What it does Trains a neural network to recognize text in captcha images without chopping them into individual characters first. Feed it 128×48 pixel images named like LABEL_*.jpg, run through tfrecord conversion, training, and evaluation scripts. Supports single GPU or multi-GPU training.

The interesting bit The 99.7% accuracy claim comes from a specific generator (lepture/captcha) with 50k samples and 20k steps. Switch generators to Gregwar/CaptchaBundle and accuracy drops to 52.1% even with 100k samples and 200k steps — a useful reminder that “captcha” covers a lot of visual ground.

Key highlights

• End-to-end recognition: no character segmentation preprocessing
• Python 2.7 + TensorFlow 1.1 + Anaconda2 4.3.1 (pinned, dated stack)
• Multi-GPU training script included
• Eval script produces accuracy numbers; recognize script prints predicted strings
• Works with custom datasets or default generated captchas

Caveats

• Python 2.7 and TensorFlow 1.1 are firmly in legacy territory
• Accuracy varies wildly by captcha generator; the 99.7% figure is not universal
• README lacks detail on model architecture or why segmentation-free works here

Verdict Worth a look if you’re researching captcha-breaking techniques or need a baseline end-to-end OCR pipeline to adapt. Skip it if you need modern dependencies or production-ready code — this is a 2017-era research snapshot.