Deep learning-based detection and evaluation of pavement surface damage

• Introduces the PavementDamagesG-7 dataset, featuring high-resolution (1920×1080) images with pixel-wise annotations of diverse pavement damage types under varied environmental conditions. • Combines DeepLabV3+ with MobileNet backbones and Optuna hyperparameter optimization to improve segmentation accuracy, boundary precision, and robustness against class imbalance. • Integrates semantic segmentation with Pavement Condition Index (PCI) evaluation, providing a scalable, engineering-aligned framework for quantitative road condition assessment from pixel-level predictions. Timely and accurate detection of pavement surface damage is essential for maintaining road safety and optimizing maintenance strategies. This study proposes a deep learning-based methodology for the semantic segmentation and evaluation of pavement distress, leveraging DeepLabV3+ with ResNet-101 and MobileNet backbone networks to achieve this. The models were fine-tuned using Optuna for hyperparameter optimization, and Deep Ensembles were applied to assess uncertainty and improve interpretability. Segmentation focused on four classes: cracks, potholes, alligator cracking, and pavement marking. This refinement led to enhanced segmentation performance, with the MobileNet backbone achieving an mIoU of 89.25% and an F1-Score of 92.09%. A notable contribution of this study is the availability of the PavementDamagesG-7 dataset, comprising 920-resolution (1920×1080) images with pixel-wise segmentation masks for various damage types. The methodology includes damage evaluation using the Pavement Condition Index, which enables alignment with the standardized engineering practices. Comparative analysis between AI-based and manual PCI assessments showed high consistency, particularly in scenarios in which the damage was localized. The model demonstrated the ability to adjust the effective evaluation area according to the actual extent of damage, thereby offering a flexible and scalable solution for infrastructure monitoring. Overall, the proposed methodology provides a scalable and robust solution for pavement monitoring, offering spatially accurate segmentation and quantifiable damage evaluation while reducing subjectivity and operational overhead compared to traditional inspection methods.