An integrated detection-semantic fusion and near-infrared system for food-delivery packaging waste

Abstract

The rapid increase in food-delivery packaging waste poses major problems to sustainable waste management. Current recycling techniques are limited by the intricate spectral properties of multi-material packaging and challenges in detection in unstructured settings. Traditional vision-based methodologies are impeded by restricted cross-modal integration and insufficient multi-scale analysis, obstructing the advancement of a closed-loop circular economy. To address these limitations, this study introduces FDPWaste, the first dataset that combines annotated images with near-infrared spectra of five plastics, capturing the complexity of real recycling contexts. A new detection model, CFD-YOLO, is further developed with enhanced feature attention, achieving 93.3 % ${\mathrm{mAP}}_{50}$. Furthermore, the edge-aware segmentation network, ECA-UNet, generates enhanced color features that are utilized by a PSO-SVM material classifier, achieving an accuracy of 97.1 %. These components are integrated into an automated sorting system with robotic control and real-time decision-making. Overall, the framework provides practical tools for advancing circular recycling of food-delivery packaging.