Luogu CAPTCHA Recognition Model (Validated Best Edition)
Lightweight CNN model trained on langningchen/luogu-captcha-dataset to recognize 4-character Luogu CAPTCHAs.
Model Performance
- Validation Captcha Accuracy (Whole Sequence): 94.26% (FP16/FP32) / 94.34% (INT8 Quantized)
- Character Level Accuracy: ~98.5%
- Input Dimensions: 35x90x3 (Normalized to [0, 1])
- Alphabet:
abcdefghijklmnopqrstuvwxyz123456789(Length: 35)
Available Formats & File Sizes
| Format | File Path | Size | Description |
|---|---|---|---|
| ONNX (INT8) | onnx/luogu-captcha-int8.onnx |
1.70 MB | Recommended for CPU inference: static quantized, ~75% smaller with no accuracy loss (94.34%). |
| ONNX (FP16) | onnx/luogu-captcha-fp16.onnx |
3.15 MB | Recommended for GPU/DirectML/WebGPU: native half-precision (94.26%). |
| ONNX (FP32) | onnx/luogu-captcha-fp32.onnx |
6.22 MB | Clean standard single-precision graph for maximum platform compatibility. |
| TF.js (uint8) | tfjs/ |
~720 KB | Quantized layers format for zero-dependency in-browser recognition. |
| Keras H5 | luogu-captcha-mobile.h5 |
~6.5 MB | Standard Keras model file. |
- Downloads last month
- 55
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support