Leveraging Synthetic Data for Deep-Learning-Based Road Crack Segmentation from UAV Imagery

This presentation was delivered at the DELTA 2025 - International Conference on Deep Learning Theory and Applications, and forms part of the research activities of the EvoRoads Horizon Europe project (Grant Agreement ID: 101147850). The work contributes to the project’s broader objective of advancing intelligent, data-driven solutions for road infrastructure monitoring and maintenance.

The respective paper authors are Andriani Panagi and Christos Kyrkou from the KIOS Research and Innovation Center of Excellence (KIOS CoE).

The presentation and paper investigate automated road crack detection using Unmanned Aerial Vehicles (UAVs) combined with deep learning–based image segmentation. UAV-based inspection offers clear advantages over traditional methods, as it enables large-scale monitoring without disrupting traffic or requiring costly on-site interventions. However, a major bottleneck in developing robust deep learning models for this task is the limited availability of high-quality, annotated datasets of road cracks captured in real-world conditions.

To address this challenge, the study explores synthetic data generation as a cost-effective and scalable alternative to manual annotation. Multiple synthetic datasets are generated using different strategies to vary realism and structural characteristics. These datasets are then used to train and evaluate two state-of-the-art segmentation architectures: UNet with a MobileNet encoder and UNet with an EfficientNet encoder. In addition, the impact of three commonly used loss functions - Dice Loss, Focal Loss, and Weighted Binary Cross-Entropy Loss (WBCEL) - is systematically assessed.

The models are evaluated on real crack imagery to assess their ability to generalize beyond synthetic training data. Results show that the UNet–MobileNet configuration trained with WBCEL achieves the best performance, reaching a mean Intersection over Union (mIoU) of 63.52%. A detailed analysis highlights how both the realism of synthetic data and the choice of loss function significantly influence segmentation accuracy.

Overall, this work demonstrates that carefully designed synthetic datasets, combined with appropriate model and loss function choices, can substantially improve generalization to real-world road inspection scenarios, supporting scalable and automated infrastructure monitoring within the EvoRoads framework.

Conference official website is accessible here.

Paper via the official conference proceedings is accessible here.