An Analysis of Lightweight Models for Document Image Machine Translation

Bose A, Gorges T, Hüttner L, Schneider LS, Seuret M, Wu F, Christlein V (2026)


Publication Type: Conference contribution

Publication year: 2026

Journal

Publisher: Springer Science and Business Media Deutschland GmbH

Book Volume: 16973 LNCS

Pages Range: 143-160

Conference Proceedings Title: Lecture Notes in Computer Science

Event location: Vienna, AUT

ISBN: 9783032360328

DOI: 10.1007/978-3-032-36033-5_9

Abstract

Document Image Machine Translation (DIMT) translates text in document images into a target language while preserving reading order. This task is challenging due to complex layouts, multi-column structures, tables, and OCR noise. While large multi-modal models achieve strong results, their high computational cost limits practical deployment, especially in resource-constrained settings. We therefore explore whether lightweight alternatives can offer competitive performance more efficiently. We present a comparative study of lightweight modular pipelines and multi-modal approaches (one- and two-stage) for English-to-Chinese DIMT. Specifically, we evaluate: (1) lightweight pipelines combining compact document text extraction or reading-order generation models (TrOCR, Mistral OCR, and fine-tuned LayoutLMv3+T5-Small) with MarianMT; (2) large-model pipelines using few-shot prompting; and (3) multi-modal systems, including direct end-to-end generation (Pixtral) and DeepSeek-OCR-2-based variants, to examine trade-offs between simplicity, controllability, and performance. Experiments on the ICDAR 2025 DIMT benchmark show that fine-tuned lightweight pipelines achieve strong translation quality with far fewer parameters: LM3-T5-V2 → MarianMT reaches 0.3627 BLEU, outperforming Gemma-3 → MarianMT (0.3225). Pixtral achieves 0.2928 BLEU in the direct setting, remaining below the best modular pipeline, while DeepSeek-OCR-2 reaches 0.0275 BLEU directly but improves to 0.3023 when paired with MarianMT. Overall, our results highlight practical trade-offs between accuracy and computational cost, providing guidance for deploying DIMT systems under different resource constraints.

Authors with CRIS profile

How to cite

APA:

Bose, A., Gorges, T., Hüttner, L., Schneider, L.-S., Seuret, M., Wu, F., & Christlein, V. (2026). An Analysis of Lightweight Models for Document Image Machine Translation. In Gernot A. Fink, Alicia Fornés, Koichi Kise, Daniel Lopresti (Eds.), Lecture Notes in Computer Science (pp. 143-160). Vienna, AUT: Springer Science and Business Media Deutschland GmbH.

MLA:

Bose, Abantika, et al. "An Analysis of Lightweight Models for Document Image Machine Translation." Proceedings of the 20th International Conference on Document Analysis and Recognition, ICDAR 2026, Vienna, AUT Ed. Gernot A. Fink, Alicia Fornés, Koichi Kise, Daniel Lopresti, Springer Science and Business Media Deutschland GmbH, 2026. 143-160.

BibTeX: Download