highest accuracy across all image types and CNN models, surpassing longitudinal cracking and edge cracking, which performed similarly and significantly outperformed transverse cracking and potholes. YOLOv5 provided clear visual explanations (Eigen-CAM) across all image types. In conclusion, fusion images could be an accurate, efficient, and reliable alternative solution for pavement distress detection.