Hybrid Attention Transformer enabled intelligent recycling-oriented sorting of multi-type beverage containers under degraded waste conditions for circular economy systems

The rapid increase in beverage container waste has created significant challenges for automated recycling systems, particularly under real-world conditions involving contamination, deformation, illumination variation, and partial occlusion. Existing convolutional neural network (CNN)-based waste sorting systems often experience reduced performance in degraded waste-stream environments and lack material-oriented recycling intelligence. To address these limitations, this study proposes a novel Hybrid Attention Transformer (HAT) framework for intelligent recycling-oriented classification and sorting of multi-type beverage containers. The proposed framework integrates patch-based embedding, multi-head spatial self-attention, channel-wise attention refinement, and a dual-head multi-task decoder for simultaneous object-level and material-level prediction. A new Waste Condition Level (WCL) variable is introduced to model clean, moderately soiled, and highly degraded waste conditions for robustness evaluation. Experiments were conducted using 15,607 annotated beverage container images under extensive augmentation settings. The proposed HAT model was implemented in PyTorch and optimized using AdamW with multi-task weighted loss functions. Experimental results demonstrated that the proposed framework achieved 99.8% classification accuracy with 97.2% precision, 94% recall, and 94% F1-score, outperforming existing CNN- and YOLO-based approaches under degraded waste conditions. The model also maintained stable confidence and robust material-level prediction across all WCL levels. The findings confirm that the proposed HAT framework provides an effective, scalable, and deployment-oriented solution for intelligent waste sorting and circular economy applications.