A Reproducible Benchmark of AdamW-Augmented Lightweight Models for Trash Classification

Abstract

Global waste generation is projected to reach 3.40 billion tonnes by 2050, creating urgent demands for automated waste classification systems that can overcome the limitations of manual sorting methods. Current deep-learning research on waste classification lacks standardised evaluation protocols, preventing meaningful architectural comparisons and hindering the progress of reproducible research. This paper establishes a reproducible benchmark framework for lightweight neural network models designed explicitly for trash classification research applications. Lightweight models are designed for optmised architecture and computation cost while maintain accuracy. Four representative lightweight models, including MobileNet V3 Large, Vision Transformer (ViT) Small, EfficientFormer, and ShuffleNet V2, were systematically evaluated on the TrashNet dataset using identical training protocols. All models employed AdamW optimisation with a learning rate of 1 × 10^-4, weight decay of 1 × 10^-4, and CosineAnnealingLR scheduling through 5-fold stratified cross-validation on RTX 2080 Ti hardware. Experimental results demonstrate that ViT Small achieved the highest classification accuracy at 0.815 but required 21.67M parameters, while MobileNet V3 Large delivered superior computational efficiency with 0.768 accuracy and 0.72ms inference time using only 4.21M parameters. Statistical analysis revealed significant performance differences across models (p = 0.0002), with hardware-aware architectural optimisations proving more critical than raw parameter reduction for computational performance on data centre GPU hardware. The standardised evaluation framework and open-source implementation provide rigorous baselines for advancing automated waste classification research.

Manuscript received: 12 Jun 2025 | Revised: 7 Aug 2025 | Accepted: 11 Aug 2025 | Published: 30 Nov 2025