Enhancing facial image clarity: Deblurring gaussian blur with UNET++ architecture

Abstract

In this paper, we present a comprehensive study on enhancing the clarity of facial images through Gaussian deblurring using the UNET++ architecture. Employing the high-quality FFHQ dataset comprising 512x512 RGB images, we embark on an investigation of varying Gaussian kernel sizes, specifically 3x3, 5x5, and 7x7, applied to blurred facial images. To effectively address this deblurring task, we train three distinct UNET++ models, each tailored to deblur images corresponding to a specific kernel size. Our experimental results showcase the efficacy of our approach, with the trained models achieving an impressive Peak Signal-to-Noise Ratio (PSNR) of 39.143 dB, a Structural Similarity Index (SSIM) of 0.983 and a Multiscale SSIM (MS-SSIM) of 0.998 for a 3x3 Gaussian kernel. This substantiates the potential of employing UNET++ architecture for Gaussian deblurring tasks, offering a promising avenue for enhancing the visual quality of facial images and potentially benefiting a range of applications, from image restoration to facial recognition systems.