Enhancing facial image resolution: Leveraging UNET++ for super-resolution using deep learning

Abstract

This paper presents a novel and robust method for achieving 4X upscaling of facial images, leveraging the UNET++ architecture and applied to the widely recognized Flickr-Faces-HQ (FFHQ) dataset comprising high-resolution 512x512 images. The study systematically explores the performance of four distinct UNET++ model sizes, each offering a unique balance between computational efficiency and upscaling quality. Through extensive experimentation, the proposed approach demonstrates remarkable prowess in preserving intricate facial features, enhancing texture details, and maintaining visual realism during the upscaling process. Notably, the best-performing model exhibits an impressive average Peak Signal-to-Noise Ratio (PSNR) of 30.465 dB, a Structural Similarity Index (SSIM) of 0.859 and a Multiscale SSIM (MS-SSIM) of 0.976, establishing a new standard of excellence in facial image upscaling. These findings contribute to the advancement of image processing techniques, particularly in the context of facial imagery, and hold profound implications for a wide range of applications, including computer graphics, image enhancement, and face recognition systems.