An Automated Segregation System for Self-Sustainable Smart Cities
Proper waste segregation is a cornerstone of sustainable urban development. This project utilizes Deep Learning and Computer Vision to automate the identification of household waste, categorizing materials to reduce recycling contamination.
The system was developed using Transfer Learning on the InceptionV3 architecture, achieving significant improvements in generalization compared to baseline lightweight models.
- Multi-Scale Detection: Uses Inception modules to identify waste items of various shapes and sizes.
- Fine-Tuned Accuracy: Optimized through partial unfreezing of deep layers to specialize in waste textures.
- Interactive UI: A real-time web dashboard for image uploading and classification.
- Smart City Logic: Integrated decision-making to label items as "Recyclable" or "General Trash."
I performed architecture benchmarking to determine the best model for this use case.
| Metric | MobileNetV2 (Baseline) | InceptionV3 (Final) |
|---|---|---|
| Training Accuracy | 80.2% | 75.9% |
| Validation Accuracy | 47.9% | 70.0% |
| Key Advantage | Resource Efficient | Superior Generalization |
- Language: Python 3.10
- Deep Learning: TensorFlow 2.15+, Keras (Legacy Support via
tf-keras) - Architecture: InceptionV3
- Frontend: Streamlit
- Deployment: Virtual Environments (venv)