Dual-Stage Face Restoration: Hybrid Loss Optimization and Lightweight Super-Resolution

Swjtu Leeds Joint School, Southwest Jiaotong University, Chengdu, Sichuan, China

qq1299964802@my.swjtu.edu.cn

Abstract

This study aims to address the issue of poor restoration quality in facial image restoration tasks under conditions of occlusion, low resolution, and blurriness. In complex scenarios such as digital cultural heritage preservation and security monitoring, facial restoration often faces problems of poor quality and insufficient clarity. To overcome the limitations of traditional models with single loss functions and fixed training structures, this paper proposes a dual-stage optimization framework based on Generative Adversarial Networks (GANs) to tackle issues of occlusion, low resolution, and blurriness in facial image restoration tasks. This framework first achieves preliminary semantic completion through a hybrid loss function that combines keypoint loss and perceptual loss, and then introduces the FSRCNN super-resolution network as a post-processing module to improve detail representation. Experimental results show that the face restoration model based on dual-stage optimization performs excellently on the CelebA-HQ dataset, with specific metrics being: average MSE of 0.009692, PSNR of 20.136043 dB, and SSIM of 0.631508. Compared to traditional restoration models, this method significantly alleviates common issues such as lack of clarity, organ distortion, and loss of detail. Moreover, the model also shows significant improvements in semantic coherence and detail quality, demonstrating its effectiveness in high-quality facial restoration tasks.