Text Enhancement for Historical Handwritten Documents

This paper presents a text enhancement method for historical handwritten documents. Text enhancement is a sub-field of super-resolution focused on document images and text. Our work is based on generative adversarial networks (GAN), and we have introduced a new evaluation metric for document image enhancement. Our approach denoises historical documents and holistically increases their resolution using GANs. We modified The generator structure of our GAN model by replacing Batch Norm layers with Residual-in-Residual Dense Blocks (RRDB) and adopting a discriminator based on the Relativistic GAN. Our evaluation metric for text enhancement focuses on text quality based on the magnitude of gradients at text edges to assess the improvement of the generated images. We tested our method on three degraded handwritten historical datasets of two languages and obtained excellent results. In addition, We compare our approach with SRGAN, Nearest, and Bi-Cubic interpolations and show that our method performs much better than these methods on the three datasets. The proposed approach can handle various types of noise while applying text enhancement up to 16 times the input image.