Advancing scene text image super-resolution via edge enhancement priors

Scene text image super-resolution aims to simultaneously enhance the readability and resolution of low-resolution text images. Despite significant progress in this field, the issue of text and background image blending remains unresolved. Existing methods that utilize pre-trained text recognizers to guide reconstruction through text priors often overlook contextual semantics and are susceptible to interference from redundant information during the integration with text image features, leading to misguidance in text reconstruction. To address these challenges, we propose a network based on edge enhancement priors (EEP). EEP initially introduces the Canny operator and employs a pixel attention module to obtain edge-enhanced feature maps, thereby avoiding the problem of text-background blending. The edge features further enhance the text priors, aiding in the enhancement of contextual semantic information. Subsequently, we propose a novel sequence reconstruction module based on edge-enhanced priors, which reduces the impact of redundant information on the image reconstruction process and achieves superior super-resolution effects. Extensive experiments demonstrate that our EEP model can achieve remarkable performance compared to other state-of-the-art deep learning methods, with a 0.6% improvement in detection accuracy using the ASTER recognizer and a PSNR increase of 0.53 dB on the TextZoom dataset.