Single Face Image Super-Resolution Reconstruction with Wasserstein Generative Adversarial Networks
Yuquan Gao, Guoxi Sun, Xinzhuo Zhao · 2024
Face image super-resolution reconstruction is a crucial task for enhancing low-quality surveillance videos. Most state-of-the-art approaches employ generative adversarial networks (GANs). However, GANs are known to suffer from training instability issues. We propose a single face image super-resolution method called Single Face Image Super-Resolution Reconstruction (SFSR), based on Wasserstein GANs and residual dense networks. Without requiring any prior information, our model can directly generate high-resolution face images at 4x or 8x scale from arbitrary low-resolution inputs, outperforming traditional bilinear upsampling. To further enhance image quality, we utilize a multi-level residual dense generator with skip connections. Our loss function combines pixel-wise, perceptual, and adversarial losses to simultaneously optimize fidelity and realism. Extensive experiments on CelebA and CNBC datasets demonstrate state-of-the-art quantitative and qualitative performance, validating the efficacy of our approach for stabilizing GAN training and producing sharp, artifact-free super-resolved face images.