Document Image Binarization Using Dual Discriminator Generative Adversarial Networks

For document image analysis, image binarization is an important preprocessing step. Also, binarization can help in improving the readability of old and historical manuscripts. Such documents are generally degraded due to various reasons such as bleed-through, faded ink, or stains. Achieving good binarization performance on these documents is a challenging task. In this letter, a deep learning based model for document image binarization has been proposed, comprising a Dual Discriminator Generative Adversarial Network (DD-GAN) which uses Focal Loss as generator loss. The DD-GAN consists of two discriminator networks - one looks for the global similarity i.e. on the whole image, and another one explores the image in small patches i.e. local similarity. At the final stage, simple thresholding is performed on the generated images. The method has been tested on five recent DIBCO datasets. It has been found that the method is robust and it provides results comparable with state-of-the-art methods. The code for this letter is available at https://github.com/anuran-Chakraborty/BinarizationDualDiscriminatorGAN.