Flowers Classification: Transfer Learning & Fine-Tuning

This repository contains a complete Deep Learning pipeline for classifying flower species with high precision. The project demonstrates the power of Transfer Learning and Fine-Tuning using the MobileNetV2 architecture to build a robust classifier even with limited data.

📂 Dataset Overview

The model is trained on a dataset containing 5 classes:

• Daisy
• Roses
• Sunflowers
• Tulips
• Dandelion

This repository presents a complete end-to-end deep learning pipeline for high-accuracy flower species classification.
The project leverages transfer learning and fine-tuning with the MobileNetV2 architecture to build a robust image classifier, even with a relatively limited dataset.

The implementation demonstrates best practices in data pipeline optimization, model training strategy, and performance evaluation.

---

📂 Dataset Overview

The model is trained on a multi-class image dataset containing five flower categories:

• Daisy
• Roses
• Sunflowers
• Tulips
• Dandelion

---

🛠️ Technical Highlights

🔹 Efficient Data Pipeline

• Built using tf.data.Dataset for scalable and high-performance data loading
• Includes image decoding, preprocessing, batching, and prefetching
• Eliminates I/O bottlenecks during training

🔹 Exploratory Data Analysis (EDA)

• Interactive visualizations using Plotly
• Class distribution and dataset balance analysis

🔹 Model Architecture

• Feature Extractor: MobileNetV2 (pretrained on ImageNet)
• Custom Classification Head
  • GlobalAveragePooling2D
  • Dense output layer (Softmax)

🔹 Two-Stage Training Strategy

Phase 1 — Feature Extraction

• Base model frozen
• Training classification head only

Phase 2 — Fine-Tuning

• Top layers of MobileNetV2 unfrozen
• Very low learning rate
• Improved domain-specific feature learning

---

📊 Final Results

The two-stage training strategy significantly improved generalization and predictive performance.

Metric	Training	Validation
Final Accuracy	~99%	~91%
Final Loss	~0.05	~0.25

Key Observations

✔ Smooth convergence of training and validation loss
✔ Stable accuracy improvement across epochs
✔ No severe overfitting after fine-tuning
✔ Strong generalization on unseen images

---

🧬 Model Structure

# Architecture Summary
- Input (224, 224, 3)
- MobileNetV2 (Base)
- GlobalAveragePooling2D
- Dense (128 units, ReLu )
- Dense (5 units, Softmax)

🎯 Project Objectives

• Demonstrate practical transfer learning workflow

• Implement optimized TensorFlow input pipelines

• Apply structured fine-tuning strategy

• Provide transparent evaluation and visualization

🚀 How to Run

1. Environment: Open the .ipynb file in Google Colab or any Jupyter environment.
2. Dependencies: Ensure you have tensorflow, plotly, pandas, and matplotlib installed.
3. Data: The notebook is designed to work with images stored in Google Drive; update the drive.mount path.

👨‍💻 Author

Samir Mohamed , AI and Computer vision Engineer .