디퓨전 기반 데이터 증강 평가

Abstract

In this paper, we evaluate the performance of diffusion models as image data augmentation tools for deep learning (DL) tasks. Traditional data augmentation techniques are mostly dependent on simple transformations, such as rotations, flips, or color adjustments. This approach may not fully provide with the complex variations necessary for deep learning models. Diffusion models offer the potential to generate high-quality, diverse, and realistic synthetic images. We explore their use for augmenting image datasets and the validity of the results. We demonstrate that diffusion models can provide significant results in image generation.