EAST

Why we included this project

EAST is the reference implementation of the Efficient and Accurate Scene Text Detector from the 2017 paper, and it remains a common first stop for OCR work. Instead of assuming text sits on clean horizontal lines, it finds text regions at arbitrary angles and returns rotated bounding boxes, which is exactly what you want for photos, scanned documents, and street scenes. The repo includes a pre-trained model that scores around 80.8 F1 on the ICDAR 2015 incidental text challenge, plus a C++ locality-aware NMS that keeps detection fast on modest hardware. Training and demo scripts make it straightforward to adapt the detector to your own data, whether you are digitizing documents, reading license plates, or pulling text out of photos. One caveat: it is built on TensorFlow 1.x, so expect some porting work if you are on a modern stack.