Evaluating Data Augmentation Techniques for GAN Performance Enhancement

GANs are one of the prominent generative AI models for artificial data generation with applications such as video generation, image generation and more.To unlock the full potential of GANs, training dataset should be carefully chosen as GANs are very sensitive to quality and diversity of training dataset. This study employs different data Augmentation techniques to mitigate these problems. The study systematically evaluates the impact of 7 different data augmentation techniques on GAN performance evaluated on CIFAR-10 dataset. The techniques used are geometric transformations, Noise Injection, Color jittering, Frequency Domain Transformation, Cutout, Mixup and CutMix. The performance of these methods is assessed using the Fréchet Inception Distance (FID), a standard metric for GAN evaluation. Results Indicate that noise injection gives the best improvement over baseline performance an improvement of 10.61% is observed, geometric and Cutmix augmentations also displayed marginal improvements. Other methods did not perform as expected and negatively affected the FID score. These findings displays the importance of selecting appropriate augmentation strategies tailored to the specific needs of GAN training. This paper provides a bedrock for optimizing data augmentation techniques to enhance the stability and performance of GANs, contributing valuable insights to the field of generative modelling.