Robust near-infrared-based plastic classification with relative spectral similarity pattern

Abstract

Sensor-based material flow characterization techniques, particularly hyperspectral imaging in the near-infrared (NIR) range, can recognize materials quickly, accurately, and economically. When identifying materials using NIR hyperspectral imaging, extracting influential features from high-dimensional wavelength information is essential for effective recognition. However, spectral noise from the rough and contaminated surfaces of objects (especially un-shredded waste) degrades the feature-extraction performance, which in turn deteriorates the material classification performance. In this study, we propose a real-time feature-extraction method, named relative spectral similarity pattern color mapping (RSSPCM), to robustly classify materials in noisy environments, such as plastic waste sorting facilities. RSSPCM compares relative intra- and inter-class spectral similarity patterns, instead of individual similarity, to class-representative spectra alone. Recognition targets have similar chemical makeups that are applied to feature extraction as an intra-class similarity ratio. The proposed model is robust owing to the remaining relative similarity trends found in a contaminated spectrum. We evaluated the effectiveness of the proposed method using noisy samples obtained from a waste-management facility. The results were compared with two spectral groups obtained at different noise levels. Both results showed high accuracy as there was an increased number of true positives for low-reflectance regions. The average F1-score values were 0.99 and 0.96 for low- and high-noise sets, respectively. Furthermore, the proposed method showed minimal F1-score variations between classes (standard deviation of 0.026 for the high-noise set).