Page 209 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

WEE is composed of the following key modules:

• Object Detection Module: Detects and localizes visible waste objects from images
captured at the recycling stations.
• Feature Matching Module: Extracts and analyzes fine-grained visual features such as Change 4.2-Climate
texture, color, shape, and labeling patterns to differentiate between different material
types.
• Fine-grained Classification Module: Maps the detected objects to specific recycling
categories according to region-specific regulations, based on subtle visual differences.
• Object Tracking and Counting Modules: Tracks detected objects over time and counts
items for operational reporting. These modules use pretrained models optimized for
visual tracking tasks and are not subject to retraining.
• Adaptive Feedback Loop: Collects operational data (e.g., misclassifications, novel
material types) and uses it to refine detection and classification modules through periodic
retraining.

Pipelines for Training, Deployment, and Update

• Training:
WEE is trained through supervised learning on a large, annotated dataset of real-world
waste images, capturing various material types, contamination levels, and region-specific
labeling variations.
• Deployment:
After training, models are containerized and deployed on edge devices equipped with
GPU acceleration for high-speed inference directly at recycling sites.
• Update:
New data collected from field operations – especially misclassified items or newly
encountered materials – is aggregated, annotated, and incorporated into retraining
cycles. This ensures that WEE continuously adapts to operational conditions without
disrupting service.

Model Verification in Relation to Local Regulations

Before deployment, WEE models are rigorously verified against the recycling rules and
regulations specific to the target region:
• A reference dataset reflecting local regulatory standards is curated.
• Model outputs are validated by comparing classification results against the reference set.
• Pilot deployments are conducted, during which model predictions are manually audited
by local waste management authorities or regulatory partners.
• This ensures that WEE consistently aligns with local recycling standards and can
accommodate regulatory updates over time.

Fine-tuning for Country-Specific Requirements

To meet the needs of diverse regulatory environments, WEE is customized and fine-tuned for
each country:
• Label Mapping Adjustment:
The classification schema is modified to reflect material categories defined by local
recycling rules.
• Feature Matching Enhancement:
In Canada, for example, WEE was specifically enhanced to differentiate between:

o Non-alcohol cans (e.g., soda cans)

173