Super-resolution of scene text image via probabilistically selective fusion and context enhancement

Scene Text Image Super-Resolution (STISR) refers to enhance the resolution of text images captured in natural scenes. Text images taken from a distance,in motion,or under poor lighting conditions are often difficult to discern. However,most existing methods have not adequately considered the distinctive characteristics of scene text images. For instance,they often fail to simultaneously integrate text-specific prior knowledge with techniques tailored for super-resolution in scene text images. To address these issues,we introduce a network based on Probabilistically Selective Fusion and Context-Enhanced mechanisms in this paper,which consists of three key components:the Probabilistically Selective Fusion (PSF) module,the Text-Image Fusion (TIF) module,and the Context-Enhanced Residual (CER) module. The PSF module enhances the robustness of text priors from text recognizers through probabilistic selection,followed by the TIF module that fuses textual and visual features. Finally,the CER module refines fine text details and high-frequency information. Experimental results on the TextZoom dataset show that our method outperforms existing mainstream methods,validating its effectiveness in the applications of the super-resolution of scene text image.