Enhancing automated health monitoring of road infrastructure through a hierarchical robust deep learning approach

Abstract

The life-cycle monitoring of road health conditions is crucial for precise and automated maintenance of road infrastructure. Despite this importance, existing machine vision-based methods for identifying pavement distresses are constrained in their susceptibleness to interference from factors such as shadows. To bridge this gap, the article introduces a novel hierarchical pavement structure monitoring model, which significantly enhances the task’s performance. This proposed model integrates a de-shadowing module based on generative adversarial networks and an anchor-free instance segmentation model to provide pixel-level detection results, encompassing distress categories and morphological masks. A large-scale pavement image dataset was constructed, and a generative model was developed to expand pavement shadow samples. Experimental results show that the model achieved an average precision metric of 78.2%, surpassing existing baseline models. The effectiveness of the de-shadowing module was also validated through the significant improvements observed in various evaluation indicators. The proposed method demonstrates superior accuracy and robustness, holding promise for real-world road infrastructure monitoring systems.