TurboOCR

TurboOCR is a self-hosted OCR server engineered for speed and accuracy. It wraps the PP-OCRv5 weights from PaddleOCR in a native C++ runtime on NVIDIA TensorRT, exposing both HTTP and gRPC endpoints from a single binary. On a single RTX 5090 it sustains 270 img/s on FUNSD with 11 ms p50 latency and an F1 score of 90.2%, outperforming the Python PaddleOCR equivalent. It is designed for developers and teams who need high-throughput, low-latency OCR in production. The server accepts images and PDFs via POST requests, returning structured JSON with text, confidence scores, bounding polygons, and optional layout regions. PDF pages are rendered by a PDFium worker pool and fed into the same GPU pipeline as images. Layout detection using PP-DocLayoutV3 (25 region classes) is available opt-in per request via ?layout=1. The server is configured to run on any NVIDIA Turing-or-newer GPU on Linux. Deployment is one-line Docker: the TensorRT engines build on first start and cache in a named volume. TurboOCR ships as a Docker image at ghcr.io/aiptimizer/turboocr and is MIT licensed with source on GitHub. Prometheus metrics are exposed on /metrics, offering request counters, latency histograms, and VRAM usage. What makes TurboOCR different is its focus on raw performance via native C++ and TensorRT, while maintaining high accuracy (F1 = 90.2% on FUNSD). It is not a SaaS; it is a self-hosted binary that gives users full control and no per-request costs. It suits high-volume document processing pipelines where every millisecond counts.