Hyperspectral Data Acquisition and Its Application for Face Recognition

Current face recognition systems are rife with serious challenges in uncontrolled conditions: e.g., unrestrained lighting, pose variations, accessories, etc. Hyperspectral imaging (HI) is typically employed to counter many of those challenges, by incorporating the spectral information within different bands. Although numerous methods based on hyperspectral imaging have been developed for face recognition with promising results, three fundamental challenges remain: 1) low signal to noise ratios and low intensity values in the bands of the hyperspectral image specifically near blue bands; 2) high dimensionality of hyperspectral data; and 3) inter-band misalignment (IBM) correlated with subject motion during data acquisition.

This dissertation concentrates mainly on addressing the aforementioned challenges in HI. First, to address low quality of the bands of the hyperspectral image, we utilize a custom light source that has more radiant power at shorter wavelengths and properly adjust camera exposure times corresponding to lower transmittance of the filter and lower radiant power of our light source.

Second, the high dimensionality of spectral data imposes limitations on numerical analysis. As such, there is an emerging demand for robust data compression techniques with lows of less relevant information to manage real spectral data. To cope with these challenging problems, we describe a reduced-order data modeling technique based on local proper orthogonal decomposition in order to compute low-dimensional models by projecting high-dimensional clusters onto subspaces spanned by local reduced-order bases.

Third, we investigate 11 leading alignment approaches to address IBM correlated with subject motion during data acquisition. To overcome the limitations of the considered alignment approaches, we propose an accurate alignment approach ( A₃) by incorporating the strengths of point correspondence and a low-rank model. In addition, we develop two qualitative prediction models to assess the alignment quality of hyperspectral images in determining improved alignment among the conducted alignment approaches. Finally, we show that the proposed alignment approach leads to promising improvement on face recognition performance of a probabilistic linear discriminant analysis approach.