Federated continual learning for vision-based plastic classification in recycling.

Area of Science:

• Materials Science
• Computer Science
• Environmental Science

Background:

• Accurate plastic classification is vital for high-purity recycling and sustainable waste management.
• Current methods like hyperspectral imaging are expensive, and centralized machine learning models lack adaptability.
• Decentralized, privacy-preserving AI solutions are needed to overcome these limitations in real-world recycling facilities.

Purpose of the Study:

• To investigate vision-based plastic particle classification using a Federated Continual Learning (FCL) framework.
• To address the lack of public datasets by creating a novel image dataset of six major plastic types.
• To evaluate the performance of FCL against conventional methods for plastic sorting.

Main Methods:

• Development of a novel image dataset of six common plastic types from local recycling facility waste.
• Implementation and benchmarking of various Federated Learning (FL) and Continual Learning (CL) technique combinations.
• Experimental evaluation of the Federated Continual Learning approach for decentralized, adaptive plastic classification.

Main Results:

• Federated Continual Learning achieved an overall plastic classification accuracy of 83.68%.
• Per-class accuracies ranged significantly, from 74.8% to 99.8%, demonstrating robust performance across different plastic types.
• FCL performance was comparable to traditional Transfer Learning methods, validating its effectiveness.

Conclusions:

• Federated Continual Learning provides a robust, scalable, and privacy-preserving solution for real-world plastic sorting.
• This approach supports adaptive model performance, crucial for evolving waste streams and diverse recycling facilities.
• FCL contributes to cost-effective operations and advances circular economy practices through improved recycling efficiency.