zaltahir/kaptchaocr-det

KaptchaOCR synthetic glyph-detection dataset

Synthetic CAPTCHA images with per-character bounding boxes, used to fine-tune the glyph detectors in KaptchaOCR.

This dataset contains no real CAPTCHA images and no personal data. Every image and box is generated by the project's gen_det generator. The generator knows each glyph's ink box and tracks it through wave distortion and global rotation, so the labels are exact and no manual annotation is involved.

Contents

m2_dataset.zip

The full training kit. Unzip it and use the splits directly with PaddleOCR det training or the KaptchaOCR eval harness.

Split	Layout	Language	Images	Seed	Purpose
train	scatter	English + Chinese	20,000	11	fine-tune
val	scatter	English + Chinese	200	1	comparison set
val2	scatter	English + Chinese	200	21	second reference
val_line	line	Latin	200	2	optional line evaluation
val_clean	line	Latin	50	5	pipeline sanity check

Each split contains:

<split>/
├── images/000000.png ...
├── Label.txt   # PaddleOCR det format
└── gt.json     # glyph boxes used by eval_det

Label.txt format:

images/000000.png<TAB>[{"transcription":"X","points":[[x,y],[x,y],[x,y],[x,y]]}, ...]

gt.json format:

{"000000.png": [{"ch": "X", "quad": [[x, y], [x, y], [x, y], [x, y]]}]}

unseen/test/

The 400-image, 1,800-box benchmark split used in the model card. It was generated after model selection with a frozen seed (20261103) and was never used for training or checkpoint selection. It contains images/, Label.txt and gt.json.

Image format

• 320×160 PNG canvases
• scattered glyphs, English + Chinese characters
• random wave distortion, rotation, color, noise and background

License

Apache-2.0. The images and boxes are generated by the KaptchaOCR project. No third-party CAPTCHA images are redistributed here. Fonts are used only to render the synthetic images; no font files are included.

Links

• GitHub repository
• Models on Hugging Face
• Benchmarks