Metal-Dielectric Object Classification by Combining Polarization Property and Surface Spectral Reflectance

We propose a method for automatically classifying multiple objects in a natural scene into metal or dielectric. We utilize polarization property in order to classify the objects into metal and dielectric, and surface-spectral reflectance in order to segment the scene image into different object surface regions. An imaging system is developed using a liquid crystal tunable filter for capturing both polarization and spectral images simultaneously. Our classification algorithm consists of three stages; (1) highlight detection based on luminance threshold, (2) material classification based on the spatial distribution of the degree of polarization at the highlight area, and (3) image segmentation based on illuminant-invariant representation of the spectral reflectance. The feasibility of the proposed method is examined in detail in experiments using real-world objects.