From edges to pages: boundary-aware binarization and two-stage reconstruction of historical documents

Abstract

Historical documents often suffer from severe degradations such as bleed-through, stains, fading, and physical losses, which compromise both human readability and machine analysis. Historical document restoration must therefore recover legible text and faithful backgrounds while ensuring structural authenticity and usability for archival workflows. Conventional pipelines either blur fine strokes or fail to maintain visual integrity, limiting their effectiveness in practice. This thesis introduces a two-part restoration framework that jointly optimizes stroke fidelity and background reconstruction at scale. In the first part, we propose BA-GAN (Boundary-Aware Generative Adversarial Network), a robust end-to-end framework for restoring heavily degraded historical document images. BA-GAN features a single generator guided by two discriminators: one focused on objectlevel content and another on contour-level information. By leveraging both global and local information concurrently, the model improves stroke edge extraction, enhances binarization results, and ensures precise reconstruction of text boundaries. Experiments on HDIBCO 2017/2018 demonstrate state-of-the-art performance, achieving, for example, DIBCO 2018 metrics: Fm 89.28, PSNR 18.44 dB, and DRD 4.10. Beyond binarization, BA-GAN integrates a full document reconstruction framework that restores both text and background. A two-stage inpainting strategy is employed: initial background estimation via pixel-based interpolation, followed by deep learning-based GAN inpainting to seamlessly reconstruct missing content, remove noise, and correct ink bleed-through artifacts. Experiments on READ 2016 using VDQAM scores show higher evaluation scores after reconstruction, demonstrating improved visual fidelity and text legibility. This approach enables robust reconstruction of entire historical documents while preserving structural integrity and historical authenticity. Key contributions include: (i) a novel adversarial binarization framework modeled as a three-player game; (ii) a dual-discriminator cGAN architecture enabling superior stroke edge preservation; (iii) state-of-the-art performance on DIBCO benchmarks; and (iv) a documentcentric restoration pipeline combining binarization with inpainting, validated on real-world degraded manuscripts. While challenges remain in ultra-low-contrast and cross-bleed scenarios, future directions include multispectral fusion, self-supervised pretraining, and stronger contentpreservation priors.

Date	24 Nov 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Mohamed Cheriet (Supervisor)