Development of low rank based algorithms for denoising and classification of hyperspectral images

Hyperspectral images comprise several bands (often over hundreds of bands) where each band caters to a continuous band of wavelengths in the electromagnetic spectrum. These bands also include regions in the electromagnetic spectrum that lie outside the visible range of human vision and hence HSIs are capable of conveying information which cannot be perceived from ordinary color/grayscale images. HSIs can be corrupted by various kinds of noise viz. Gaussian noise, Poisson noise, dead lines and stripes etc. This affects not just the visual appearance but the applications of these images too. The objective of my research is to exploit the correlations present in the underlying clean data in order to restore severely degraded hyperspectral images. Some results from the AVIRIS Indian Pines hyperspectral datset are shown below.

(a) Band 103

(b) Restored

(c) Band 105

(d) Restored