DannyLuna/recaptcha-classification-57k

recaptcha-classification-57k

YOLO based image classification model for reCAPTCHA images. This model is used by vision-ai-recaptcha-solver to solve reCAPTCHA challenges.

Model summary

• Task: image classification
• Format: ONNX (recaptcha_classification_57k.onnx) and PyTorch (recaptcha_classification_57k.pt)
• Architecture: Ultralytics YOLO classification
• Labels: 14 classes (13 target classes + other)
• Trained on a 57k images dataset.

Labels

Target classes in this model:

• bicycle
• bridge
• bus
• car
• chimney
• crosswalk
• fire hydrant
• motorcycle
• mountain
• palm tree
• stairs
• tractor
• traffic light

The other class represents non-target/background tiles and is intentionally not treated as a match by the solver.

Intended use

• Classifying individual reCAPTCHA tile images as part of research or testing workflows.
• Integrated use with the vision-ai-recaptcha-solver package.

Training data

Trained on the dataset: DannyLuna/recaptcha-57k-images-dataset.

Training procedure

Trained with Ultralytics YOLO classification and exported to ONNX with dynamic input shapes.

How to use

Ultralytics (recommended)

from ultralytics import YOLO

model = YOLO("recaptcha_classification_57k.onnx", task="classify")
results = model("tile.jpg")

probs = results[0].probs
top1_idx = int(probs.top1)
print(model.names[top1_idx], float(probs.top1conf))

vision-ai-recaptcha-solver

from vision_ai_recaptcha_solver import RecaptchaSolver, SolverConfig

with RecaptchaSolver(SolverConfig()) as solver:
    result = solver.solve(
        website_key="your_site_key",
        website_url="https://example.com/your-page-with-recaptcha",
    )
    print(result.token)

Model files

• recaptcha_classification_57k.onnx
• recaptcha_classification_57k.pt

Limitations

• The reCAPTCHA system changes over time; accuracy may degrade on new variants.
• False positives are possible on visually similar objects.

Ethical considerations

Use this model only for lawful, authorized purposes such as research, benchmarking, or testing in controlled environments.

License

MIT License. See the repository LICENSE file for details.