Teaching OCR to un-warp text before reading it

What it does

ASTER is a scene text recognizer that reads cropped text from natural images even when the text is warped or curved. It rectifies distorted text into a canonical form before running recognition, using an attention-based architecture described in an IEEE TPAMI paper. The repository provides pretrained weights, training scripts, and evaluation tools for standard benchmarks.

The interesting bit

The model does not assume your text is already deskewed; it learns to spatially transform and straighten text as a first-class step in the pipeline. That makes it a single system that both irons out geometry and reads the characters.

Key highlights

• Built on TensorFlow 1.4 and reuses components from the TensorFlow Object Detection API
• Ships with pretrained models and data preparation helpers for popular scene-text datasets
• A third-party PyTorch port is available and noted in the README
• MIT-licensed, but the authors explicitly flag potential patent concerns for commercial use and prefer academic research usage

Caveats

• Tightly coupled to TensorFlow 1.4; the README warns higher versions may not work
• Requires building custom C++ operators and protobuf definitions as part of setup
• The authors caution against commercial deployment without contacting them first due to possible patent issues

Verdict

A solid reference implementation if you’re reproducing or extending the ASTER paper. Otherwise, the stale TensorFlow version and academic licensing caveats make it a poor fit for modern production OCR pipelines.