Pick your own eyes and brain for document OCR

What it does

docTR is a deep-learning OCR library that pulls text out of PDFs, images, and even web pages. It runs a two-stage pipeline: a detection model locates words, then a recognition model reads them. The output is a structured document tree—pages, blocks, lines, words—that you can visualize or export as JSON.

The interesting bit

Instead of a monolithic black box, docTR lets you mix and match detection architectures (DBNet, FAST, LinkNet) with recognition architectures (CRNN, PARSeq, ViTSTR). It also offers a KIE predictor that detects semantic classes like dates or addresses rather than just dumping raw text, plus a synthesis mode that rebuilds the original page from predictions to sanity-check results.

Key highlights

• Modular pipeline: pair detection and recognition models independently.
• Handles rotated text with options for straight or angled bounding boxes.
• Structured output hierarchy (Page, Block, Line, Word) exportable to nested dictionaries.
• KIE predictor for class-specific extraction beyond plain OCR.
• Ships with pretrained weights, GPU-ready Docker images, and a live Hugging Face demo.

Caveats

• Web page ingestion requires an extra dependency (weasyprint).
• Docker GPU images need a host CUDA version of at least 12.2.

Verdict

Reach for docTR if you want fine-grained control over OCR architecture choices or structured document parsing. Skip it if you just need a quick, opaque API call and do not care which model is doing the reading.