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.
