Using a custom dataset of 1,000 images in 10 waste classes, the study compared three ResNet backbones, three classifier heads and three losses. ResNet50 with a DenseNet201-based head and ArcFace achieved 98.4% accuracy, 98.3% F1 and 98.1% retrieval mAP, remaining stable under rotation and scaling. ResNet34 with a fully connected head and ArcFace offered a lighter performance-complexity trade-off.
Key findings
- ResNet50, DenseNet201 head and ArcFace produced the top scores. • ArcFace improved both classification and retrieval features. • ResNet34 offered a lighter compromise.
Why this matters globally
Automated recognition could support recycling, but high scores on a small curated dataset do not guarantee performance on cluttered, contaminated real conveyor streams.
Thai researcher contribution
A Rajamangala University of Technology team links classification and image retrieval for waste-management AI.
Limitations to consider
The custom 1,000-image dataset lacks reported external or field testing, leaving substantial domain-shift and overestimation risk.