Combined Masking Strategies for the Identification of Black Plastic Samples Through Mid-Wavelength Infrared Hyperspectral Imaging

Black plastic materials pose a significant challenge in industrial sorting and recycling due to their invisibility to conventional Near-Infrared (NIR) detection systems. Recently, Mid-Wavelength Infrared Hyperspectral Imaging (MWIR HSI) has emerged as a promising technology to overcome this limitation by enabling spectral discrimination of black polymers. However, the low emission intensity in the MWIR range and the presence of surface reflections and complex geometries hinder the effective spatial identification of such materials. This study proposes a dual masking strategy that relies solely on MWIR hyperspectral data, without support from external sensors, to achieve accurate spatial localization and classification of black plastic samples. The first mask employs a strict threshold-based selection on specific spectral bands to retain only reliable pixels for classification. The second mask applies K-means clustering to Lab∗ and RGB reconstructions of the hyperspectral image to recover spatial information and define object boundaries. The combination of these two masks enables robust, pixel-wise classification while preserving spatial completeness, making the approach suitable for in-line applications. Results demonstrate that the method effectively identifies black plastic samples down to sizes of 1-2 cm, with inter-sample distances above 0.5 cm. Below these thresholds, higher-resolution imaging or pre-trained models may be required.