Integrating Spectral and Visual Data for Material Identification and Segmentation

Abstract

Accurate sorting and identification of materials is crucial in real-world applications such as recycling, manufacturing, and quality control, yet traditional classification based solely on visual information often fails to distinguish materials with similar appearances but different properties. In this study, we investigate the identification and segmentation of materials by combining spectral and visual information. Unlike typical 2D spectral images, our data consist of one-dimensional 18-channel spectral signals per object, complemented by high-quality object masks produced by the SAM 2 model. We evaluate both analytical methods, based on distance metrics to reference spectra, and machine learning approaches capable of modeling more complex distributions.