VISPec‐DNet: Multi‐Modal Decision Fusion of Hyperspectral and Red–Green–Blue Images for Plastic Waste Detection and Sorting

The global plastic waste crisis is exacerbated by low recycling rates and the diminished value of recycled materials, largely resulting from inaccurate sorting of plastics based on the polymer type. This study investigates the fusion of red–green–blue (RGB) and hyperspectral near‐infrared (NIR) imaging of waste plastic objects to improve the accuracy of plastic waste sorting. We propose a novel decision‐level image fusion framework, VISPec‐DNet, which integrates complementary information from both modalities to improve the detection and sorting accuracy. The system is capable of detecting 6 waste plastic categories across high‐density polyethylene (HDPE), polyethylene terephthalate (PET), and polypropylene (PP) polymer types. The system achieves an impressive 99.1% accuracy in polymer classification, with the top‐performing model reaching a detection accuracy of 97.4%. Object detection is performed using You Only Look Once (YOLO) models (versions 5, 8, 11, and 12), while Light Gradient‐Boosting Machine (LightGBM) is employed for pixel‐level classification of hyperspectral and multispectral data. Results demonstrate the potential of multimodal image fusion to significantly advance plastic waste sorting, offering a scalable solution for improving recycling efficiency in industrial applications.