SwinUnetSR: A Transformer-based Encoder-Decoder Structure with a Lightweight Upsampler for Face Super Resolution

Face Super-Resolution (FSR) is a critical task aimed at enhancing low-resolution (LR) face images into high-resolution (HR) ones while preserving essential facial features. This task has broad applications, including identity recognition in surveillance systems and facial detail restoration for biometric authentication. In this paper, a novel hybrid architecture, SwinUnetSR, is proposed for FSR tasks. Built on the SwinV2-B transformer, the model integrates it within a Convolutional Networks for Biomedical Image Segmentation (U-Net) framework, followed by a lightweight upsampler for HR image reconstruction. The encoder, based on SwinV2-B, leverages hierarchical attention mechanisms to efficiently process global contextual information and down-sample facial features. U-Net serves as the decoder, where skip connections fuse the compressed features with SwinV2-B outputs. A lightweight upsampler then upscales the feature maps into HR images. Experimental results demonstrate that SwinUnetSR achieves high Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) scores, indicating its effectiveness. While computational resource limitations and data scarcity prevent it from outperforming state-of-the-art models, the high evaluation scores confirm the feasibility of SwinUnetSR for FSR. Furthermore, the model holds promise for future expansion to more complex scenarios. Code is available at: https://github.com/KbKuuhaku/swin-unet-sr.