PET Perplexity

An Automated Polymer Segregator

AI-powered computer vision system for automated PET bottle sorting, grading, and analytics in recycling facilities.

---

Overview

PET recycling facilities handle high-volume mixed bottle streams where manual sorting is slow, inconsistent, and difficult to audit.
PET Perplexity automates bottle analysis from conveyor-belt images to classify:

• Color category
• Material type (PET vs non-PET)
• Size/capacity class
• Brand
• Estimated weight

It also provides a dashboard for batch-level quality, composition, and pricing insights.

---

Problem Statement

Mixed bottle bales reduce recycling output value when contaminants and colors are not separated correctly.

• Clear PET can fetch higher rates (INR 40-50/kg)
• Colored PET often fetches lower rates (INR 25-38/kg)
• Non-PET contamination can reduce full-batch value

This project aims to increase sorting accuracy, consistency, and traceability with AI-driven visual inspection.

---

MVP Features

1) Color Code Identification

Classifies bottles into:

• Clear / Transparent
• Light Blue
• Green
• Brown / Amber
• Mixed / Other

2) Material Type Detection (PET vs Non-PET)

Detects PET characteristics (e.g., recycle code, optical appearance) and flags contaminants with confidence scores.

• PET / PETE / Code 1
• Non-PET examples: HDPE (2), PVC (3), PP (5)

3) Bottle Size Classification

Approximate capacity buckets:

• 200-300 ml
• 500-600 ml
• 1 L
• 1.5-2 L

• 2 L (bulk)

4) Brand Identification

Identifies common brands using logo, label pattern, shape cues, and cap-color combinations.

Examples:

• Water: Bisleri, Kinley, Aquafina
• Soft drinks: Coca-Cola, Pepsi, Sprite, Fanta
• Juice: Minute Maid, Tropicana
• Generic / Unbranded

5) Weight Estimation

Estimates per-bottle weight from size and visual thickness profile.

Size	Typical Weight	Height	Diameter
200 ml	10-12 g	120-140 mm	50-55 mm
500 ml	18-25 g	180-210 mm	60-70 mm
1 L	35-45 g	250-280 mm	80-90 mm
2 L	55-70 g	320-350 mm	95-105 mm

6) Analytics Dashboard

Batch-level visual analytics:

• Total bottles detected
• Color distribution (pie chart)
• PET vs non-PET ratio
• Size distribution (bar chart)
• Brand-wise count
• Total estimated weight
• Quality grade (% clear PET)
• Export report (PDF/Excel)

---

Suggested System Architecture

• Input Layer: Image/video ingestion from conveyor cameras
• Detection Layer: Bottle localization and instance extraction
• Classification Layer: Multi-head model for color, material, size, brand
• Estimation Layer: Weight regression/class-based estimator
• Analytics Layer: Batch aggregation + KPI computation
• Dashboard Layer: Interactive web app for operations and reporting

---

Tech Stack (Suggested)

• Computer Vision / DL: Python, OpenCV, PyTorch or TensorFlow
• Model Serving: FastAPI / Flask
• Data & Analytics: Pandas, NumPy, Plotly/Matplotlib
• Dashboard: Streamlit / Dash / React + backend API
• Storage: PostgreSQL / MongoDB + object storage for images
• Deployment: Docker, on-prem edge GPU or cloud inference

---

Dataset Requirements

Collect labeled images across:

• Lighting conditions (day/night, indoor/outdoor)
• Conveyor motion blur levels
• Bottle deformation/compression states
• Label-on / label-off cases
• Brand variants and counterfeit packaging
• Contaminants and non-PET lookalikes

Annotation Schema (Recommended)

For each bottle instance:

• Bounding box / segmentation mask
• color_class
• material_class
• size_class
• brand_class
• estimated_weight_target (if supervised)
• Optional OCR tags (PET, recycle codes)

---

Evaluation Metrics

• Detection: mAP@IoU
• Classification: Accuracy, Precision, Recall, F1 (per class)
• Material Rejection Quality: False acceptance / false rejection rates
• Weight Estimation: MAE / RMSE
• Operational KPI: Throughput (bottles/min), dashboard latency

---

Business Impact

• Reduced manual sorting effort
• Higher bale quality and resale value
• Early contaminant rejection
• Better inventory transparency
• EPR compliance through brand-wise tracking

---

Future Enhancements

• Spectral/NIR fusion for higher material certainty
• Active learning loop from operator corrections
• Real-time PLC integration for actuator-based sorting
• Dynamic price optimization module by batch composition
• Multi-camera 3D geometry for better volume/weight estimation

---

Project Status

MVP Definition Complete
Next phase: data collection, annotation pipeline, baseline model training, and dashboard prototype.

---