Federated continual learning for vision-based plastic classification in recycling

Accurate classification of plastic types is essential for high-purity sorting in recycling processes, a key requirement for sustainable waste management. However, current industrial solutions often rely on expensive hyperspectral imaging systems or on centralized machine learning models that struggle to generalize across facilities and adapt to evolving waste streams. This study addresses these limitations by investigating vision-based particle classification using a Federated Continual Learning framework, which enables decentralized training across multiple recycling plants while preserving data privacy and supporting model adaptation over time. In this domain, the lack of publicly available image datasets is significant. To support this approach, a novel image dataset containing six major plastic types was created from material collected at a local recycling facility. Experimental evaluations benchmarked various combinations of Federated Learning and Continual Learning techniques. Federated Continual Learning achieved an overall classification accuracy of 83.68%, with per-class accuracies ranging from 74.8% to 99.8%, performing comparably to conventional Transfer Learning methods. These results demonstrate that Federated Continual Learning can offer robust, scalable, and privacy-preserving solutions for real-world plastic sorting, contributing to more cost-effective operations and supporting the development of circular economy practices.