Semantically Guided Efficient Attention Transformer for Face Super-Resolution Tasks

Face super-resolution generates high-resolution face images from low-resolution inputs, supporting face recognition in challenging environments. While deep learning methods, like Stable Diffusion and origin Transformer-based framework, have advanced super-resolution, they require heavy computation, making them difficult for tasks like face recognition. However, recognition accuracy only needs images of size 112x112, reducing the necessity for extremely large outputs. Existing methods often rely on convolutional attention, limiting receptive fields and performance. To address this, we propose SETFSR, a Semantically Guided Efficient Attention Transformer Face Super-Resolution. Our model leverages efficient self-attention for global feature extraction and incorporates face parsing constraints for structural accuracy. Experiments on Helen, CelebA, and FFHQ datasets show that SETFSR outperforms state-of-the-art models in PSNR, SSIM, and identity preservation.