Automatic detection and location of pavement internal distresses from ground penetrating radar images based on deep learning

Abstract

Ground penetrating radar (GPR) has been prevailingly applied in nondestructive testing of internal distresses within asphalt pavement. However, the interpretation of abnormal echo signals in massive pavement GPR images is time-consuming and labor-intensive, and prone to misjudgment. To address this issue, a three-step method was proposed for the automatic detection and location of the internal distresses (eg: crack and debonding) echo features from GPR images. The on-site and numerical simulated GPR images of the asphalt pavement together formed the dataset required for the subsequent deep learning models. The first step is that You Only Look Once version 3 (YOLOv3) model predicted rectangular boxes for enclosing the internal distress echo features from GPR images. The second step involves developing the U-net models to segment the internal distress echo feature pixels in the cropped rectangular boxes. The last step is that the median points of the segmentation of the internal distress echo feature were fitted with a theoretical curve equation, to estimate the location of the internal distress. The proposed method has shown that the comprehensive detection accuracy of the internal distress echo features can reach 96.99%, the semantic segmentation accuracies of the internal distress echo feature pixels are not less than 0.856, and the average deviation of the estimated depths of the internal distresses is 3.25 cm. The research method makes further advances in accurately and automatically detecting and locating the internal distresses of asphalt pavement.