Fully Automatic Virtual Unwrapping Method for Documents Imaged by X-Ray Tomography

The study of historical documents faces challenges due to aging, particularly when rolled or folded, risking damage during unfolding. While computer tomography enables 3D digital replicas, direct examination is inconvenient. To facilitate content analysis, various virtual unfolding methods have been proposed. We present a groundbreaking, fully automated system for virtual unfolding/unrolling, employing a neural network to generate a binary document mask and perform skeletonization on 2D sections of the 3D volume. Additional algorithms address artifacts, false loops, branching, and discontinuities. Introducing a unified coordinate system for skeletal sections allows the generation of an unfolded document image. Performance is assessed on the CT-OCR-2022 dataset, utilizing a novel criterion for geometric distortion evaluation. Enriched with marker coordinates, the dataset facilitates future algorithm assessments. Accompanying source codes for the proposed algorithms and evaluation criterion are publicly available.