Cost-Sensitive Lightweight Waste Classification Using Modified EfficientNetB0 and Parallel CNNs for Resource-Constrained Environments

With the rapid expansion of industrialization and the rise of smart cities, the need for efficient waste collection, sorting, and management has become increasingly urgent. Effective recycling hinges on precisely identifying and processing waste materials, crucial for reducing pollution and promoting environmental sustainability. Deep learning (DL) techniques have recently been adopted to enhance various aspects of waste management. Yet, much of the research has relied on high computational resources and often overlooked the class imbalance challenge. This study addresses these limitations by introducing an innovative and lightweight waste classification model based on cost-sensitive learning, integrated with a modified EfficientNetB0 architecture. The model incorporates Convolutional Block Attention Modules (CBAM), which prioritize significant features while minimizing irrelevant ones, and merge parallel Convolutional Neural Network (CNN) branches to extract more complex and refined features from waste data. The proposed model, when tested on the publicly available TrashNet Dataset, is found to be effective in terms of accuracy, training time, inference time, and the total number of parameters compared to other state-of-the-art research, contributing to advancements in waste management practices. Therefore, this research offers a versatile and accurate approach to waste classification, particularly well-suited for remote locations with limited computational resources, marking a significant advancement in global waste management practices.