<![CDATA[Automated waste classification is a critical component of intelligent recycling systems, where model selection must balance predictive performance and computational efficiency. This study benchmarks three representative convolutional neural network (CNN) backbones—ResNet50, EfficientNet-B0, and MobileNetV3-Large—for eight-class waste classification under controlled augmentation and unified optimization protocols. Using a fixed 80/10/10 split (7,747 training, 969 validations, and 960 testing images), all models are evaluated across multiple random seeds to ensure statistical reliability. Performance is assessed using macro-F1, precision, recall, and GPU-based inference latency to characterize the accuracy–efficiency trade-off. EfficientNet-B0 achieves the highest macro-F1 (0.9676 ± 0.0034), while MobileNetV3-Large delivers comparable performance (0.9669 ± 0.0023) with substantially lower latency—approximately six times faster than ResNet50. Augmentation sensitivity analysis further reveals architecture-dependent robustness under occlusion-based perturbation. These results demonstrate that lightweight architectures can achieve near-optimal classification performance with significantly reduced computational cost, providing deployment-oriented guidelines for practical waste sorting systems.]]>
Copyrights © 2026
