docTR

docTR performs OCR in two stages, detecting text regions and then recognising them, on PDFs, images and even web pages. It copes with rotated pages, detects region types such as title, text, table and page header, and exports the result as structured JSON with the position of every word. There is a local demo app, a FastAPI service template and Docker images with GPU support. It suits engineers and developers who want to build a pipeline that extracts data from forms, test certificates or scanned data sheets. Limitation: it is a developer library with no graphical interface, and needs Python 3.11 or later and PyTorch.