|
|
--- |
|
|
language: en |
|
|
license: mit |
|
|
tags: |
|
|
- vision |
|
|
- document-ai |
|
|
- donut |
|
|
- ocr-free |
|
|
- image-to-text |
|
|
--- |
|
|
|
|
|
# document_parsing_donut_v1 |
|
|
|
|
|
## Overview |
|
|
This model is an implementation of the **Donut** (Document Understanding Transformer) architecture. Unlike traditional OCR-based systems, this model is OCR-free, meaning it maps raw document images directly to structured JSON outputs. It is fine-tuned to parse complex layouts such as invoices, receipts, and technical forms without a separate text recognition step. |
|
|
|
|
|
|
|
|
|
|
|
## Model Architecture |
|
|
The model utilizes a vision-encoder text-decoder framework: |
|
|
- **Encoder**: A Swin Transformer that processes high-resolution images into visual features. |
|
|
- **Decoder**: A BART-based multi-lingual transformer that generates text tokens in a sequence-to-sequence manner. |
|
|
- **Objective**: The model is trained using a cross-entropy loss to predict the next token based on both the visual input and preceding tokens: |
|
|
$$\mathcal{L} = -\sum_{t=1}^{T} \log P(y_t | y_{<t}, \mathbf{x})$$ |
|
|
|
|
|
## Intended Use |
|
|
- **Automated Data Entry**: Extracting key-value pairs from digitized business documents. |
|
|
- **Layout Analysis**: Identifying structural components (headers, tables, footers) in multi-page PDFs. |
|
|
- **Archival Digitization**: Converting historical scanned documents into searchable, structured data. |
|
|
|
|
|
## Limitations |
|
|
- **Resolution Sensitivity**: Performance drops significantly if images are scaled below 960x1280 pixels. |
|
|
- **Language Bias**: While capable, accuracy is highest for Latin-script documents; CJK and Arabic scripts require specialized fine-tuning. |
|
|
- **Handwriting**: The model is optimized for printed text and may struggle with highly cursive or disorganized handwriting. |
