Lightweight Heterogeneous Convolutional Neural Network for Trash Classification

Trash classification plays a crucial role in effectively reducing environmental pollution, maximizing resource recycling, and lowering the cost of trash disposal, with profound implications for achieving sustainable development and constructing a clean and healthy society. However, current trash classification heavily relies on traditional manual classification, facing challenges in meeting the growing demands for efficiency and accuracy. The application of deep learning technologies can significantly enhance the accuracy and efficiency of trash classification, enabling automated trash classification. In this study, we constructed a lightweight heterogeneous convolutional neural network (LHN et) specifically designed for the classification of TrashN et. By employing heterogeneous kernel-based convolutions (HetConv) to extract features from the trash images and enhancing feature interaction through channel shuffle, LHNet demonstrated outstanding performance in comparison with multiple models. It achieved an accuracy of 88.13%, a recall of 86.90%, a precision of 85.10%, a kappa coefficient of 85.23%, and an fl-score of 85.77%. Simultaneously, LHNet exhibited a computational complexity (FLOPs) of 1291.4688M and a parameter count (Params) of 9.6902M. LHNet showed significant improvements in accuracy and efficiency, providing a robust approach for achieving automated and efficient trash classification. This not only addresses the escalating demands for trash processing but also contributes valuable technological support to promote sustainable development and the establishment of a clean and healthy society.