Edge-Deployed Waste Classification System Using Transfer Learning on Custom IoT Camera Data with Data Augmentation

Abstract

Rapid urbanization and inefficient waste segregation have intensified the global waste management crisis, with Municipal Solid Waste projected to reach 3.8 billion tonnes annually by 2050. This study proposes a smart automated waste classification system integrating a lightweight MobileNetV2 convolutional neural network with an IoT-based Raspberry Pi 4 platform to enable real-time, cloud-independent automated waste sorting across five categories which are Metal, Plastic, Paper, Organic, and None (not_waste). A dataset of 671 manually collected images was used to train and compare two models: a baseline model achieving 71.20% accuracy and an augmentation-enhanced model achieving 93.60% accuracy with a macro F1-score of 0.9344, where offline data augmentation proved to be the single most important factor, dramatically improving Organic class recall from 0.22 to 0.96. The deployed system captures waste images in real time, classifies them on-device, and automatically activates a sorting mechanism, confirming that lightweight CNN architectures combined with balanced training data offer a practical and low-cost solution for smart waste management in resource-constrained real-world environments.