Real-time polymer type detection from hyperspectral images of mixed plastic waste using deep learning for automated segregation

Abstract

Efficient plastic waste management relies on the accurate identification and sorting of materials, which are essential for advancing recycling and upcycling efforts. While laser-induced breakdown spectroscopy is commonly used for polymer detection, hyperspectral imaging (HSI) offers a non-destructive alternative with enhanced spectral resolution, enabling differentiation of visually similar materials. However, HSI’s high-dimensional data poses challenges for real-time processing. This study aims to develop a real-time deep learning-based detection model inspired by YOLOv2 (You Only Look Once). The proposed model classifies and localizes four common polymer types, namely; polypropylene, polyethylene terephthalate, high-density polyethylene, and polystyrene in mixed plastic waste streams using HSI data (900–1700 nm). By utilizing the YOLOv2 framework to automatically learn and extract discriminative spectral–spatial features, the model effectively processes the spectral content of HSI data for rapid and accurate inference. Experimental results show a mean average precision of 94.5% at an Intersection over Union threshold of 0.50 and a processing time of 0.086 s per image, with an average of 23 polymers per image. Although YOLOv3 offered slightly higher accuracy, YOLOv2 was chosen for its superior speed-accuracy tradeoff, making it more suitable for real-time industrial applications. These findings demonstrate that integrating hyperspectral imaging with lightweight deep learning models can deliver scalable, high-speed, and high-accuracy solutions for plastic waste sorting—potentially enhancing recycling efficiency and supporting circular economy goals.