XNews

docTR: A PyTorch-based OCR library for full-page text extraction from PDFs and images

docTR: A PyTorch-based OCR library for full-page text extraction from PDFs and images

docTR is a PyTorch-based text recognition library that extracts full-page text from PDFs or images with minimal code. It operates in two stages: first locating each word's position on the page, then recognizing characters within those regions, with both stage models being independently replaceable. When layout recognition is enabled, the system separately identifies titles, body text, tables, headers, and footers, preserving document structure. The library includes dedicated processing options for skewed and rotated pages common in scanned documents, with output in straight or tilted bounding boxes. A live demo is available on Hugging Face along with Colab notebooks. Originally developed by Mindee, the project is now maintained by the t2k team with recent code commits.

Source posts