Text Gestalt: Stroke-Aware Scene Text Image Super-resolution

In the last decade, blossom of deep learning has witnessed rapid development scene text recognition. However, recognition low-resolution images remains a challenge. Even though some super-resolution methods have been proposed to tackle this problem, they usually treat as general while ignoring fact that visual quality strokes (the atomic unit text) plays an essential role for According Gestalt Psychology, humans are capable composing parts details into most similar objects guided by prior knowledge. Likewise, when observe image, will inherently use partial stroke-level recover appearance holistic characters. Inspired we put forward Stroke-Aware Scene Text Image Super-Resolution method containing Stroke-Focused Module (SFM) concentrate on internal structures characters in images. Specifically, attempt design rules decomposing English and digits at stroke-level, then pre-train recognizer provide attention maps positional clues with purpose controlling consistency between generated image high-resolution ground truth. The extensive experimental results validate can indeed generate more distinguishable TextZoom manually constructed Chinese character dataset Degraded-IC13. Furthermore, since SFM is only used guidance training, it not bring any time overhead during test phase. Code available https://github.com/FudanVI/FudanOCR/tree/main/text-gestalt.