A modern, retro-styled web application for recognizing CAPTCHA images using a hybrid CRNN (Convolutional Recurrent Neural Network) architecture with attention mechanisms. The application features CPU-based ONNX inference for fast, cross-platform deployment without GPU requirements.
The UltraCaptureV3 application features a retro-styled interface with a vintage aesthetic, displaying the project title, accuracy metrics, and creator information on the home page.
- High Accuracy: Achieves 95.08% Character Accuracy and 86.37% Sequence Accuracy on test sets.
- Advanced Architecture: Hybrid CRNN combining ResNet-style CNN backbone, CBAM attention, Bi-LSTM, and Transformer Encoder.
- CPU-Based Inference: Uses ONNX Runtime for fast, optimized CPU inference (30-100ms per image) without GPU requirements.
- Retro/Vintage Design: Beautiful single-page application with retro aesthetic, dark/light theme toggle, and responsive layout.
- Easy Deployment: Simple setup with PowerShell scripts for automated environment configuration.
- Web-Based Interface: Intuitive drag-and-drop image upload with real-time predictions and inference timing.
- Cross-Platform: Works on Windows, macOS, and Linux with Python 3.11+ and Node.js.
The Inference section provides an interactive drag-and-drop interface for uploading CAPTCHA images, displaying predictions, and showing inference timing information in real-time.
The model is a custom-built Convolutional Recurrent Neural Network (CRNN) that integrates multiple advanced concepts for maximum accuracy:
- Stage 1 - Feature Extraction: Deep CNN with ResNet-style residual blocks enhanced with CBAM (Convolutional Block Attention Module) for channel and spatial attention.
- Stage 2 - Sequence Modeling: Bidirectional LSTM processes feature sequences, refined by Transformer Encoder for long-range dependencies.
- Loss Function: Connectionist Temporal Classification (CTC) for unsegmented sequence learning.
Key Specifications:
- Input Size: 64×256 pixels (RGB).
- Hidden Size: 512.
- Attention Heads: 8.
- Transformer Layers: 4.
- Charset: 62 characters (0-9, A-Z, a-z).
- Model Size: ~273MB (ONNX format).
The architecture section displays the CRNN model components in a retro-styled card layout, showing the convolutional backbone, CBAM attention mechanism, Bi-LSTM, and Transformer Encoder with their respective descriptions.
The model was trained on the Huge CAPTCHA Dataset that's been made by the author, Mridankan Mandal, which has been made available on Kaggle. This dataset contains a large number of CAPTCHA images with alphanumeric characters, where the label for each image is its filename.
The model achieves high accuracy on both the validation and unseen test sets.
| Metric | Validation Set | Test Set |
|---|---|---|
| Sequence Accuracy | 86.4% | 86.37% |
| Character Accuracy | 95.1% | 95.08% |
The performance metrics table displays the model's accuracy on validation and test sets, showing both sequence accuracy and character accuracy metrics.
- Python 3.11 (recommended) or 3.8+.
- Node.js 18.0 or higher.
- Git (optional, for cloning the repository).
The easiest way to set up the application is using the provided PowerShell script:
.\setup.ps1This script will automatically:
- Check for Python and Node.js installation.
- Create a Python virtual environment.
- Install all backend dependencies (ONNX Runtime, Flask, etc.).
- Install all frontend dependencies.
- Verify the ONNX model exists.
Start both the backend and frontend servers with:
.\start.ps1Then open your browser and navigate to:
http://localhost:5173
For detailed setup and usage instructions, see:
- Installation and Setup Guide - Complete setup instructions.
- Usage Guide - How to use the web application.
- API Documentation - Backend API reference.
- Codebase Index - Project structure and file descriptions.
ProjectAAWebsite/
├── frontend/ # React + Vite web application
│ ├── src/
│ │ ├── App.tsx # Main application component
│ │ ├── App.css # Application styles
│ │ ├── index.css # Global styles
│ │ └── main.tsx # Entry point
│ ├── public/ # Static assets
│ ├── package.json # Frontend dependencies
│ ├── vite.config.ts # Vite configuration
│ └── tailwind.config.js # Tailwind CSS configuration
│
├── backend/ # Flask API with ONNX Runtime
│ ├── app.py # Flask application
│ ├── models/
│ │ └── best_model.onnx # ONNX model for inference
│ ├── config.json # Model configuration
│ ├── requirements.txt # Backend dependencies
│ └── export_to_onnx.py # ONNX export script (reference)
│
├── InstallationAndSetup.md # Installation guide
├── Usage.md # User guide
├── API.md # API documentation
├── CodeBaseIndex.md # Codebase structure
├── setup.ps1 # Automated setup script
├── start.ps1 # Application startup script
└── README.md # This file
- GitHub Repository: CRNN_Captcha_Recognition.
- Dataset: Huge CAPTCHA Dataset on Kaggle.
- Creator: Mridankan Mandal (RedZapdos123, WhiteMetagross).
Frontend:
- React 19 with TypeScript.
- Vite 7 (build tool).
- Tailwind CSS 3 (styling).
- Axios (HTTP client).
Backend:
- Flask 3.0.0 (web framework).
- ONNX Runtime 1.17.1 (CPU-based inference).
- Pillow 10.1.0 (image processing).
- NumPy 1.24.3 (numerical operations).
This project is created by Mridankan Mandal as part of Project AA.
The model was trained on the Huge CAPTCHA Dataset available on Kaggle.
© 2025 UltraCaptureV3. All rights reserved.
For setup issues, see the Installation and Setup Guide.
For usage questions, see the Usage Guide.
For API-related questions, see the API Documentation.
For technical details, see the Codebase Index.