📸 Deep Learning: Image Classification using CNN and Transfer Learning

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

1. Building a Convolutional Neural Network (CNN) from scratch
2. Implementing Transfer Learning with MobileNet

The project demonstrates how custom-built CNNs compare to transfer learning models in terms of performance, training time, and generalization ability.

---

📌 1. General Description

• Built a CNN model to classify CIFAR-10 images into predefined categories.
• Implemented a transfer learning approach using the pre-trained MobileNet model.
• Compared both models using evaluation metrics and visualizations.

---

📊 2. Project Overview

🔍 Features

• Performs image classification on the CIFAR-10 dataset (60,000 images, 10 classes).
• Implements two modeling approaches: CNN from scratch and MobileNet transfer learning.
• Evaluates and compares models using accuracy, loss, and confusion matrices.

🎯 Problem Solved

• Automates classification of small-scale images into meaningful categories.
• Demonstrates practical differences between training models from scratch and leveraging pre-trained architectures.

🌍 Potential Impact

• Education: Useful for learning CNNs and transfer learning.
• Industry: Shows how transfer learning accelerates training and improves results.
• Research: Provides a foundation for further image classification experiments.

---

📁 3. Dataset Description

• Dataset: CIFAR-10
• Images: 60,000 (32x32 color images, 10 classes).
• Split: 50,000 training / 10,000 testing.
• Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck.
• Use case: Benchmark dataset for computer vision tasks.

---

🎯 4. Research Goal

• Classify CIFAR-10 images into 10 categories.
• Compare CNN vs. MobileNet transfer learning.
• Evaluate using accuracy, loss, and visualizations.

---

⚙️ 5. Steps Taken

1. Data Preprocessing

• Normalized image pixel values.
• Applied data augmentation to improve generalization.

2. Model Building

• Custom CNN with convolutional, pooling, and dense layers.
• MobileNet Transfer Learning, fine-tuning the top layers.

3. Training

• Used Adam optimizer and early stopping.
• Tuned learning rate for stability.

4. Evaluation

• Compared test accuracy and loss.
• Visualized training history and confusion matrices.

---

🔍 6. Key Findings

• MobileNet transfer learning outperformed the CNN trained from scratch.
• Transfer learning achieved higher accuracy with fewer epochs.
• The custom CNN worked but required more parameter tuning.
• MobileNet showed better generalization to unseen data.

---

🧪 7. How to Reproduce

Requirements

• Python: 3.11.5
• Libraries:

  import tensorflow as tf
  from tensorflow.keras import datasets, layers, models
  import matplotlib.pyplot as plt
  import numpy as np

Files

Project_3_(CNN_LN).ipynb → CNN + MobileNet implementations Project_3 Presentation.pdf → Project presentation slides requirements.txt → Dependencies

🚀 8. Next Steps / Improvements

Try other pre-trained models (ResNet, VGG, EfficientNet). Perform hyperparameter tuning. Add advanced data augmentation and regularization. Explore deploying the trained model as a web/app service.

🗂️ 9. Repository Structure

File/Folder Description Project_3_(CNN_LN).ipynb Notebook containing CNN and MobileNet models Project_3 Presentation.pdf Slide presentation of project findings requirements.txt Python dependencies required to run project

📢 Contact Information

Project Lead: Nekky Lung Email: nekkytang@gmail.com LinkedIn: linkedin.com/in/nekkytang GitHub: https://github.com/nekkylung Repository: Deep-Learning-Image-Classification-using-CNN-and-Transfer-Learning

📜 License

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

🙏 Acknowledgments

This project was developed as part of a collaborative team-based analysis, simulating real-world analytics scenarios and demonstrating practical applications of machine learning in computer vision and financial markets.