This project provides an extremely fast, GPU-accelerated document parsing server, addressing the need for high-throughput local processing of complex documents. It targets engineers, researchers, and power users requiring efficient OCR, layout analysis, table extraction, and formula recognition, offering significant speed advantages over traditional OCR engines and VLM-based parsers.

How It Works

TurboOCR employs a unified C++/CUDA/TensorRT engine built upon PP-OCRv6 for text detection and recognition, PP-DocLayoutV3 for layout analysis, SLANet+ for table extraction to HTML, and PP-FormulaNet-S for formula extraction to LaTeX. This multi-stage pipeline is optimized for single-GPU execution using TensorRT FP16, enabling local, low-latency document parsing without external APIs or large vision-language models, achieving remarkable throughput.

Quick Start & Requirements

• Primary install: Docker (ghcr.io/aiptimizer/turboocr:latest).
• Prerequisites: Linux, NVIDIA driver 595+, Turing+ GPU (RTX 20-series / GTX 16-series+), ~4 GB VRAM for text-only, ~8 GB for full pipeline.
• Estimated setup time: ~90 seconds for initial TensorRT engine build on RTX 5090 (up to an hour on older GPUs); subsequent starts are instant via cache.
• Links: Docs, Benchmarks, API.

Highlighted Details

• Performance: Achieves up to 559 images/s on receipts and 520 images/s on forms with a single RTX 5090.
• Accuracy: Competitive with leading engines on forms and receipts (e.g., FUNSD 92%, CORD 93% word-F1).
• Multi-script OCR: PP-OCRv6 supports Latin, Chinese, Japanese; PP-OCRv5 models extend support to Arabic, Cyrillic, Korean, Thai, Greek.
• Advanced Features: Native PDF parsing, layout analysis, table extraction to HTML, and formula extraction to LaTeX are available as opt-in stages.
• Deployment: Single-line Docker deploy with auto-built TensorRT engines and Prometheus metrics.

Maintenance & Community

Sponsorships from Miruiq and DiaIQ are noted. No explicit community links (Discord, Slack, etc.) are provided in the README.

Licensing & Compatibility

• License: MIT.
• Compatibility: Permissive for commercial use.

Limitations & Caveats

Requires a compatible NVIDIA GPU (Turing architecture or newer) and sufficient VRAM (minimum 4GB, recommended 8GB for full pipeline). Initial TensorRT engine compilation can be time-consuming on older hardware. Input size caps are enforced for images and PDF pages.