🧠 Image Classification using CNN vs MobileNetV2

📌 Project Overview

This project compares two deep learning approaches for image classification of products:

• A Custom Convolutional Neural Network (CNN) built from scratch
• A MobileNetV2-based Transfer Learning model (🏆 Best Performing)

It aims to solve challenges related to class imbalance, small dataset size, and overfitting, using real-world product images from multiple categories.

---

📂 Project Resources

🔹 Dataset (Product Images): 1,909 images in 5 classes (Not uploaded due to size)
🔹 Project Code (.ipynb): View Notebook
🔹 Project Report (DOCX File): Download Report

---

🧾 Dataset Overview

• Total Images: 1,909
• Classes: Product_1 to Product_5
• Images per class:
  • Product_1: 510
  • Product_2: 14 (⚠️ Highly Imbalanced)
  • Product_3: 400
  • Product_4: 385
  • Product_5: 600
• Preprocessing:
  • Resized to 224x224 pixels
  • Normalized pixel values to [0,1]
  • Real-time augmentation: rotation, shift, shear, zoom, flip
  • Class weights for imbalance handling

---

🛠 Model Architectures

1️⃣ Custom CNN

• Conv2D (32 → 64 → 128) + MaxPooling
• Dense(128) + Dropout(0.5)
• Softmax output layer
• Optimizer: Adam (lr=1e-3)
• EarlyStopping and ModelCheckpoint
• Epochs: 25

2️⃣ MobileNetV2 Transfer Learning (🏆 Best Model)

• Pretrained on ImageNet (without top layers)
• Classification head: GAP → Dense(128) → Dropout → Dense(5, softmax)
• Training Strategy:
  • Phase 1: Freeze base, train head for 10 epochs
  • Phase 2: Unfreeze last 20 layers, fine-tune for 5 epochs (lr=1e-5)
• EarlyStopping enabled in both phases

---

📈 Model Performance

Model	Accuracy	Test Loss	Remarks
Custom CNN	92.68%	0.25	Overfitting observed
MobileNetV2	97.21%	0.07	Generalized well, faster training

📌 Note: Both models failed to correctly classify Product_2 due to very limited training data.

---

📊 Evaluation Metrics

• Accuracy
• Loss
• Precision, Recall, F1-Score (per class)
• Confusion Matrix
• Training & Validation Curves

---

🧠 Key Observations

• Transfer learning significantly improved accuracy and generalization
• CNN showed signs of overfitting and required more memory
• Severe class imbalance led to 0% F1 for Product_2 in both models
• MobileNetV2 had ~165K trainable parameters vs CNN’s ~11M

---

🚀 Future Scope

🔹 Apply advanced synthetic augmentation (e.g., GANs) for minority classes
🔹 Explore EfficientNet or ResNet50 for better transfer performance
🔹 Consider active learning for iterative data improvement
🔹 Deploy the model via Flask or Streamlit for real-time classification

---

🏗 Project Highlights

This project demonstrates a complete deep learning pipeline including:
✔️ Data Preprocessing & Augmentation
✔️ CNN & Transfer Learning Model Development
✔️ Evaluation with Visualization
✔️ Addressing Class Imbalance using Weights
✔️ Academic Report & Documentation

---

👥 Authors

• Sunny Rao Karegam
• Sandeep Kumar Kandagatla
• Srikanth Kannamoni
• Alphin Stivi John

🎓 MSc Data Analytics – Dublin Business School (2025)

---

📌 Note

This is an academic project submitted as part of the Machine Learning & Pattern Recognition module.
The dataset is not included in this repository due to space/privacy restrictions.
Please contact the authors for access or use a similar open-source dataset for replication.

---