Evaluation of Generative Data Augmentation Approaches in Enhancing Aircraft Damage Detection Models

Abstract

The scarcity of annotated datasets poses a major challenge to the robustness and predictive accuracy of deep learning models in detecting aircraft surface defects. This study investigates generative data augmentation as a solution to limited annotated data in aircraft surface damage detection. Carried out as part of the CINNABAR 2 project in collaboration with the DLR MRO Institute in Hamburg, Germany, the study compares Generative Adversarial Networks (GANs) and Diffusion Models in their ability to generate realistic synthetic images. The real dataset was collected using multiple acquisition systems, including smartphones and Digital Single Lens Reflex (DSLR) cameras. Synthetic image quality was evaluated through the Learned Perceptual Image Patch Similarity (LPIPS) metric, and other metrics commonly used in object detection tasks. The findings show that Diffusion Models surpass GANs, achieving lower LPIPS scores and improving detection performance by 12%, indicating greater realism, diversity, and effectiveness for training deep learning models.