Comparative Optimization of EfficientNetB3, MobileNetV2, and ResNet50 for Waste Classification

Abstract

Waste management has become a critical challenge in efforts to maintain environmental sustainability and public health. Poorly managed waste can cause environmental pollution, reduce quality of life, and complicate recycling processes. To address this issue, this study aims to classify waste based on images while optimizing several deep learning architectures, namely EfficientNetB3, MobileNetV2, and ResNet50, to identify the best model for waste classification. The research methodology includes data collection, preprocessing, data augmentation, model development, and performance evaluation using accuracy, precision, recall, and F1-score metrics. The dataset, obtained from the Kaggle platform, consists of 4,650 images divided into six categories: battery, glass, metal, organic, paper, and plastic. The results show that EfficientNetB3 with the Adam optimizer achieved the best performance, with accuracy, precision, recall, and F1-score all at 93%, followed by ResNet50 at approximately 91%, and MobileNetV2 ranging from 70–73% depending on the optimizer. The use of different optimizers was found to influence model performance, and data augmentation helped improve generalization, especially for classes with limited samples. Limitations of this study include the relatively limited dataset coverage. Future research is recommended to expand the dataset and explore alternative or hybrid architectures. These findings demonstrate the potential of deep learning–based systems in supporting sustainable waste management.