arkhabbazan/ocr-digits-captcha

OCR Digits Captcha

A TrOCR model fine-tuned to recognize short numeric captchas of 4 or 5 digits such as 4829 and 50942.

Model details

• Architecture: Vision Encoder-Decoder / TrOCR
• Base model: microsoft/trocr-base-printed
• Task: Image-to-text OCR
• Target geometry: 145x56 and 120x48 RGB images
• Output: Digit strings of 4 or 5 digits

Training configuration

Parameter	Value
Epochs	6
Learning rate	2e-5
Batch size	8
Gradient accumulation	2
Weight decay	0.01
Generation beams	4
Maximum output length	16

Evaluation

The model was trained and evaluated on synthetic captchas generated at the target size and format. Real-world accuracy was not measured on a held-out real set, so the scores below do not prove it.

Split	Samples	Exact match	Character accuracy
Validation	1500	99.93%	99.98%
Synthetic test	750	100.00%	100.00%

Usage

from PIL import Image
from transformers import TrOCRProcessor, VisionEncoderDecoderModel

repo_id = "arkhabbazan/ocr-digits-captcha"
processor = TrOCRProcessor.from_pretrained(repo_id)
model = VisionEncoderDecoderModel.from_pretrained(repo_id).eval()

image = Image.open("captcha.png").convert("RGBA")
white = Image.new("RGBA", image.size, (255, 255, 255, 255))
white.alpha_composite(image)
pixel_values = processor(images=white.convert("RGB"), return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values, num_beams=4, max_length=16)
prediction = processor.batch_decode(
    generated_ids, skip_special_tokens=True
)[0]
print(prediction.replace(" ", ""))

Limitations

• Training data is primarily synthetic.
• Unseen fonts and layouts may reduce accuracy.
• The model is intended for short digit strings, not document OCR.
• Usage must be authorized and compliant with applicable policies.