Automatic Road Subsurface Distress Recognition from Ground Penetrating Radar Images using Deep Learning-based Cross-verification
Summary: Ground penetrating radar (GPR) has emerged as a rapid and non-destructive technique for detecting road subsurface distress (RSD). However, the process of recognizing RSD from GPR images remains labor-intensive and heavily relies on expert inspectors. Recent advancements in deep learning have provided new avenues for automatic RSD recognition, yet challenges persist in terms of defect recognition capabilities.
This article discusses a novel cross-verification strategy designed to leverage the complementary strengths of region proposal networks in recognizing objects from various perspectives of GPR images. The study utilizes three YOLO-based models to identify RSD, specifically voids and loose structures, as well as manholes. Each model is trained on a distinct view of a 3D GPR dataset consisting of 2134 meticulously validated samples obtained through field scanning.
Introduction
Ground penetrating radar has gained traction as an effective method for assessing road conditions without causing any disruption to the infrastructure. Traditional methods of distress recognition require significant human intervention, which is not only time-consuming but also introduces the potential for human error. To overcome these limitations, researchers have turned to deep learning techniques that can automate the recognition process. However, the effectiveness of these methods can be limited by the complexity of the data.
Methodology
The study proposes a cross-verification strategy that enhances the recognition capabilities of deep learning models. By employing three YOLO-based models, each trained on a different perspective of the GPR dataset, the researchers aimed to improve the accuracy of distress detection.
- Dataset: The 3D GPR dataset consists of 2134 rigorously validated samples, including a variety of distress types.
- Model Training: Each YOLO model was trained on a specific view of the GPR images, allowing for a comprehensive understanding of the data from multiple angles.
- Cross-verification: The models utilized a cross-verification strategy to combine their outputs, thereby enhancing the overall accuracy of RSD detection.
Results
The implementation of the cross-verification strategy yielded remarkable results, achieving a recall rate of over 98.6% in tests conducted with real field-scanning data. This high level of accuracy demonstrates the potential of deep learning-based automatic recognition systems in reducing the reliance on human inspectors.
Conclusion
Field tests indicate that the integration of deep learning techniques in RSD recognition can significantly diminish the labor associated with road inspections, reducing the need for human involvement by approximately 90%. This advancement not only streamlines the inspection process but also enhances the reliability of distress detection, ultimately contributing to better-maintained road infrastructure.
As the field continues to evolve, further research into the optimization of deep learning models and their applications in road maintenance is essential. The promising results from this study pave the way for future innovations in non-destructive testing methods and automated inspection systems.