Hyperspectral Imaging and Analysis for Sparse Reconstruction and Recognition

Hyperspectral imaging, also known as imaging spectroscopy, captures a data cube of a scene in two spatial and one spectral dimension. Hyperspectral image analysis refers to the operations which lead to quantitative and qualitative characterization of a hyperspectral image. This thesis contributes to hyperspectral imaging and analysis methods at multiple levels. In a tunable filter based hyperspectral imaging system, the recovery of spectral reflectance is a challenging task due to limiting filter transmission, illumination bias and band misalignment. This thesis proposes a hyperspectral imaging technique which adaptively recovers spectral reflectance from raw hyperspectral images captured by automatic exposure adjustment. A spectrally invariant self similarity feature is presented for cross spectral hyperspectral band alignment. Extensive experiments on an in-house developed multi-illuminant hyperspectral image database show a significant reduction in the mean recovery error. The huge spectral dimension of hyperspectral images is a bottleneck for efficient and accurate hyperspectral image analysis. This thesis proposes spectral dimensionality reduction techniques from the perspective of spectral only, and spatio-spectral information preservation. The proposed Joint Sparse PCA selects bands from spectral only data where pixels have no spatial relationship. The joint sparsity constraint is introduced in the PCA regression formulation for band selection. Application to clustering of ink spectral responses is demonstrated for forensic document analysis. Experiments on an in-house developed writing ink hyperspectral image database prove that a higher ink mismatch detection accuracy can be achieved using relatively fewer bands by the proposed band selection method. Joint Group Sparse PCA is proposed for band selection from spatio-spectral data where pixels are spatially related. The additional group sparsity takes the spatial context into account for band selection. Application to compressed hyperspectral imaging is demonstrated where a test hyperspectral image cube can be reconstructed by sensing only a sparse selection of bands. Experiments on four hyperspectral image datasets including an in-house developed face database verify that the lowest reconstruction error and the highest recognition accuracy is achieved by the proposed compressed sensing technique. An application of the proposed band selection is also presented in an end-to-end framework of hyperspectral palmprint recognition. An efficient representation and binary encoding technique is proposed for selected bands of hyperspectral palmprint which outperforms state-of-the-art in terms of equal error rates on three databases.