ODTr: Transformer Integrating OCR Auxiliary Map and Image Depth Information for Document Image Unwarping

Published: 01 Jan 2024, Last Modified: 12 Apr 2025ICPR (31) 2024EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: In this study, we introduce a novel application of transformer for document image unwarping, leveraging depth information and Optical Character Recognition (OCR) results. Our proposed model integrates two key modules: a depth module and an OCR module, into the transformer framework tailored for document image processing. The depth module predicts the relative depth of each pixel in the document image, thereby providing crucial spatial context for unwarping. Concurrently, the OCR module identifies regions suitable for OCR, acting as preferences for the unwarping process. This hybrid approach aims to mitigate text distortion inherent in document image unwarping, consequently enhancing OCR accuracy, although it may reduce our model’s capability in structural image unwarping. Experimental results showcase our model’s effectiveness, achieving a Character Error Rate (CER) of 24.81%, marking a significant 6.2% absolute enhancement compared to the baseline method, DocTr.
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview