An Approach to Automation in Smart Waste Classification Based on Machine Learning for Embedded Devices

Abstract

This study aims to develop automated smart waste classification systems using machine learning deployed on resource-constrained hardware, such as Raspberry Pi, with the goal of integrating these models into the initial stage of waste sorting—that is, the trash bin. This study proposes two models: ConvNeXt and an improved CNN model, which is referred to as PrQ-CNN. The improved PrQ-CNN model utilizes pruning and quantization techniques to reduce its size. Both models were trained on a dataset comprising two classes: Non-recyclable and Recyclable. The data were split into a training set and a testing set with a 90:10 ratio. The models were integrated into a Raspberry Pi embedded computer to compare metrics such as model size, inference time, accuracy, missed detection rate (MDR), and false discovery rate (FDR). Experimental results show that the PrQ-CNN model achieved an accuracy of 92%, while its size was 12 times smaller than that of the original model before pruning and quantization. The ConvNeXt model achieved an accuracy of 98%. The results from the improved PrQ-CNN and ConvNeXt models were compared with VGG16, ResNet, EfficientNet, and MobileNet.