Printed Document Optical Character Recognition With Vision Encoder Decoder Architecture

via “printed-document optical character recognition with vision-encoder-decoder architecture”

image-to-text model by undefined. 7,67,977 downloads.

Unique: Uses a vision-encoder-decoder architecture (combining CNN visual encoding with transformer text decoding) specifically trained on printed documents, enabling character-level accuracy through attention-weighted spatial feature maps rather than traditional OCR heuristics. The base model achieves this through end-to-end differentiable learning of visual-to-textual mappings.

vs others: Outperforms traditional rule-based OCR engines (Tesseract) on printed documents with complex layouts and varied fonts due to learned visual representations, while being more lightweight and faster than large multimodal models (GPT-4V) for document-specific tasks.