Replastify

An image-based plastic classification system. Upload or photograph a plastic item and get its resin type, health/safety info, recycling instructions, and eco-friendly alternatives — powered by a fine-tuned EfficientNet-B0 model and the Gemini API.

---

What It Does

• Classifies plastics into 6 types: HDPE, LDPE, PET, PP, PS, PVC
• Returns confidence score and full probability distribution across all 6 classes
• Provides static structured info: common uses, health concerns, recyclability, decomposition time
• Generates disposal suggestions via the Gemini 2.0 Flash API (with a static fallback if unavailable)
• Flags low-confidence predictions (below 70%) with an uncertainty warning
• Supports both file upload (drag & drop) and live webcam capture

---

Tech Stack

Layer	Technology
Backend	FastAPI + uvicorn (Python 3.10+)
ML	PyTorch + torchvision — EfficientNet-B0
AI Suggestions	Google Gemini 2.0 Flash API
Image Processing	Pillow (PIL)
Frontend	HTML5 + CSS3 + Vanilla JS
Config	pydantic-settings + python-dotenv

---

Project Structure

replastify/
├── backend/
│   ├── app/
│   │   ├── main.py              # FastAPI app + startup
│   │   ├── config.py            # Settings (model path, thresholds, Gemini key)
│   │   ├── routes/predict.py    # POST /predict, GET /health, GET /plastic-types
│   │   ├── services/
│   │   │   ├── classifier.py    # Model loading + inference
│   │   │   ├── gemini.py        # Gemini API integration (async, fallback)
│   │   │   └── plastic_info.py  # Static knowledge base for all 6 types
│   │   └── utils/image_utils.py # Image validation + preprocessing
│   ├── models/
│   │   └── best_efficientnet_b0.pth  # Trained model weights (~17 MB)
│   ├── .env                     # API keys (gitignored)
│   └── requirements.txt
├── frontend/
│   ├── index.html
│   ├── css/style.css
│   └── js/app.js
├── notebooks/
│   ├── 01_data_exploration.py
│   ├── 02_split_dataset.py
│   ├── 03_train_model.py        # 2-stage transfer learning pipeline
│   └── 04_evaluate_model.py     # Metrics, confusion matrix, visualizations
├── data/
│   ├── raw/                     # Original labeled images (gitignored)
│   └── processed/               # train / val / test splits (gitignored)
└── docs/
    ├── results_summary.md          # Concise metrics summary
    └── results_analysis.md         # Detailed model evaluation report

---

Setup & Run

Prerequisites: Python 3.10+, pip

# 1. Clone and enter the repo
git clone <repo-url>
cd replastify

# 2. Create and activate virtual environment
python -m venv .venv
source .venv/bin/activate       # Linux/macOS
# .venv\Scripts\activate        # Windows

# 3. Install dependencies
pip install -r backend/requirements.txt

# 4. Configure environment
cp backend/.env.example backend/.env
# Edit backend/.env — add your GEMINI_API_KEY (optional; static fallback works without it)

# 5. Run the server
cd backend
uvicorn app.main:app --reload --port 8000

• App: http://localhost:8000
• API docs (Swagger): http://localhost:8000/docs

The backend serves the frontend as static files — no separate frontend server needed.

---

Model

Property	Value
Architecture	EfficientNet-B0 (ImageNet pretrained)
Custom head	Dropout(0.4) → Linear(1280→256) → ReLU → Dropout(0.3) → Linear(256→6)
Input	224×224×3 (resize 256, center crop 224, ImageNet normalization)
Classes	HDPE, LDPE, PET, PP, PS, PVC
Training	2-stage: frozen backbone (10 epochs) then partial unfreeze (up to 20 epochs)
Optimizer	AdamW with CosineAnnealing LR schedule
Model file	backend/models/best_efficientnet_b0.pth (~17 MB)

The trained model file is gitignored (large binary). To run the app, retrain using the notebooks on a GPU or obtain the weights from the project author.

Test Set Results

Metric	EfficientNet-B0	ResNet50 (baseline)
Test Accuracy	88.94%	87.23%
Macro F1	0.8641	0.8278
Cohen's Kappa	0.8636	0.8422
Top-3 Accuracy	97.45%	97.87%
PP F1 (worst class)	0.6061	0.4286

Macro F1 is the primary metric — it weights all 6 classes equally and is more meaningful than accuracy on the imbalanced dataset. EfficientNet-B0 is deployed.

Note on PP: Polypropylene has the lowest F1 (0.606) due to limited training data (119 images, 6.5% of dataset). The API appends an extra verification reminder for PP predictions.

---

API

POST /predict

Upload a plastic image (jpg/png/webp, max 10 MB). Returns classification results, static plastic info, and disposal suggestions.

Request: multipart/form-data with field file

Response:

{
  "prediction": {
    "plastic_type": "PET",
    "full_name": "Polyethylene Terephthalate",
    "resin_code": 1,
    "confidence": 0.94,
    "is_uncertain": false,
    "uncertainty_message": null,
    "top3": [{"type": "PET", "confidence": 0.94}, ...],
    "all_probabilities": {"PET": 0.94, "HDPE": 0.03, ...}
  },
  "info": {
    "common_uses": [...],
    "recyclability": "Widely recycled",
    "recyclability_score": 5,
    "health_concerns": "...",
    "decomposition_years": 450,
    "warning": null,
    "fun_fact": "..."
  },
  "suggestions": {
    "recycling_tips": [...],
    "reuse_ideas": [...],
    "eco_alternatives": [...],
    "environmental_note": "...",
    "source": "ai"
  }
}

suggestions.source is "ai" when Gemini responds successfully, "static" when the fallback is used.

GET /health

Returns model load status, class count, and whether Gemini is enabled.

GET /plastic-types

Returns the full static knowledge base for all 6 supported plastic types.

Full interactive API documentation: http://localhost:8000/docs

---

Dataset

Split	Images
Train	1,834 (~80%)
Val	226 (~10%)
Test	235 (~10%)
Total	2,295

• 6 classes: HDPE, LDPE, PET, PP, PS, PVC (the "Other" category was excluded — insufficient samples)
• Source: curated academic plastic classification dataset
• Split: stratified 80/10/10, seed=42
• Class imbalance handled via WeightedRandomSampler during training

---

Environment Variables

# backend/.env
GEMINI_API_KEY=your_key_here   # Leave empty to use static suggestions only

All other settings (model path, confidence threshold, file size limit) are in backend/app/config.py.

---

Known Limitations

• PP classification is limited — F1=0.606 due to only 119 training images. Verify PP items by checking the resin code (#5) on the packaging.
• Test set is small — 235 images total, 16 for PP. Metric estimates have moderate variance.
• Distribution gap — Model was trained on curated images. Real-world phone photos (varied lighting, dirty/crushed plastics, complex backgrounds) will likely reduce accuracy by 5–15%.
• No "Other"/"Mixed" class — Items that aren't clearly HDPE, LDPE, PET, PP, PS, or PVC will be misclassified as the closest match.
• No calibration testing — Softmax confidence scores are not calibrated; a 90% confidence prediction is not guaranteed to be correct 90% of the time.