Learning Generative Structure Prior for Blind Text Image Super-resolution

LR TBSRN TATT Ours (a) A real-world LR text image cropped from a Chinese invoice (b) A real-world LR text image cropped from a signboard (a) Real-world LR text image cropped from Chinese invoice (b) Real-world LR text image cropped from plaque Figure 1. Results obtained with TBSRN [7], TATT [41] and our approach on the segments of Chinese text image cropped from a real lowresolution invoice and a signboard. Both baselines are re-trained with elaborated degradations as in BSRGAN [75] and Real-ESRGAN [66].