Hierarchical Visual Feature Aggregation for OCR-Free Document Understanding

TL;DR

This paper introduces an OCR-free document understanding framework that uses multi-scale visual features and a hierarchical aggregation module to improve efficiency and accuracy in processing diverse document images with large language models.

Contribution

The paper proposes the Hierarchical Visual Feature Aggregation (HVFA) module and a new instruction tuning task, enhancing multi-scale visual feature integration and text reading capabilities in LLM-based document understanding.

Findings

HVFA reduces input tokens while maintaining information quality

The approach achieves superior performance on document understanding tasks

Effective handling of varying document image sizes

Abstract

We present a novel OCR-free document understanding framework based on pretrained Multimodal Large Language Models (MLLMs). Our approach employs multi-scale visual features to effectively handle various font sizes within document images. To address the increasing costs of considering the multi-scale visual inputs for MLLMs, we propose the Hierarchical Visual Feature Aggregation (HVFA) module, designed to reduce the number of input tokens to LLMs. Leveraging a feature pyramid with cross-attentive pooling, our approach effectively manages the trade-off between information loss and efficiency without being affected by varying document image sizes. Furthermore, we introduce a novel instruction tuning task, which facilitates the model's text-reading capability by learning to predict the relative positions of input text, eventually minimizing the risk of truncated text caused by the limited capacity of LLMs. Comprehensive experiments validate the effectiveness of our approach, demonstrating superior performance in various document understanding tasks.