♻️ VisionSort

VisionSort is a deep learning project for automated waste classification using MobileNetV2 and a custom Loss-Feedback Cosine Warm Restart Scheduler. The model classifies waste into 12 different categories to support smart recycling and waste management. :contentReference[oaicite:0]{index=0}

---

📌 Features

• 12-class waste image classification
• Transfer Learning with MobileNetV2
• Two-phase training (Frozen Backbone → Fine-tuning)
• Custom Loss-Feedback Cosine Warm Restart Scheduler
• Class-weighted training for imbalanced data
• Model checkpointing and resume support
• Training and validation visualization

---

🗂️ Dataset

• Dataset: Garbage Classification
• Source: Kaggle
• Total Images: 15,515
• Classes:
  • Battery
  • Biological
  • Brown Glass
  • Cardboard
  • Clothes
  • Green Glass
  • Metal
  • Paper
  • Plastic
  • Shoes
  • Trash
  • White Glass

---

🛠️ Technologies Used

• Python
• TensorFlow / Keras
• MobileNetV2
• NumPy
• Pandas
• Scikit-learn
• Matplotlib

---

📊 Model Performance

Metric	Value
Validation Accuracy	97%
Macro F1 Score	96%
Weighted F1 Score	97%
Best Epoch	16

---

📈 Training Strategy

• Phase 1: Train classification head
• Phase 2: Fine-tune top MobileNetV2 layers
• AdamW Optimizer
• Label Smoothing
• Early Stopping
• Model Checkpointing
• Adaptive Learning Rate Scheduler

---

📷 Results

Training Curves

• Training Loss decreases consistently.
• Validation Accuracy reaches 97%.
• Adaptive learning rate improves convergence.

Confusion Matrix

Most classes achieve high classification accuracy with minor confusion between visually similar classes like plastic and glass.

---

📂 Project Structure

VisionSort/
│
├── notebook.ipynb
├── README.md
├── checkpoints/
├── images/
├── results/
└── requirements.txt

---

🚀 Future Improvements

• Deploy using TensorFlow Lite
• Add Grad-CAM visualization
• Improve glass category classification
• Build a web/mobile interface
• Test on real-world recycling images