doctr

Why we included this project

docTR is a document OCR library that folds text detection and recognition into one predictor object, so you can point it at a PDF or image and get structured text back in a few lines of code. The two stages stay separate underneath, which lets you pair a detection backbone that finds each word with a recognition backbone that transcribes it, mixing in pretrained models to suit your speed and accuracy needs. It runs under both PyTorch and TensorFlow, so teams already committed to either framework can use it as-is, and beyond PDFs and images it can handle rendered web pages. There is also a layout detection mode that labels regions such as titles, tables, and headers, which is worth having when you care where text sits, not just what it says.