Abstract:The efficient fine sorting of domestic waste plastics is crucial for enabling high-value recycling, improving the quality of recycled materials, and enhancing overall energy-conversion efficiency. This study addresses a common limitation of single-sensor systems—the difficulty of simultaneously identifying both material type and color, which constrains sorting accuracy. To overcome this limitation, we propose a multi-feature fusion identification method that integrates RGB vision with near-infrared (NIR) spectroscopy. Polypropylene (PP) and polyethylene terephthalate (PET) were selected as representative materials for experiments. A data-acquisition platform was built to synchronously capture 640×640-pixel color images and hyperspectral data in the 935.9–1722.5 nm range. Spectral data were preprocessed by black-white frame correction, Savitzky-Golay filtering, and standard normal variate (SNV) transformation to remove noise. A feature-band selection strategy based on spectral extrema reduced 204-dimensional spectral data to a key band interval (1641.4–1687.2 nm). We developed a dual-branch fusion model that combines a Yolact instance-segmentation network (ResNet50-FPN backbone) in the RGB branch to extract color and contour features, with a support vector machine (SVM) in the NIR branch to classify material types using the selected spectral bands. Decision-level fusion was used to integrate the two branches′ outputs. Experimental results show that the proposed method achieves a precision of 97%, a recall of 97%, and an overall accuracy of 96% when classifying six categories of waste plastics (transparent, white, and colored PP and PET). These results represent an overall improvement of 6–7 percentage points across the evaluated metrics compared with an RGB-only baseline (Yolact). The feature-band selection strategy compresses spectral dimensionality by 99.2%, effectively reducing model complexity and the risk of overfitting. The proposed method provides a general and adaptable framework for algorithm development in automated sorting equipment. Validated on real-world production waste samples, the fusion architecture can be extended to identify other domestic waste plastics (e.g., PS, PVC, HDPE), thereby contributing to improved energy efficiency and environmental benefits through precise material and color sorting. We note a limitation: black plastics, which strongly absorb NIR light, were not included in this study. Future work will consider integrating mid-infrared spectroscopy or X-ray sensing for black-plastic identification, evaluating lightweight backbone networks for real-time processing, exploring joint optimization of band selection and feature extraction, and investigating adaptive dynamic fusion strategies to enhance robustness in complex scenarios. The framework's adaptability and the demonstrated performance gains underscore its potential for practical industrial deployment in waste management systems. Close-