Abstract
Recent advances in Deep Generative Models (DGM) have demonstrated their remarkable ability to synthesize highly realistic images. This work examines the effectiveness of DGM produced synthetic images in the context of training AI models for visual asset inspection, which is plagued by a host of challenges including limited data volumes, extremely imbalanced datasets owing to rarely occurring defects, poor quality and consistency of images, spiraling costs and time for data labeling, to name a few. This paper presents an approach to using DGMs, specifically Generative Adversarial Network (GAN) models, to create synthetic images containing defects and damage. The objective of the approach goes beyond generating good quality images and extends to synthesizing a distribution of images with a variety of defect features without significantly altering the object structure.
The paper concludes by discussing multiple metrics to assess the effectiveness of such synthetic data in training downstream asset inspection models and presents the results of multiple experiments on real-world data.
