Bi-ESRGAN: A New Approach of Document Image Super-Resolution Based on Dual Deep Transfer Learning

Abstract

This paper proposes a new super-resolution approach for low-resolution document images based on dual deep transfer learning and GAN Architecture. It is an improvement of an already existing method but constrained by its poor caliber by its low quality on document images. In these images of complex types, it is necessary to preserve the most details and outlines of text and graphic areas. These constraints were the target of our contribution, which aims to improve the ESRGAN method. The proposed approach is called “Bi-ESRGAN”. It is based on the combination of two ESRGAN networks. The networks act in double focal on two different image maps (full image and details on the contour map) with a collaborative decision. Our approach has been tested and compared on our document image dataset that we built from document images presenting different challenges, categories, complexity levels and degradation kinds. The experimental results carried out are encouraging and confirmed the superiority of our approach compared to more than sixteen existing approaches with and without learning.