Deep learning architectures for hyperspectral imaging applications

Abstract

A typical consumer camera captures three bands of light corresponding to red, green and blue colors. A hyperspectral imager captures dozens or even hundreds of bands. A depth sensing camera captures the distance to the target at each pixel. Imaging spectra and depth opens new possibilities for extracting information about the target, and these kind of imagers have already been used in applications in agriculture, astronomy, forestry, medical imaging and other industries. The captured high-dimensional data volumes are large, and extracting meaningful information from them requires advanced and efficient processing methods. Previously, the need for expert manual work has limited the utilization of data in large scale. This research introduces neural network models for solving these problems in a few case applications. It also demonstrates hyperspectral measurement methods: one for radiance approximation and another for angular reflectance measurement by combining a depth camera with a hyperspectral camera.