SANet: Face super-resolution based on self-similarity prior and attention integration

Ling Li, Yan Zhang, Lin Yuan, Xinbo Gao

Abstract: Highlights•An iterative dual-branch FSR framework based on self-similarity prior is proposed.•An attention integration mechanism is designed to exploit feature complementarity.•The self-similarity search space is expanded to mine three kinds of long-range dependencies.•A symmetric nearest neighbor sampling (SNNS) strategy is tailored to enhance the search efficiency and precision.•The cycle fully connected layer (CFCL) and extended channel projection (ECP) are optimized for efficient restoration.