Filip Tuch
Supervisor(s): Zuzana Černeková
Comenius University
Abstract: The digitisation of historical archives is frequently hindered by physical degradation of documents, requiring costly manual restoration or automated approaches based on deep learning. Although deep learning offers powerful tools for restoration, its application in this domain raises significant safety concerns about semantic fidelity, specifically the risk of 'hallucinating' false characters that alter the meaning of the document. This paper evaluates and compares the safety and efficacy of three distinct deep learning architectures: a Generative Adversarial Network (Pix2Pix) and two Convolutional Neural Networks (DnCNN and DRUNet) in the restoration of degraded text documents. We introduce a synthetic dataset generation pipeline that simulates various types of degradation commonly found in historical documents, along with the extraction of precise textual ground truth for accurate evaluation. Beyond traditional image quality metrics, we propose a safety assessment framework that distinguishes between safe errors (deletions or data loss) and unsafe errors (substitutions or insertions) in the context of text restoration. Our results demonstrate that while the generative Pix2Pix improves visual quality and general readability, it is unsafe for document restoration due to the frequent introduction of 'hallucinated' characters. On the contrary, DnCNN and DRUNet exhibit safer restoration capabilities by minimising both safe and unsafe errors, achieving superior results in both semantic safety and image quality metrics.
Keywords: Computer Vision
Full text: Year: 2026
