Face Super Resolution Based on Identity Preserving V-Network

Abstract

Numerous super-resolution methods have been developed to restore and upsample low-resolution and low-detail images to higher resolutions. Specifically, face super-resolution studies aim to restore various degradations in facial images while enhancing their resolution and preserving details. This study proposes the VNet architecture, which consists of a deep learning-based convolutional network for converting low-resolution and degraded facial images into high-quality and detailed images, and a pre-trained FaceNet model to preserve identity (biometric) information. The architecture leverages the advantages of the Encoder-Decoder structure bidirectionally to maintain details and recover lost information. In the initial stage, the Encoder module compresses the image representation, filtering out unnecessary information. The Decoder module then reconstructs the high-resolution and restored image from the compressed representation. The use of residual connections in this process helps minimize information loss while preserving details. The final stage utilizes the identity loss feedback from the FaceNet model to enhance the image without deviating from the original identity context. Tests conducted on various facial datasets demonstrate that VNet achieves high metric performance in both super-resolution and restoration tasks. The results indicate that the proposed architecture is effective in producing realistic and high quality versions of low-resolution and degraded facial images.

Full Text