| Title |
Face Super-Resolution via Restormer Attention and Feedback-enhanced Facial Prior Integration |
| Authors |
(Huimin Chang) ; (Qihui Ding) |
| DOI |
https://doi.org/10.5573/IEIESPC.2025.14.5.616 |
| Keywords |
Face super-resolution; Restormer; Recurrent neural networks; Self-attention mechanisms; ; Image reconstructionㅍ |
| Abstract |
Face Super-Resolution (FSR) methods based on deep learning have made significant progress in recovering severely degraded facial images. However, existing approaches still face challenges when dealing with extremely low-resolution and noisy inputs, particularly in preserving facial structures and fine details. This paper introduces a Restormer-based Face Super-Resolution (RFSR) method that integrates the robust feature extraction capabilities of Transformers with the temporal processing advantages of recurrent neural networks. The RFSR architecture comprises four key components: an initial feature extraction module (G1), a Restormer module, a Recurrent SuperResolution module (RecurrentSRModule), and a final reconstruction module (G2). The Restormer module employs multi-head transposed self-attention mechanisms to capture long-range dependencies, effectively extracting global facial features. The RecurrentSRModule refines and enhances image details through multiple iterations. This iterative collaboration mechanism enables the network to improve reconstruction quality progressively, particularly when processing challenging low-quality inputs. Additionally, the network incorporates a residual connection that adds the upsampled original input to the network output. This design allows the main network to focus on learning highfrequency details and image enhancement while preserving low-frequency information from the original input. This approach improves reconstruction stability. It also enhances detail fidelity in the super-resolved images. Extensive quantitative and qualitative experimental results demonstrate that the proposed RFSR method outperforms existing state-of-the-art FSR approaches in recovering high-quality facial images, especially when processing extremely low-resolution and heavily noisy inputs. Our method effectively restores facial structures and texture details while maintaining identity consistency and subtle expression variations.ㅍ |