DocFusion Open-source Document Parsing Framework - Solve Complex System Problems and Provide an Efficient and Unified Solution

Docfusion

Developed by sc22mc

DocFusion is a unified framework for document parsing tasks, aiming to solve the problems of system complexity and limited performance in existing document parsing methods and providing a more efficient and unified solution for document parsing.

Text Recognition

PyTorch

Open Source License:Apache-2.0 #Unified framework for document parsing #Lightweight model #Layout element detection

Downloads 107

Release Time : 1/28/2025

Model Overview

DocFusion is a unified document parsing framework that can simultaneously handle layout element detection and recognition tasks in document parsing, featuring lightweight and high performance.

Model Features

Unified processing ability

Proposed Gaussian-Kernel CrossEntropy Loss (GK-CEL) to enable the generative framework to simultaneously handle layout element detection and recognition tasks in document parsing.

Lightweight model

As a unified document parsing model, DocFusion only has 0.28B parameters.

High-quality dataset

Built the DocLatex-1.6M dataset to provide high-quality data support for model training.

Excellent performance

Performed excellently in four core document parsing tasks, verifying the effectiveness of the unified method.

Model Capabilities

Document layout element detection

Document element recognition

Unified document parsing

Use Cases

Document processing

Document layout analysis

Automatically detect layout elements in the document, such as titles, paragraphs, tables, etc.

Performed excellently in four core document parsing tasks

Document element recognition

Recognize the specific element content in the document, such as text, images, formulas, etc.

High-quality data support improves the recognition accuracy

🚀 DocFusion: A Unified Framework for Document Parsing Tasks

DocFusion is a unified framework for document parsing tasks. It addresses the limitations of existing methods by introducing the Gaussian - Kernel CrossEntropy Loss (GK - CEL), enabling simultaneous handling of layout element detection and recognition. With only 0.28B parameters, it shows competitive performance across four core document parsing tasks.

🚀 Quick Start

The text doesn't provide specific quick - start steps, so this section is skipped.

✨ Features

Unified Approach: DocFusion can handle both layout element detection and recognition tasks simultaneously, reducing system complexity and maintenance overhead.
Gaussian - Kernel CrossEntropy Loss (GK - CEL): This loss function enables generative frameworks to manage the differences in data representations between detection and recognition tasks.
Small Parameter Size: With only 0.28B parameters, it is efficient in resource utilization.
High - Quality Dataset: The DocLatex - 1.6M dataset provides high - quality training support.

📚 Documentation

Problem Background

Document parsing involves layout element detection and recognition, which are crucial for information extraction. Existing methods often use multiple models for these tasks, increasing system complexity and maintenance overhead. Some models attempting to unify detection and recognition fail to address the differences in data representations, limiting performance in document processing.

Key Innovations

Our research found that recognition relies on discrete tokens while detection relies on continuous coordinates, causing challenges in gradient updates and optimization. To solve this, we proposed the Gaussian - Kernel CrossEntropy Loss (GK - CEL), allowing generative frameworks to handle both tasks at the same time.

Model Introduction

Based on GK - CEL, we proposed DocFusion, a unified document parsing model with only 0.28B parameters.

Dataset

We constructed the DocLatex - 1.6M dataset to offer high - quality training support.

Experimental Results

Experimental results show that DocFusion, equipped with GK - CEL, performs competitively across four core document parsing tasks, validating the effectiveness of our unified approach.

Resources

📄 License

The project is licensed under the Apache - 2.0 license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご