SupRes: Facial Image Upscaling Using Sparse Denoising Autoencoder

Abstract

Even in this era of digital images, still many images and media are hazy, pixelated, and blurry. This could be due to low-quality imaging sensors, poor image stabilization, or the image itself being old. This study proposes the Sparse Denoising Autoencoders (SDAEs) for upscaling blurry images. The performance of the proposed SDAEs is then compared with the deep learning architecture, Pix2Pix Generative Adversarial Networks (GANs) by primarily focusing on the facial images. The experimental results show that the SDAEs give slightly better results than GANs. Additionally, the SDAE architecture is computationally 30% efficient when compared to the Pix2Pix GAN.