AI Briefingindra-ocr: 49 pages per second on CPU: Ultra-lightweight OCR without PyTorch
datalab-to/indra-ocr
About the project
This is an ultra-lightweight OCR model with 1M parameters, running solely on ONNX Runtime without PyTorch. It processes 49 documents per second on CPU, achieving speeds 11 to 17 times faster than Tesseract or PaddleOCR.
It converts PDFs and images into Markdown, text, and JSON formats. By providing bounding boxes and labels for each line and layout block, it enables precise understanding of document structure. It supports a hybrid approach that prioritizes the text layer for digital PDFs and uses OCR for scans or images.
It supports OCR for 14 languages, including Latin, Cyrillic, Greek, Devanagari, Bengali, and Chinese. It has limitations with complex tables, handwriting, and low-quality scans, making it suitable for clean printed materials or bulk document processing.
datalab-to/indra-ocr
The original page has no description.
image-to-text
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.