TurboOCR
A GPU document parser that combines OCR, layout analysis, table and formula recognition into one CUDA and TensorRT pipeline, converting pages to Markdown at up to hundreds of images per second.
TurboOCR is described as a document parser rather than a plain OCR tool: it runs text detection and recognition together with layout analysis, table extraction to HTML and formula extraction to LaTeX, all on a single multi-stream CUDA and TensorRT engine, running locally without a vision-language model and serving results over HTTP and gRPC. Whole-page OCR runs at up to several hundred images per second on receipts on a single high-end GPU, and full structured parsing with layout, tables and formulas runs markedly faster than comparable vision-language-model-based document parsers in the project's own comparison.
Its detection and recognition model covers Latin, Chinese and Japanese script in one model, with additional Arabic, Cyrillic, Korean, Thai and Greek script support through retained older recognizer models, and it can export word-level bounding boxes alongside recognized text. PDFs are rendered and processed in parallel with optional page-image export and automatic rotation, and the service exposes Prometheus metrics.
It deploys through a one-line Docker container that builds its TensorRT engines on first start, requires Linux with a Turing-generation or newer NVIDIA GPU, and is written in C++ under the MIT license; Apple Metal and Intel OpenVINO backends are in testing and an AMD ROCm backend is in development.