Abstract
Garbage sorting contributes to resource recycling, mitigates environmental pollution, and promotes sustainable development. However, traditional garbage sorting methods typically require significant human labor and time resources, underscoring the necessity for automated solutions. While the convolutional neural network (CNN) has achieved significant success in garbage sorting, existing models still suffer from low computational efficiency and accuracy. In light of these challenges, this study proposes the smart recycle sort network (SRS-Net), a lightweight model with attention mechanism aimed at enhancing the efficiency and accuracy of garbage sorting processes. Lightweight networks reduce computational complexity and parameters, improving garbage sorting efficiency. We improve the ShuffleNet unit and introduce the lightweight shuffle attention module (LSAM) as the primary module of SRS-Net. On one hand, given the diverse shapes and sizes of garbage items, we replace the depthwise convolution (DWConv) in the ShuffleNet unit with heterogeneous kernel-based convolutions (HetConv) to accommodate this diversity. On the other hand, to better focus on important features of garbage images, we introduce shuffle attention (SA), a channel-spatial attention mechanism that considers the importance of inter-channel relationships and spatial positions. To validate the performance of SRS-Net, we conduct comparative experiments on two datasets, TrashNet and garbage dataset. The experimental results demonstrate that SRS-Net achieves an accuracy of 90.02% on TrashNet and 91.52% on garbage dataset, with FLOPs of 1262.0 M and Params of 9.6902 M. Our approach effectively facilitates automated garbage sorting and resource recycling.