CIFAR-10 Image Classification with CNN Ensemble 🖼️🤖

This project implements an advanced image classification system using Convolutional Neural Networks (CNNs) on the CIFAR-10 dataset. By combining three distinct CNN architectures through ensemble methods, we achieve robust and accurate image classification results. 🎯

Overview 📋

The project consists of:

• 🧠 Three CNN models with different architectures
• 🔄 Advanced data augmentation techniques
• 🤝 Ensemble methods (Simple Averaging and Majority Voting)
• 🌐 A Streamlit web interface for real-time predictions

Project Components 📦

├── CNN_model_1.h5       # Basic CNN architecture
├── CNN_model_2.h5       # Enhanced CNN with batch normalization
├── CNN_model_3.h5       # ResNet-inspired architecture
└── ensemble_classifier.py # Streamlit interface

Model Architectures 🏗️

1. CNN Model 1 (Base Model) 📊

• Foundational CNN architecture:
  • 3 convolutional layers with increasing filters
  • MaxPooling after each conv layer
  • Gradually increasing Dropout (0.25 → 0.5) for regularization
  • Dense layers: 512 → 256 → 10 (output)

2. CNN Model 2 (Enhanced) 🚀

• Sophisticated architecture:
  • 4 convolutional blocks
  • Additional filters: 256 in the last convolution blocks
  • Additional neurons for a more complicated fully-connected layer : 1024 → 512 → 10 (output)

3. CNN Model 3 (ResNet-Inspired) ⭐

• State-of-the-art features:
  • Residual connections (skip connections)
  • Identity mappings
  • Deep supervision
  • Bottleneck layers

Training Pipeline 🛠️

Data Preparation 📥

• Dataset: CIFAR-10 (60,000 32x32 color images)
  • 40,000 training images
  • 10,000 validation images
  • 10,000 testing images
• Classes: 10 (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck)
• Preprocessing:
  • Normalization: [0,1] range

Data Augmentation 🔄

augmentation = ImageDataGenerator(
    rotation_range=15,
    width_shift_range=0.1,
    height_shift_range=0.1,
    horizontal_flip=True,
    zoom_range=0.1,
    shear_range=0.1
)

Training Configuration ⚙️

• Batch size: 32
• Epochs: 100 (with early stopping)
• Optimizer: Adam (lr=0.001)
• Loss: Categorical Crossentropy
• Metrics: Accuracy, Top-5 Accuracy

Running the Interface 🖥️

Local Deployment 🏠

streamlit run ensemble_classifier.py

Google Colab Deployment ☁️

!pip install streamlit
!streamlit run ensemble_classifier.py & npx localtunnel --port 8501

Usage Guide 📱

1. 📥 Download the pre-trained models
2. 🚀 Launch Streamlit interface
3. 🖼️ Upload a 32x32 RGB image
4. 🔍 Click "Make Prediction"
5. 📊 View detailed classification results

Performance Metrics 📈

Model Performance Comparison:

┌────────────────────────┬────────────┐
│ Model                  │ Accuracy   │
├────────────────────────┼────────────┤
│ CNN Model 1 (Base)     │    84%     │
│ CNN Model 2 (Enhanced) │    86%     │
│ CNN Model 3 (ResNet)   │    89%     │
│ Ensemble (Majority)    │    87%     │
└────────────────────────┴────────────┘

Future Improvements 🔮

• 🔧 Implement more advanced ensemble techniques
• 🎯 Add model interpretability features
• 🚀 Optimize for mobile deployment
• 📊 Add confusion matrix visualization
• 🔄 Real-time data augmentation

Requirements 📋

• Python 3.7+
• TensorFlow 2.17.1
• Streamlit
• NumPy
• OpenCV
• Pillow

License 📄

This project is licensed under the MIT License - see the LICENSE file for details.

---

Made with ❤️ for the Computer Vision community