Label-Efficient Self-Supervised Representation Learning-based Road Pothole Detection from Thermal Images
Abstract
Potholes represent one of the leading causes of road accidents and vehicle damage globally, contributing to issues such as punctured tyres and wheel damage. Prompt and accurate detection of potholes through automated systems is crucial for ensuring road safety and infrastructure integrity. However, traditional pothole detection techniques that predominantly utilize supervised learning require extensive human-labelled datasets for training, which are both expensive and time-consuming. To address this challenge, we propose self-supervised representation learning (SSL) - based approach for pothole detection is reported and demands fewer labelled data. Our proposed approach involves two main phases: an initial pretext phase and a subsequent downstream fine-tuning phase. In the pretext phase, the model autonomously extracts robust and informative image representations from the unlabelled dataset of pothole images. In the downstream phase, the model is fine-tuned on a smaller, labelled dataset to specifically address thermal pothole detection. Extensive experimental evaluations were performed with varying proportions of labelled data (10%, 50%, and 100%), achieving accuracies of 95.71%, 98.29% and 99.2% respectively. The findings clearly demonstrate that the proposed SSL-based approach outperforms the conventional and existing supervised learning-based methods of pothole detection. Notably, the self-supervised learning-based model trained on 10% or 50% of the labelled data surpasses fully supervised models trained on the entire dataset.