Document Binarization with Quaternionic Double Discriminator Generative Adversarial Network.

Quaternionic networks have emerged as a lightweight alternative to standard neural networks. We propose using a Quaternionic conditional Generalized Adversarial Network adapted to document image binarization. A double discriminator ensures that the output is consistent over a coarse and a finer level of resolution, while the generator is tasked with producing the binarized document. We achieve excellent binarization results, while our network is significantly smaller (4x smaller) than its real-valued counterpart.