Denoising Diffusion-Based Image Generation Model Using Principal Component Analysis

Abstract

In recent years, advancements in GPU technology and increased data collection have significantly enhanced the performance of artificial intelligence and image generation models. However, in specific areas such as medical imaging or facial images, constraints in data collection and class imbalance issues have posed challenges to improving image quality. This study proposes the integration of Principal Component Analysis (PCA) into image generation models to address these challenges. Specifically, to overcome the limitations of conventional image generation models like GANs and VAEs, we utilize the Denoise Diffusion Probabilistic Model (DDPM) as the backbone, integrating it with PCA techniques. Using the CIFAR10 and FFHQ datasets, we evaluated the image quality of the proposed PCA-DDPM, the traditional DDPM, and DCGAN. As a result, the PCA-DDPM demonstrated superior image quality and efficiency. Notably, it maintained high performance even when trained with a limited amount of data. The findings of this research contribute significantly to the advancement of image generation technology and are expected to be applied in various domains.