Skip to content

normal.md

Latest commit

 

History

History
80 lines (62 loc) · 3.85 KB

normal.md

File metadata and controls

80 lines (62 loc) · 3.85 KB
Raw

Text-Based CAPTCHAs

Complexity: ⭐ (Low)

How it looks:

Normal Captcha

How to identify:

  • Distorted, wavy, or scratched text
  • Random letters and numbers
  • Lines or noise overlay
  • Sometimes colored background

Demo to try it yourself:

Free & open-source bypass methods

Tool Description Success Rate Link
ddddocr 🔥 Best for simple CAPTCHAs, works out of box 90-99% sml2h3/ddddocr
EasyOCR Multi-language OCR, good accuracy 70-90% JaidedAI/EasyOCR
PaddleOCR Powerful OCR, great for complex text 80-95% PaddlePaddle/PaddleOCR
Tesseract Classic OCR, needs preprocessing 50-80% tesseract-ocr/tesseract
captcha_trainer Train your own model 95-99% kerlomz/captcha_trainer
MMOCR OCR toolkit from OpenMMLab with many models 80-98% open-mmlab/mmocr
TrOCR Transformer-based OCR (Hugging Face) 85-98% microsoft/trocr-base
CRNN baselines Lightweight CNN+RNN text recognizers 80-95% meijieru/crnn.pytorch

Preprocessing tricks to boost OCR

  • Convert to grayscale → adaptive threshold (e.g., Otsu) to remove background noise.
  • Deskew/rotate to horizontal text.
  • Morphology (dilate/erode) to reconnect broken strokes; remove thin noise with opening.
  • Median/gaussian blur to reduce speckles; bilateral filter to keep edges.
  • Resize 2–4× with lanczos before OCR.
  • Invert colors if foreground/background is inverted.
  • Segment characters (vertical projection) and run per-character OCR when spacing is big.

Custom training options

  • Collect 2–5k synthetic samples using fonts + noise similar to target CAPTCHA.
  • Train with captcha_trainer, MMOCR, or a CRNN in PyTorch/TensorFlow.
  • Use CTC loss for variable-length strings; augment with elastic distortions, blur, salt & pepper, random lines.
  • Export to ONNX/TensorRT for fast inference.

Cloud / vision APIs (when allowed)

Service Notes
Google Cloud Vision Strong on warped text; pay-per-use
AWS Textract Good multi-language support
Azure Read Solid printed-text OCR
GPT-4o / Gemini Vision Vision-LMMs can solve many low-complexity text CAPTCHAs

💰 Paid alternatives:

If free methods don't work, use API services:

Service Price (per 1000) Speed Link
2Captcha $0.50 - $1.00 10-30s 2captcha.com
Anti-Captcha $0.50 - $1.00 10-30s anti-captcha.com
CapSolver $0.40 - $0.80 5-20s capsolver.com
CapMonster $0.40 - $0.90 10-25s capmonster.cloud
DeathByCaptcha $1.39 15-30s [deathbycaptcha.com](https://deathbycaptcha.com/api#api_details_upload_captcha

📊 Quick decision guide

  • Try ddddocr first (fast, high accuracy on simple text).
  • If noisy/warped: preprocess (binarize, deskew, denoise) then PaddleOCR/MMOCR.
  • Need maximum accuracy or custom font: train with captcha_trainer or CRNN on synthetic data.
  • Want hands-off: use paid solvers (2Captcha/Anti-Captcha/CapSolver).