This project explores generative data augmentation to improve fine-grained image classification on a limited dataset of 75 butterfly species (5199 images). The main objective is to generate synthetic training data to boost a fixed baseline CNN classifier without modifying its architecture, optimizer, or loss function.
- João Vaz
- João Francisco Morais
We compared three generative families against the real-data baseline:
- Variational AutoEncoders (VAEs)
- Generative Adversarial Networks (GANs)
- Denoising Diffusion Probabilistic Models (DDPMs)
Generation quality matters more than augmentation quantity. The final performance hierarchy is DDPM > GAN > Baseline > VAE:
-
DDPM: Best performance (+18.85 pp macro-F1, 86.61% accuracy) driven by superior image fidelity (FID 39.48).
-
GAN: Improved performance (+6.99 pp macro-F1), heavily relying on Adaptive Discriminator Augmentation (ADA) to prevent memorization.
-
VAE: Degraded classification (-2.72 pp macro-F1) due to blurry and washed-out generated textures.