The rapid increase in global plastic waste necessitates efficient recycling, yet current manual and me chanical sorting methods suffer from low efficiency and poor working conditions. In particular, transparent PET bottles, a high-value resource, present significant detection challenges due to their lack of visual texture and severe deformation during collection, leading to critical false detections. To address this issue, we propose an improved YOLOv11 model integrated with a Large Separable Kernel Attention (LSKA) module at the leve of feature fusion stage. By decomposing large kernels into horizontal and vertical components, the LSKA mechanism secures a broad receptive field without excessive computational cost, enabling the model to effectively capture the global shape and context of transparent objects even under severe occlusion. Experimental results show that the proposed model achieves a mean Average Precision (mAP50) of 97.2%, surpassing the baseline by 0.6% under the 5-fold cross- validation experimental settings. The proposed model exhibits robustness and reliability, confirming its potential for accurate automated plastic sorting systems where high-quality classification is critical.
Lee et al. (Sat,) studied this question.