🔍 CapSnap

A pure-Python CAPTCHA solver & OCR engine — built from scratch, zero dependencies

No Tesseract. No OpenCV. No NumPy. No Pillow. Just Python.

---

🤔 What is CapSnap?

CapSnap is a production-quality CAPTCHA solver and OCR engine written entirely in the Python standard library with zero external dependencies. It implements every component from scratch:

• 📦 Custom PNG decoder
• 🖼️ Grayscale conversion & Otsu's binarization
• 🔗 Connected-component labeling for character isolation
• 🧠 Multi-Layer Perceptron (MLP) neural network — trained on real CAPTCHA images
• ✂️ 5-way forced vertical slicing for noisy CAPTCHAs

It ships with a pre-trained model specifically tuned for noisy, handwritten-style CAPTCHA characters used by real-world websites — and it can be retrained on any new CAPTCHA style in minutes.

---

✨ Key Features

Feature	Details
🚫 Zero dependencies	Pure Python standard library only — nothing to pip install
🔓 CAPTCHA solver	Built-in mode for 4–5 character noisy CAPTCHAs
🧠 Neural network OCR	Custom MLP with ReLU hidden layer and softmax output
📸 Universal input	File path, pathlib.Path, raw bytes, or base64 data-URL
🏋️ Self-trainable	Retrain on any CAPTCHA dataset — answers extracted from API tokens automatically
🖥️ CLI included	capsnap captcha.png works out of the box
⚡ Lightweight	< 2 MB installed, starts instantly, no warmup
🔒 Privacy-first	Runs 100% locally — no API calls, no cloud, no data leaks

---

📦 Installation

pip install capsnap

No extra dependencies. No system libraries. Works on Python 3.11+.

---

🚀 Quick Start — CAPTCHA Solving

from capsnap import OCR

# Initialize in CAPTCHA mode
ocr = OCR(mode="captcha")

# Solve from a file path
result = ocr.read("captcha.png")
print(result.text)        # e.g. "RF3rH"
print(result.confidence)  # e.g. 0.923

Feed it anything

# From a URL / API response (base64 data-URL)
result = ocr.read("data:image/png;base64,iVBORw0KGgo...")

# From raw bytes (e.g. requests response)
import urllib.request
with urllib.request.urlopen("https://example.com/captcha") as r:
    result = ocr.read(r.read())

# From a pathlib.Path
from pathlib import Path
result = ocr.read(Path("captcha.png"))

Full example — solve a live CAPTCHA from an API

import json
import urllib.request
import base64
from capsnap import OCR

url = "https://example.com/api/captcha"
req = urllib.request.Request(url, headers={"accept": "application/json"})
with urllib.request.urlopen(req) as response:
    data = json.loads(response.read().decode())

# Pass the base64 image directly to CapSnap
ocr = OCR(mode="captcha")
result = ocr.read(data["image"])  # data:image/png;base64,...
print(result.text)

---

🖥️ CLI Usage

# Solve a CAPTCHA image
capsnap --mode captcha captcha.png

# Standard document OCR
capsnap document.png

---

🔄 All Input Methods

from capsnap import OCR
ocr = OCR(mode="captcha")

ocr.read("captcha.png")                    # file path string
ocr.read(Path("captcha.png"))              # pathlib.Path
ocr.read(open("captcha.png", "rb").read()) # raw bytes
ocr.read("data:image/png;base64,...")      # base64 data-URL

# Explicit methods still available
ocr.read_path("captcha.png")
ocr.read_bytes(raw_bytes)
ocr.read_base64(b64_string)

---

🏗️ Architecture

Input (path / bytes / base64)
          │
          ▼
   ┌─────────────┐
   │ PNG Decoder │ ← pure Python, no Pillow
   └──────┬──────┘
          │
          ▼
   ┌─────────────┐
   │  Grayscale  │ ← luminance-weighted conversion
   └──────┬──────┘
          │
          ▼
   ┌─────────────┐
   │    Otsu     │ ← auto threshold binarization
   └──────┬──────┘
          │
     ┌────┴─────────────────────┐
     │ CAPTCHA mode             │ Document mode
     │ 5-way vertical slicing   │ Connected-component labeling
     └────────────┬─────────────┘
                  │
                  ▼
         ┌──────────────┐
         │ Feature Ext. │ ← normalized 20×20 patch
         └──────┬───────┘
                │
                ▼
         ┌──────────────┐
         │  MLP Network │ ← ReLU hidden → Softmax output
         └──────┬───────┘
                │
                ▼
          OCRResult
       (text, confidence)

---

🎯 Retraining on Any CAPTCHA

CapSnap can retrain its neural network on any CAPTCHA style automatically if the API exposes the answer in the response token. No manual labeling needed:

# Fetch 1000 real CAPTCHAs and retrain the model
PYTHONPATH=. python tools/train_from_tokens.py

The training script:

1. Fetches live CAPTCHAs from the target API
2. Decodes the correct answer from the API's JWT/token
3. Extracts and normalizes character patches
4. Trains the MLP for 100 epochs
5. Saves the updated capsnap/model.capsnap automatically

---

📁 Project Structure

capsnap/
├── capsnap/               ← installable package
│   ├── api.py             ← OCR class (universal entry point)
│   ├── model.py           ← MLP neural network (pure Python)
│   ├── model.capsnap      ← pre-trained weights (bundled)
│   ├── recognize.py       ← character recognition
│   ├── decoder/           ← pure-Python PNG decoder
│   ├── components.py      ← connected-component labeling
│   ├── features.py        ← patch extraction + resize
│   ├── grayscale.py       ← RGB → grayscale
│   ├── threshold.py       ← Otsu's binarization
│   ├── morphology.py      ← noise removal
│   ├── segmentation.py    ← line/word grouping
│   └── cli.py             ← CLI entry point
├── tests/                 ← offline unit tests
├── examples/              ← usage examples
├── assets/                ← images & donation QR codes
└── .github/workflows/     ← CI + auto-publish to PyPI

---

🧪 Testing

pip install -e ".[dev]"
pytest tests/ -v

All 8 tests run offline — no network, no API calls required.

---

💖 Support & Donations

CapSnap is free and open-source. If it saved you time, helped bypass a pesky CAPTCHA, or powered your project, consider buying me a coffee! ☕

🇧🇩 bKash / Nagad / Rocket (Bangladesh)

---

💎 Telegram Stars / TON (Gram)

---

🪙 Litecoin (LTC)

---

💵 USDT (BEP20)

---

Every contribution — big or small — keeps this project alive. Thank you! 🙏

---

📄 License

MIT License — see LICENSE for details.

---

Made with ❤️ by @anbuinfosec · Pure Python · Zero Dependencies