A U-Net based Self-Supervised Image Generation Model Applying PCA using Small Datasets

Abstract

Generative Adversarial Networks (GAN) is a research-based on deep learning technology that synthetically generates, combines, and transforms images similar to the original images. The main focus of GAN existing work has been to improve the quality of generated images and to generate high-resolution images by changing the training scheme or devising more complex models. However, these models require a large amount of data and are not suitable for training with a small amount of data. To address these challenges, this paper aims to improve the quality of images and the stability of training with a small dataset by proposing a novel training method for generating real-world images by using PCA and Self-Supervised GAN. Previously, PCA was applied to DCGAN to generate images with a small dataset, but some images showed poor results. By preparing quantitatively different datasets, we show that the quality of generated image with a small dataset is equivalent, or even better when compared to the quality of the image generated with a large dataset.