Large-scale blind face super-resolution via Edge-guided Frequency-aware Generative Facial Prior Networks

Abstract

Large-scale blind face super-resolution is a significant image processing task with high practical value and wide applications. The task is more challenging than conventional image super-resolution since the degradation is more complex and the amount of useful information is very small inside the image. To address this problem, we propose Edge-guided Frequency-aware Generative Prior Network (EFGPN), a super-resolution method based on edge guidance, frequency constraint, and generative priors. The proposed edge guidance could aid the network in retaining more structural information. Frequency constraints suppress artifacts while allowing the network to generate good features across all frequency bands. The use of generative priors enables the network to produce photorealistic faces. The proposed EFGPN can blindly super-resolve low-resolution faces to high-quality faces. Experimental results indicate that our proposed EFGPN outperforms the state-of-the-arts under subjective and objective evaluations.