Status: ✅ Deployed
A deep learning-powered system to automatically classify household waste into six recyclable categories using state-of-the-art transfer learning with advanced CNN backbones.
This project benchmarks EfficientNetV2 (S, M, B2, L), EfficientNetB0, MobileNetV2, and ResNet50 — identifying the best model for real-world deployment.
An interactive Gradio web app is hosted live on Hugging Face Spaces for anyone to test!
This classifier can recognize:
- 📦 Cardboard
- 🧪 Glass
- ⚙️ Metal
- 📄 Paper
- 🧴 Plastic
- 🚮 Trash
Try it instantly in your browser — no setup needed:
👉 Launch the Gradio App on Hugging Face Spaces
- ✅ Multiple CNN Backbones: Benchmarks EfficientNetV2S, EfficientNetV2M, EfficientNetV2B2, EfficientNetV2L, EfficientNetB0, MobileNetV2, and ResNet50.
- 🔬 Transfer Learning: Leverages pretrained ImageNet weights for faster convergence and better generalization.
- ⚖️ Class Imbalance Handling: Uses computed class weights to balance the dataset effectively.
- 📊 Clear Performance Metrics: Visualizes training curves, final test accuracy, classification reports, and confusion matrix.
- 🏆 Top Model: EfficientNetB0 achieved 91.97% test accuracy, outperforming larger variants while remaining computationally efficient.
- ☁️ Deployed on Hugging Face Spaces: Zero local setup — fully serverless with Gradio.
- 🧩 Modular Code: Clean notebooks for data prep, training, evaluation, and deployment.
- 🖼️ Sample Predictions: Visualizations of real test images with predicted vs. true labels.
| Model | Test Accuracy |
|---|---|
| EfficientNetV2S | 89.16% |
| EfficientNetV2M | 91.57% |
| EfficientNetV2B2 | 88.76% |
| EfficientNetV2L | 91.57% |
| EfficientNetB0 | 🏆 91.97% |
| MobileNetV2 | 19.68% |
| ResNet50 | 85.54% |
Conclusion:
EfficientNetB0 offers the best balance of accuracy and efficiency for this dataset — making it the ideal choice for real-world deployment.
├── notebooks/ # Jupyter Notebooks: training, evaluation, deployment
├── README.md # This file
├── models/ # Saved models (.keras)
├── gradio_app.py # Gradio inference script
├── requirements.txt # Dependencies