CIFAR-10 Image Classification using CNN and Transfer Learning

Project Overview

This project explores image classification on the CIFAR-10 dataset using both:

1. A custom Convolutional Neural Network (CNN)

2. Transfer Learning models:

   • MobileNetV2
   • DenseNet121
   • EfficientNetB0

The project was developed and trained using Google Colab with TensorFlow and Keras.

The goal of this project was to:

• Understand image classification workflows
• Compare traditional CNNs with modern pretrained architectures
• Evaluate transfer learning performance on CIFAR-10
• Analyze model accuracy, overfitting, validation stability, and computational efficiency

---

Dataset

CIFAR-10

CIFAR-10 is a widely used benchmark dataset for image classification.

Dataset Details

• Total Images: 60,000
• Training Images: 50,000
• Testing Images: 10,000
• Image Size: 32 × 32 × 3
• Number of Classes: 10

---

Technologies Used

Development Environment

• Google Colab
• Python 3

Libraries and Frameworks

• TensorFlow
• Keras
• NumPy
• Matplotlib
• Seaborn
• Scikit-learn
• Pandas

---

Project Structure

CIFAR10-Project/
│
├── cnn_model.ipynb
├── mobilenet_model.ipynb
├── densenet_model.ipynb
├── efficientnet_model.ipynb
├── comparison_notebook.ipynb
├── saved_models/
│   ├── cnn_model.keras
│   ├── mobilenet_model.keras
│   ├── densenet_model.keras
│   └── efficientnet_model.keras
└── README.md

---

1. Custom CNN Model

A custom CNN architecture was designed from scratch.

Architecture Components

• Convolutional Layers
• Batch Normalization
• Max Pooling
• Dropout
• Global Average Pooling

---

Training Strategy

For transfer learning models:

1. Base model layers were frozen
2. Only the custom classifier was trained
3. Adam optimizer was used
4. Validation accuracy was monitored

Fine-Tuning

After initial convergence:

• Selected deeper layers were unfrozen
• Learning rate was reduced
• Additional training epochs were applied

Fine-Tuning Benefits

• Improved validation accuracy
• Better feature adaptation to CIFAR-10
• Reduced train-validation performance gap

---

Challenges Faced

1. RAM Usage

Initial preprocessing pipelines caused high RAM consumption due to resizing all images simultaneously.

Solution

Implemented TensorFlow dataset pipelines for dynamic preprocessing.

---

2. Overfitting

Training accuracy increased while validation accuracy plateaued.

Solutions

• Dropout layers
• Smaller classifier heads
• Fine-tuning selected layers only
• Early stopping

---

3. Runtime Disconnections in Colab

Google Colab occasionally disconnected during long training sessions.

Solutions

• Saved models to Google Drive
• Reduced unnecessary epochs
• Used lightweight architectures

---

Future Improvements

Potential future enhancements include:

• Data augmentation
• Hyperparameter optimization
• Ensemble learning
• Advanced fine-tuning strategies
• Learning rate schedulers
• Larger image resolutions
• Training on larger datasets

---

Conclusion

This project demonstrated the effectiveness of transfer learning for image classification tasks on CIFAR-10.

Key conclusions:

• Transfer learning significantly outperformed the custom CNN baseline.
• MobileNetV2 provided excellent efficiency and fast training.
• DenseNet121 achieved stronger feature extraction.
• EfficientNetB0 delivered the best overall balance between accuracy and efficiency.
• Fine-tuning pretrained layers improved validation performance.
• TensorFlow dataset pipelines improved memory efficiency and training stability.

Overall, EfficientNetB0 achieved the strongest generalization performance among the tested architectures.

---

Team Members

Member	Model
Laxmi	MobileNetV2
Jan	DenseNet121
Bruno	EfficientNetB0

---

References

• TensorFlow Documentation
• Keras Applications Documentation
• CIFAR-10 Dataset
• ImageNet Pretrained Models
• Google Colab Documentation