Enhancing deep learning in structural damage identification with 3D-engine synthetic data

Abstract

Structural damage identification is crucial for maintaining infrastructure safety and durability. While deep learning-based computer vision has shown promise in this process, the scarcity of high-quality annotated data remains a challenge. To address this, synthetic data has emerged as a promising solution, enabling the creation of large and diverse datasets. This paper presents an approach that uses a 3D engine to generate synthetic crack images with controlled variations in morphology and environment, including automatic annotations. The synthetic dataset, calibrated to match real-world scales, was used to train models and significantly improved performance in detection and segmentation tasks. Experimental results showed nearly double the detection accuracy and over 2.5 times improvement in segmentation precision compared to models trained only on real data. These results demonstrate the potential of simulation-based synthetic data to improve generalization in data-scarce scenarios. This paper offers a scalable solution for structural damage detection in civil infrastructure monitoring.