aster

Why we included this project

ASTER is a scene text recognizer built around a two-stage design that still shows up in production OCR: a spatial transformer first straightens curved or skewed text, then an attention-based recognizer reads the result. The repo ships pretrained models and a demo script, so you can run a sample image and see both the rectified output and the recognition result within a few minutes. Training and data-preparation scripts for standard benchmark datasets are included, which makes it a useful reference if you are assembling your own recognition stack. One caveat: the code targets TensorFlow r1.4, so expect to port it on a modern setup. If your goal is understanding the research, the paper is linked in the README, and a maintained PyTorch port exists for practical use.