PlasticSense is an AI-powered waste classification system developed at Sorbonne Université ( Fablab Laboratory ). An image of the waste is fed to a lightweight neural network that classifies it as recyclable or non-recyclable in real time.
For a complete and detailed breakdown of the methodology, dataset, model architecture, training pipeline, and results:
| Language | Report | Details |
|---|---|---|
| 🇬🇧 English | PlasticSense_report_EN.pdf | Transfer learning, MobileNetV3-Small fine-tuning, augmentations, Early Stopping, CosineAnnealingLR — full results: 97.7% test accuracy |
| 🇫🇷 Français | PlasticSense_rapport_FR.pdf | Jeu de données, méthodologie complète, progression V1→V4, discussion des résultats |
Everything is covered in detail in both reports, from the raw dataset construction to the final deployed model.
- 🎯 Test Accuracy : 97.7%
- 📈 Best Validation Accuracy : 98.4%
- ⚖️ F1 Score : 0.977
- 🚩 Initial Target : 90% (exceeded by +7.7 points)
| Version | Accuracy | What changed |
|---|---|---|
| V1 | 68.4% | Baseline model |
| V2 | 73.6% | First dataset cleanup |
| V3 | 84.6% | Added real iPhone photos |
| V4 | 97.7% | Full optimized pipeline |
Image input
↓
MobileNetV3-Small classifies
↓
Recyclable ♻️ / Non-recyclable 🚯
git clone https://github.com/adnanegrb/PlasticSense.git
cd PlasticSense
pip install -r requirements.txtTrain the model:
python train.pyRun inference on a single image:
python inference.py --image path/to/image.jpgOutput example:
♻️ RECYCLABLE
Confidence : 98.3%
The dataset is not included in this repo. See dataset/README.md for download instructions.
8,832 images split into two classes: recyclable (plastic, paper, cardboard, metal) and non-recyclable (biological waste, mixed trash). Split: 80% train / 10% val / 10% test.
MobileNetV3-Small pretrained on ImageNet, with the last 3 blocks unfrozen for fine-tuning. Trained with AdamW optimizer, CosineAnnealingLR scheduler, WeightedRandomSampler, and Early Stopping (patience = 3).
Mohammed Adnane Garab — Sorbonne Université
GitHub
MIT License