What it does

This repo implements the classic CRNN architecture from Shi et al.’s 2015 paper: convolutional features fed into a Bi-LSTM, trained end-to-end with CTC loss. It reads text in the wild—street signs, scanned documents, whatever you throw at it—and spits out character sequences without needing per-character alignment in the training data.

The interesting bit

The author didn’t just port the paper; they trained and released two full sets of weights: one on Synth90k (English synthetic text) and a separate Chinese model on an unnamed 2.7-million-image dataset with 5,824 character classes. There’s even a small PDF recognition demo. Plus TensorFlow Serving export scripts, which you don’t see in every research repo from 2017.

Key highlights

• Pretrained English model: 97.4% per-character precision, 93.3% full-sequence accuracy on 891k Synth90k test images
• Pretrained Chinese model with 5,824 classes, sequence length 70, image size 280×32
• Multi-GPU training support via --multi_gpus 1
• TensorFlow Serving integration with both REST and gRPC client examples
• Training resumption from snapshots; optional decoded output metrics (slower but informative)

Caveats

• Locked to TensorFlow 1.15 and Python 3.5; the README explicitly warns about incompatibility with newer versions
• The Chinese dataset source is unattributed—the author admits they don’t know who owns it
• No online demo yet; TODO has been unchecked for years

Verdict

Worth a look if you need a battle-tested CRNN baseline or Chinese OCR weights and can tolerate legacy TensorFlow. Skip it if you want modern TF2/PyTorch or a maintained, production-ready pipeline.