OCR for equations that finally learned some Chinese

What it does

LaTeX_OCR_PRO takes images of mathematical formulas—handwritten, printed, or mixed with Chinese—and transcribes them into LaTeX source using an encoder-decoder pipeline. It builds on the classic Harvard image-to-LaTeX work, swapping in attention and beam search to handle more diverse input. The project also bundles dataset generation, training scripts, and attention visualizations so you can watch the model decide where to look.

The interesting bit

The real twist is the claimed support for Chinese-mixed and handwritten formulas, territory the original Harvard dataset largely ignored. The author also structures the output around a LaTeX abstract syntax tree, though the README is vague on exactly how that “simple symbol reasoning” works.

Key highlights

• Seq2seq with attention and beam search, evaluated on a 70,000+ formula dataset
• EditDistance and BLEU-4 scores in the low-to-mid 90s on the test set
• Attention visualization via TensorBoard and Jupyter notebooks
• Dataset hosted on HuggingFace for faster bootstrapping
• Ships with a bare-bones Django wrapper for browser-based demos

Caveats

• Locked to Python 3.5 and TensorFlow 1.12.2, a stack that is well past end-of-life
• Exact match accuracy sits below 50%; the author notes it could climb higher, but only with significantly more training time
• “Simple symbol reasoning” and AST-based data structures are advertised but not explained in the README

Verdict

Worth a look if you need a starting point for non-English formula OCR or want to study attention mechanisms on a concrete seq2seq model. Skip it if you need production-grade recognition or a modern dependency stack.