capsolve-sp

Model artifacts for a compact five-character CAPTCHA recognizer.

Training and inference code is maintained separately on GitHub:

• https://github.com/nakasyou/capsolve-sp

Quick start

Install the CPU inference dependencies:

pip install huggingface-hub onnxruntime pillow numpy

Run the calibrated INT8 model on a 175x60 CAPTCHA image:

from huggingface_hub import hf_hub_download
from PIL import Image
import numpy as np
import onnxruntime as ort

REPO_ID = "nakasyou/capsolve-sp"
CHARSET = "0123456789abcdefghijklmnopqrstuvwxyz"

model_path = hf_hub_download(REPO_ID, "model.onnx")
session = ort.InferenceSession(
    model_path,
    providers=["CPUExecutionProvider"],
)

with Image.open("captcha.png") as source:
    image = source.convert("L")
    if image.size != (175, 60):
        raise ValueError("captcha image must be 175x60 pixels")
    pixels = np.asarray(image, dtype=np.float32)

# The model expects white background as 0 and dark ink as 1.
input_tensor = ((255.0 - pixels) / 255.0)[None, None, :, :]
logits = session.run(["logits"], {"image": input_tensor})[0]
text = "".join(CHARSET[index] for index in logits[0].argmax(axis=-1))
print(text)

To use the FP32 model instead, change model.onnx to model-fp32.onnx.

Files

• model.onnx: calibrated INT8 ONNX, recommended for fast CPU inference
• model-fp32.onnx: FP32 ONNX
• model.safetensors: PyTorch architecture weights
• config.json: architecture and preprocessing configuration
• metadata.json: training and evaluation metadata

Model

• Architecture: residual CNN with adaptive pooling into five character cells
• Parameters: 486,861
• Input: grayscale 175x60 image
• Preprocessing: invert intensity so white is 0 and ink is 1
• Character set: 0123456789abcdefghijklmnopqrstuvwxyz
• Output: exactly five characters

The model was fine-tuned on 13,231 frames from 11,754 CAPTCHA groups. Validation used 3,355 frames from 2,938 disjoint groups.

Validation metric	Result
Character accuracy	96.94%
Exact five-character accuracy	87.00%

Across all 14,332 valid labeled 0.png images:

Runtime model	Exact accuracy
FP32 ONNX	85.047% (12,189 / 14,332)
INT8 ONNX	85.033% (12,187 / 14,332)

FP32 and INT8 predictions agreed on 99.874% of images. On an AMD Ryzen AI 7 PRO 350, INT8 reached a median of approximately 1,573 images/second at batch size 8 and eight threads.

Limitations

• Supports only five-character images using the configured character set.
• Performance can be lower on CAPTCHA generators not represented in training.
• OCR-derived labels may contain residual annotation errors.
• Use only where automated image recognition is authorized.

License

The model artifacts are licensed under the MIT License. See LICENSE.