Unsupervised Document Binarization of Engineering Drawings via Multi Noise CycleGAN

• Main Authors: Rosli, L.H., Hooi, Y.K., Bin, O.K.
• Format: Article
• Published: 2023
• Online Access: http://scholars.utp.edu.my/id/eprint/37611/; https://www.scopus.com/inward/record.uri?eid=2-s2.0-85168803705&doi=10.14569%2fIJACSA.2023.0140791&partnerID=40&md5=af01ea6cb9d491b43f810d3933911448

The task of document binarization of degraded complex documents is tremendously challenging due to the various forms of noise often present in these documents. While the current state-of-the-art deep learning approaches are capable for the removal of various noise types in documents with high accuracy, they employ a supervised learning scheme which requires matching clean and noisy document image pairs which are difficult and costly to obtain for complex documents such as engineering drawings. In this paper, we propose our method for document binarization of engineering drawings using �Multi Noise CycleGAN�. The method utilizing unsupervised learning using adversarial and cycle-consistency loss is trained on unpaired noisy document images of various noise and image conditions. Experimental results for the removal of various noise types demonstrated that the method is able to reliably produce a clean image for any given noisy image and in certain noisy images achieve significant improvements over existing methods. © 2023, Science and Information Organization. All Rights Reserved.