Applying semi-supervised learning algorithms in real-environment solid waste classification

doi:10.1016/j.resconrec.2026.108806

Applying semi-supervised learning algorithms in real-environment solid waste classification

2026-01-23

0

PRE

AI

DOI：10.1016/j.resconrec.2026.108806

Original

Original request for help

Save

Abstract

En 中文

To improve intelligent solid waste identification, this study developed semi-supervised learning (SSL) models utilizing pseudo-labeling and consistency regularization techniques. Based on 10,261 image patches covering multiple categories of recyclable solid waste, we evaluated four SSL models (FixMatch, FlexMatch, FreeMatch, and SoftMatch models) across varying labeled data ratio (10%–90%) against a supervised learning (SL) baseline. Results demonstrated that the classification accuracy for the test dataset significantly improved (p < 0.05) as the proportion of labeled data increased in the training dataset, rising from 0.843 ± 0.012 to 0.954 ± 0.004. In the external validation, SSL models exhibited notably stronger robustness and generalizability compared to the SL model which achieved an accuracy of only 0.653 ± 0.122. The FixMatch model, in particular, attained an accuracy of 0.810 ± 0.013. These findings underscore the potential of SSL-based approaches in solid waste recycling systems.

Applying semi-supervised learning algorithms in real-environment solid waste classification

Abstract

Journal

Organization