Automated damage detection for open-air and underwater navigation infrastructure using generative AI-produced training data for deep learning models

Research has been continually growing toward the development of computer vision-based inspection tools for large-scale civil infrastructure; however, many deep learning techniques require large datasets to properly train models. Collecting field data can be costly and time-consuming, or may not be feasible, which has led to efforts to leverage synthetic data to supplement field data. Recent advances in text-to-image generative artificial intelligence (AI) offer the potential to quickly create realistic synthetic images of damaged infrastructure, including the complexities of the environment found in the field. In this study, the use of text-to-image generation to create a multiclass synthetic training dataset for inland navigation infrastructure is proposed, including damage of underwater structural components. Images of steel and concrete were generated that are representative of inland navigation infrastructure components. The images were labeled for semantic segmentation, and a model was trained using open-to-air and underwater scenes. The model trained using synthetic images was tested against field images, and the performance measured using recall, precision, and intersection over union was found to be comparable to a model trained using only field images. These results demonstrate that text-to-image generative AI tools were shown to be effective for generation of synthetic images with specifically defined conditions, saving time and cost, while providing a similar performance as the use of field-collected images. While intended for damage detection in large-scale civil infrastructure, this concept could be expanded to a number of areas as the generative AI models continue to improve.