A deep learning model for classifying waste images into six categories using transfer learning with EfficientNetV2B2.
Built with TensorFlow & Keras and deployed via Gradio on Hugging Face Spaces.
The model can classify images into the following categories:
- 📦 Cardboard
- 🔮 Glass
- ⚙️ Metal
- 📄 Paper
- 🧴 Plastic
- 🚮 Trash
👉 Try the deployed model here: Garbage Classifier on Hugging Face
| Metric | Value |
|---|---|
| ✅ Accuracy | 92.97% |
| 📉 Test Loss | 0.2643 |
📌 The model performs exceptionally well on all classes except “trash”, which is underrepresented and visually ambiguous.
- 📁 Dataset loaded from TrashNet (resized & preprocessed)
- 🔍 Data Augmentation: Flip, Zoom, Contrast, Rotation
- ⚖️ Class balancing using computed weights
- 🧠 Model: EfficientNetV2B2 with top layers replaced
- 🧪 Training: Fine-tuned with EarlyStopping + ReduceLROnPlateau
- TensorFlow / Keras
- EfficientNetV2B2 (ImageNet pretrained)
- Gradio (web deployment)
- Hugging Face Spaces
- Python, NumPy, PIL
| Example Image | Predicted Class | Confidence |
|---|---|---|
| Plastic bottle | Plastic | 0.93 |
| Paper sheet | Paper | 0.91 |
| Metal can | Metal | 0.89 |
| Trash mix (wrong) | Paper ❌ | 0.61 |
⚠️ Note: The “trash” class has lower accuracy due to fewer examples and overlap with other categories.
garbage-classification/
├── app.py # Gradio interface
├── EfficientNetV2B2_model.keras # Trained model
├── requirements.txt # Dependencies
└── README.md # This file
👉 Try the live app on Hugging Face
🧑💻 Author
Made with ❤️ by Seona Ann Tom
🔗 GitHub: @seonaann
📜 License
This project is licensed under the MIT License.
🙌 Acknowledgements
Dataset inspired by TrashNet
Hugging Face for free model hosting via Spaces
Gradio for the fast, no-code interface