Robust Registration in Integrated Hyperspectral Imaging (Robuuste registratie in geïntegreerde, hyperspectrale beeldopname)
Robust Registration in Integrated Hyperspectral Imaging
Mainali, Pradip; S0209294
Hyperspectral Imaging (HSI) combines spectroscopy with imaging, enabling the identification of the chemical signature of scene materials and objects without physical contact. The potential of HSI is enormous. Although originally applied in the professional market of remote sensing, researchers are striving to develop low-cost, miniaturized and high-speed hyperspectral cameras for low-cost consumer applications such as food inspection, counterfeit detection, etc. in various, portable form factors. Hyperspectral data is collected in a 3D hyperspectral image cube/data structure with two spatial (horizontal) and one spectral (vertical) dimension. The process of stacking spectral images acquired from a 2D hyperspectral camera to form a 3D hyperspectral image cube, generally suffers from misalignment/misregistration, preventing the accurate reading of the scenes chemical-spectral signatures. To achieve a perfect hyperspectral cube, images have to be aligned with sub-pixel accuracy. The main objective of the thesis herein is two-fold. First, we develop a robust registration algorithm, immune against the severe hyperspectral image artifacts such as blur, noise and spectral reflectance variations. To meet this research objective, this thesis proposes two novel feature detection algorithms using the cosine modulated Gaussian and the tenth order Gaussian derivative filters. The proposed feature detection algorithms show a two times higher immunity against the image artifacts as compared to state-of-the-art methods. These feature detection algorithms also double the registration accuracy. Second, we also propose a low complexity implementation approach. The thesis proposes a fixed-length approximation filter, providing constant-time, low-cost image filtering implementations. Finally, uniquely combining the proposed advanced feature detection algorithms with RANdom SAmple Consensus's (RANSAC) model parameter estimation, the proposed registration algorithm reduces the registration error from hundreds of pixels in the raw hyperspectral image cube down to half-pixel registration inaccuracy, without the need of costly calibration, deblurring and/or denoising techniques in hardware or software. This eventually paves the way to a low-cost, yet reliable hardware/software co-designed hyperspectral imaging system.