Comparative Analysis of Generative Data Augmentation Techniques for Aircraft Damage Detection Algorithms: A Case Study

Abstract

This study addresses the challenge of limited annotated data in aircraft surface damage detection by evaluating generative models for data augmentation. Conducted within the CINNABAR 2 project with DLR MRO institute in Hamburg (DE), it compares Generative Adversarial Networks (GANs) and Diffusion Models for producing realistic synthetic images. Real data were collected from smartphones, DSLRs, and robotic camera systems. Image quality was assessed using the Learned Perceptual Image Patch Similarity (LPIPS) metric and visual inspection. Results indicate that Diffusion Models outperform GANs, achieving a lower LPIPS score and better detection metrics, demonstrating superior realism, diversity, and suitability for enhancing deep learning model training.