Improvement of super resolution reconstruction method for real text images

Abstract

Super resolution refers to the process of restoring a low resolution image to a high resolution image. In recent years, researchers in the field of super-resolution are not satisfied with restoring artificially defined low-resolution images, and try to restore low-resolution images in natural scenes. For this situation, a real scene text super-resolution dataset TextZoom is proposed. It contains one-to-one mapping of low-resolution and high-resolution images captured by the camera, which is more realistic and challenging than manufactured data. A super-resolution network for TextZoom, which is called TSRN is also proposed. By adding channel attention and increasing the proportion of gradient loss function, the overall network pays more attention to the restoration of text and enhances the lines, and finally improves the recognition rate of medium and difficult graded text images in the TextZoom dataset after super-resolution preprocessing.