講演情報
[AI-05]Enhancing Multi-Distress Pavement Detection Accuracy of Deep Learning Models Through Dataset Optimization.
*Ha Vinh Tran1, Thuyet Quoc Dang1, Kensuke Kimura1, Akihiro Kuno1, Kazuya Tomiyama2 (1. Taisei Group Future Technology Field/SATTE, Engineering Division, Taisei Rotec Corporation (Japan), 2. Research Fields in Civil and Environmental Engineering, Kitami Institute of Technology (Japan))
キーワード:
Pavement Distresses、Image Segmentation、Computer Vision、Dice Score、Model Performance
Pavement distress detection using Artificial Intelligence (AI) models has gained significant popularity in the domain of road inspection and maintenance. This approach allows management agencies to assess road pavement conditions with reduced time consumption and lower operational costs while minimizing traffic disruption. Despite these advantages, AI models face several practical challenges. For instance, predictors often yield low accuracy when detecting multiple pavement features simultaneously. The primary cause of this problem is the inconsistent graphical context arising from substantial variations in distress size and shape. Additionally, the difficulty in obtaining sufficient high-quality image data further hinders overall model performance. The current study aims to improve multi-pavement distress detection using a segmentation model by enhancing the image dataset. The primary results indicate that, while crack segmentation performance remained stable, the figures for patching and pothole improved by 15.4 and 17.4 percentage points in the Dice score (F1 score) during the training process, respectively. Furthermore, inference on testing data at various resolutions showed an increase of up to 12.1 and 11.4 percentage points in the Dice score for patching and pothole detection, respectively. Accordingly, the proposed method consistently exhibited superior performance. The outcomes of this study provide a simple yet effective methodology for practitioners to build high-performance systems for multi-distress pavement detection, thereby promoting the application of AI in road management systems.
