Hyperspectral Image Acquisition and Calibration with Application to Skin Detection Systems

Abstract

Hyperspectral imaging provides access to characteristic material spectra in a way not possible with standard trichromatic imaging, although relatively little has been done to exploit these data outside of remote sensing applications. In our e orts to build a hyperspectral skin detection system, several new problems were identified that had not previously been adequately addressed, suggesting that routine use of spectral imaging techniques in terrestrial (earth-based) applications is non-trivial. Previously inconsequential artefacts, such as shadowing, become extremely significant when imposing the requirement for quantitative data. A large part of this research project has been the development of methods to enable accurate and reliable processing of hyperspectral images in the terrestrial setting. Unlike broadband trichromatic cameras, imaging spectrometers are inherently susceptible to chromatic aberration given their operational spectral range and narrowband resolution. We developed an imaging spectrometer that produced optimally focused images at each wavelength band. Additionally, the near infrared (NIR) Liquid Crystal Tuneable Filter (LCTF) used on this device was found to contain extraneous transmission sidelobes at a subset of tuneable wavelengths, rendering measurements at these wavelengths highly inaccurate. We present a novel and effective method to recover accurate material reflectance using inverse methods. Further to these equipment idiosyncrasies, shadowing and illumination variations, which present significant challenges, but are often expediently ignored. A revised linear mixture model (LMM) was developed to account for these variations and a collection of techniques capable of removing these artefacts are presented. The efficacy and utility of these techniques are demonstrated on real-world hyperspectral data. Finally, a database of hyperspectral human skin images was constructed to assist the development of skin detection techniques, and the challenges faced during its construction are discussed. Preliminary analysis of these data has tied specific skin regions (e.g. nose, ears, lips and knees) to distinct endmembers and reflectance histogram modes, and that the commonly-used Gaussian random process model does not sufficiently characterise this spatial heterogeneity of skin. While a successful method for skin detection was not established, the practical tools provided by this research make significant contributions towards realising this goal.