PDCR-SR: Enhancing facial super-resolution with multi-scale prior dictionary and region-specific contrastive regularization

Facial super-resolution involves reconstructing high-quality facial images from low-resolution face images and restoring rich facial details. Existing algorithms often struggle with the restoration of global structural details and localized facial features. To address these challenges, we propose the PDCR-SR method, which introduces a Multi-Scale Prior Dictionary (MSPD) for leveraging high-quality features across scales, enhancing detail reconstruction. Additionally, the Region-Specific Contrastive Regularization Module (RSCR) focuses on improving the texture and accuracy of localized areas such as skin, eyes, nose, and mouth. Extensive comparison results prove that our model has better reconstruction performance on both synthetic faces and real wild faces, superior to other existing methods in terms of quantitative indicators and visual quality.