Deep Learning-Based Lightweight and Efficient Garbage Classification with Two-Phase Fine-Tuning of MobileNetV2

The rapid urbanization and population growth have necessitated the development of intelligent and connected solutions integrated into IoT systems to optimize waste collection , reduce processing costs, and improve the quality of the urban environment. In this article, we propose an waste classification model based on the fine-tuning of MobileNetV2, a pre-trained convolutional neural network, optimized using the Adaptive Moment Estimation (Adam) Algorithm. With a size of only 10 MB, the model is compatible with integration into IoT architectures for the detection, collection and automated centralisa-tion of waste-related data. Evaluated on a public Kaggle dataset comprising 12 classes (battery waste, biological, cardboard, clothes, glass, metal, paper, plastic, shoes, trash) the model achieves an high training and validation accuracy of 0.9995%, 0.9858% respectively , accompanied by high precision, recall and F1-score validation a test values. These results highlight the potential of the approach to automate waste sorting, reduce energy consumption and improve the efficiency of recycling systems in smart urban environments. The proposed model thus represents a promising solution to support sustainable waste management initiative.