Dual-stream context aggregation network for waste sorting

Abstract Traditional manual waste sorting has long been constrained by bottlenecks such as low efficiency, high cost, and health risks, which has driven computer vision-based automated sorting technology to become a vital research direction. However, existing methods often struggle to simultaneously capture both global semantic information and extract local detailed features in waste sorting tasks. To address the above issues, this paper proposes a d ual- s tream c ontext a ggregation net work (DS-CANet). The first important module in the network is global encoding-calibration feature extraction module that relies on the vision transformer (ViT) to model long-range semantic dependencies in images and incorporates an efficient channel attention (ECA) mechanism to adaptively enhance global features. Another key module is local fine-grained enhancement module integrated with an attention-guided deep fusion mechanism, which utilizes the grouped convolutions of ResNeXt50 to extract detailed features such as texture and edges, and employs the ECA module for channel calibration to strengthen the model’s ability to distinguish easily confused categories. Through a cascaded strategy of channel concatenation—ECA recalibration, it achieves organic integration of global and local features, thereby further optimizing the discriminative performance of the features. Experimental results on the public waste image dataset TrashNet show that the proposed method achieves a classification accuracy of 96.09%, which is higher than several compared baseline models on this benchmark. These results demonstrate the effectiveness of DS-CANet for automated waste sorting on the TrashNet dataset, while broader validation on more diverse datasets remains a direction for future work.