Edge-attention network for preserving structure in face super-resolution

Face super-resolution, commonly referred to as face hallucination, is a specific domain of super-resolution that focuses on generating high-resolution face images from their corresponding low-resolution versions. Face hallucination has received evident advances and significant attention with the expansion of deep learning methods. However, the state-of-the-art methods do not show much ability to preserve the inherent structure of the image. To address this limitation, we propose an Edge-Attention architecture that effectively maintains both local and global information in face super-resolution. Additionally, we introduce a novel loss function, the Frechet VGG Distance, which minimizes the discrepancy between the features of generated images and ground-truth images, thereby producing more realistic and high-quality results. Furthermore, we leverage the use of Unsharp Masking (UM) and Local Binary Pattern (LBP) techniques to enhance the edges of the low-resolution image. By employing the LBP method to refine the delicate edges and applying Unsharp Masking to sharpen all edges, our network excels at retrieving image structures in the super-resolution process. Experimental results provide evidence that our approach outperforms existing face super-resolution approaches in preserving facial structure and generating exceptional image quality. Moreover, our proposed method exhibits remarkable performance in enhancing the accuracy of face recognition systems when working with real-world low-resolution facial images.