Optimized Road Damage Detection Using Enhanced Deep Learning Architectures for Improved Inference Speed and Accuracy

Road infrastructure plays a critical role in the daily lives of all individuals, making regular maintenance essential for ensuring safety and longevity. However, many local governments face challenges in conducting adequate inspections due to limited financial resources. To address this issue, the Optimized Road Damage Detection Challenge 2024 (ORDCC’2024) [1] aims to advance road damage detection by utilizing datasets from six countries, with a particular focus on resource optimization through improvements in inference speed. In this study, various model architectures were explored, and extensive experiments were conducted to enhance the performance of a deep learning model for these datasets. These efforts produced significant achievements as follows: An ensemble model based on Faster R-CNN, developed for high accuracy, achieved an F1-score of 71.65% with an inference speed of 0.3012 seconds on an NVIDIA RTX 3090 Ti. Meanwhile, the YOLOv9 model, optimized for fast processing speed, reached an F1-score of 62.02% and demonstrated a significantly faster inference speed of 0.0340 seconds on an NVIDIA RTX 3090 Ti. The source code and further details can be accessed via the following repository: https://github.com/jjxxmiin/ORDDC2024.