Ensembling Lightweight Models for Intelligent Waste Recognition Based on Degraded Images

Traditional waste recycling predominantly depended on labor-intensive manual waste classification, a solution that neither efficient nor time-saving. In a bid to enhance the efficacy of recycling, relevant researchers incorporate computer vision and deep learning methodologies to facilitate rapid waste sorting. However, existing waste recognition models can not effectively recognize waste based on degraded images, nor can they be flexibly applied to different application platforms. Consequently, we introduce an intelligent approach grounded in ensemble learning for waste recognition in intricate scenarios. Initially, we construct a new waste image degradation dataset and simulate waste recycling scenarios under different illumination and motion conditions using synthetic degradation methods. Subsequently, we investigate various image recognition networks and ensemble learning methods for waste recognition, and introduce a novel query ensemble learning approach aimed at striking an optimal equilibrium between model parameters and recognition accuracy. Furthermore, we conduct relevant comparative analyses to substantiate the advantages of our methodology. The empirical outcomes indicate that our solution outperforms other deep learning models and achieves satisfactory results for accurate waste recognition based on degraded images.