aiptimizer/TurboOCR

What it solves

TurboOCR is a high-performance GPU-accelerated document parser designed to convert images and PDFs into structured Markdown. It solves the bottleneck of slow OCR (Optical Character Recognition) and document parsing, offering speeds significantly faster than traditional OCR engines or Vision Language Models (VLMs) while maintaining high accuracy for forms, receipts, and dense documents.

How it works

The project implements a multi-stage pipeline using C++, CUDA, and TensorRT for maximum efficiency. It leverages the PP-OCRv6 model family for text detection and recognition, and optionally integrates PP-DocLayoutV3 for layout analysis, SLANet-Plus for table-to-HTML conversion, and PP-FormulaNet-S for formula-to-LaTeX conversion. The entire process runs locally on NVIDIA GPUs, providing an API (HTTP and gRPC) for processing requests.

Who it’s for

It is intended for developers and organizations needing to process large volumes of documents at scale, specifically those requiring fast, structured output (Markdown, HTML, LaTeX) from scanned images or PDFs without relying on on-device VLMs or external APIs.

Highlights

• Extreme Throughput: Capable of processing up to 559 images per second on an RTX 5090 for receipts.
• Comprehensive Parsing: Handles text, layout, tables (converted to HTML), and mathematical formulas (converted to LaTeX).
• Multi-Language Support: Native support for Latin, Chinese, and Japanese, with additional support for Arabic, Cyrillic, Korean, Thai, and Greek.
• Optimized Stack: Built with C++20 and TensorRT, featuring a one-line Docker deployment and Prometheus metrics for monitoring.