Waste Type Classification Using MobileNetV2-Based Convolutional Neural Network

Abstract

Rapid urbanization and population growth have made efficient waste management a critical environmental challenge. While deep learning has advanced automated waste classification, most existing models either rely on computationally heavy architectures that are unsuitable for resource-constrained edge devices or are limited to broad, non-specific waste categories. To address this gap, this study proposes a lightweight and highly granular automated waste classification system utilizing a customized MobileNetV2 architecture. The model was trained using a supervised learning paradigm on a comprehensive dataset comprising 15.515 images categorized into 12 distinct waste classes. By leveraging transfer learning, Global Average Pooling, and dropout regularization, and class balancing, the architecture significantly reduces computational overhead while maintaining robust feature extraction. Experimental results demonstrate that through rigorous hyperparameter tuning to mitigate dataset bias, the fine-tuned model achieved an overall accuracy of 90%. The system exhibited exceptional performance in distinguishing distinct materials such as clothes (F1-score 0.98) and biological waste (F1-score 0.95), although it faced minor challenges with visually similar and reflective materials like plastic and glass. Ultimately, this research demonstrates that a single, optimized lightweight Convolutional Neural Network (CNN) can provide an optimal balance between high classification accuracy and low computational cost, offering a viable AI-based solution for real-time deployment in smart waste management systems.