VisFocus: Prompt-Guided Vision Encoders for OCR-Free Dense Document Understanding

TL;DR

VisFocus introduces a prompt-guided vision encoder for OCR-free dense document understanding, enabling better focus on relevant document parts and achieving state-of-the-art results.

Contribution

The paper proposes a novel prompt-guided vision encoder architecture with a new pre-training task, enhancing OCR-free document understanding capabilities.

Findings

Significant performance improvements on multiple benchmarks.

Effective highlighting of relevant document regions.

State-of-the-art results achieved.

Abstract

In recent years, notable advancements have been made in the domain of visual document understanding, with the prevailing architecture comprising a cascade of vision and language models. The text component can either be extracted explicitly with the use of external OCR models in OCR-based approaches, or alternatively, the vision model can be endowed with reading capabilities in OCR-free approaches. Typically, the queries to the model are input exclusively to the language component, necessitating the visual features to encompass the entire document. In this paper, we present VisFocus, an OCR-free method designed to better exploit the vision encoder's capacity by coupling it directly with the language prompt. To do so, we replace the down-sampling layers with layers that receive the input prompt and allow highlighting relevant parts of the document, while disregarding others. We pair the architecture enhancements with a novel pre-training task, using language masking on a snippet of the document text fed to the visual encoder in place of the prompt, to empower the model with focusing capabilities. Consequently, VisFocus learns to allocate its attention to text patches pertinent to the provided prompt. Our experiments demonstrate that this prompt-guided visual encoding approach significantly improves performance, achieving state-of-the-art results on various benchmarks.