An enhanced EfficientNet framework for automated waste classification using cosine annealing and label smoothing

Waste management is a crucial task due to its direct impact on environmental sustainability, public health, and resource conservation. Rapid urbanization, population growth, and industrialization have led to a significant increase in the volume and complexity of waste generated by municipal, biomedical, electronic, and hazardous waste. Therefore, technology-based waste classification is needed to address the global waste management crisis. However, deploying an accurate classification model on resource-constrained hardware demands a careful balance between model accuracy and overall computational efficiency. This research employs an EfficientNet-B1 + Transformer model along with Cosine Annealing Learning Rate Scheduling (CALRS) and optimized Label Smoothing (LS) techniques, and this method is known as EN-CALS. This unified pipeline is trained on six classes of waste dataset of 2,527 images (cardboard, glass, metal, paper, plastic, and trash). It was compared with several fine-tuned models such as ResNet-50, ViT-B/16, MobileNet-V2, MobileNet-V3-Small, MobileNet-V3-Large, ShuffleNet-V2, EfficientNet-B0/B2, and DenseNet-121. Results show us that EfficientNet-B1 with these optimizations achieves the best test accuracy of 93.68. Large-scale architectures like ResNet-50 with 87.37% accuracy, ViT-B/16 with 61.58% accuracy and DenseNet121 with 92.63% accuracy, show clear overfitting and are unsuitable for the target edge hardware. The ablation studies proved the importance of individual components in EN-CALS for the waste classification system. This research establishes a reusable model template and deployment pipeline for waste classification and is designed to scale as better datasets and hardware become feasible.