Automatic detection of moisture damages in asphalt pavements from GPR data with deep CNN and IRS method

Abstract Accurate detection and localization of moisture damage in asphalt pavements using Ground Penetrating Radars (GPR) has been attracting more and more interest in research. Existing approaches rely heavily on human efforts and expert experience and are thus both time and cost consuming and are also subject to accuracy issues caused by stochastic human errors. To address this issue, this paper presents an automated moisture damage detection and localization method by leveraging the state-of-the-art deep learning approach and newly proposed incremental random sampling (IRS) approach. First, 2.3 GHz Ground coupled GPR system was used to survey moisture damages on 16 asphalt pavement bridges to create three moisture damage datasets with different resolutions including 2135 moisture damages and 474 steel joints. On this basis, we propose mixed deep convolutional neural networks (CNN) including ResNet50 network, for feature extraction, and YOLO v2 network, for recognition, to detect and localize moisture damages. In addition, to prepare the input for the deep learning models, an IRS algorithm is proposed to generate suitable GPR images from GPR data to feed the CNN. Comprehensive experimental testing, analysis, and comparison of the proposed approaches are conducted. Experimental results demonstrated the promising performance and superiority of the proposed approaches in detecting and localizing moisture damages in asphalt pavements.