글자 인식 정보를 이용한 초해상도 글자 영상 복원

Abstract

This thesis presents a single image super-resolution (SISR) method for text image using text recognition information. Convolutional neural network (CNN) or Generative Adversarial network (GAN) based state-of-the-art SR methods have been proposed for generic image SR, but these approaches have difficulty in super-resolving text image because small text in image does not have enough information for restoring text's shape. Additionally, text images have different characteristics than natural images, so we need to consider textual information. To enhance the super-resolved text image's quality, we propose a GAN network using text recognition information. Given a input low resolution image, generator network produces sharper output super-resolve image by utilizing intermediate output and text probability map. Experimental results show that quality of our method's output is better than state-of-the-art algorithm.