Waste Classification Using Convolutional Neural Network with ShuffleNetV2 Architecture

Abstract

Internet of Things (IoT)-based smart waste sorting systems require classification algorithms that are not only accurate but also efficient in resource utilization. However, the majority of previous studies tend to focus on heavy-architecture Deep Learning models (such as VGG or ResNet) that burden edge devices, or utilize lightweight models that are limited to a few class categories. This research contributes to filling this gap by evaluating the effectiveness of the ShuffleNetV2 architecture, a lightweight CNN that optimizes Memory Access Cost (MAC), to classify 9 complex waste categories (Biological, Clothes, Glass, Plastic, Shoes, Battery, Metal, Paper, Cardboard). The research dataset was compiled through the curation and combination of three public Kaggle repositories, which were reprocessed using Roboflow, producing 19,906 augmented images to ensure visual domain variance. Empirical evaluation results show that the model achieved an accuracy of 94% with an average F1-Score of 0.93. The efficiency advantage is evidenced by the compact model size (4.99 MB) and low estimated computational load (0.30 GFLOPs) compared to conventional models. These findings indicate that ShuffleNetV2 offers an optimal performance trade-off, making it a feasible solution for implementation on mobile devices and low-power embedded systems.