Hyperspectral Imagery Denoising Based on Oblique Subspace Projection

This paper presents a hyperspectral imagery denoising algorithm based on oblique subspace projection (DOBSP), which considers the correlation between noise and signal. The algorithm first estimates the signal and noise through segmentation Gaussian filtering which can reduce more influence of the image texture than traditional Gaussian filtering. Then, signal and noise estimates are fed into principal component analysis (PCA) to identify their respective subspace basis vectors. Finally, these basis vectors are used to compute matrices of oblique subspace projection (OBSP), and the signal and noise are extracted from the original image through OBSP. We assessed the DOBSP algorithm using both simulated and real Hyperion images. The orthogonal subspace projection (OSP) which assumes that noise is independent on signal and the subspace-based striping noise reduction (SBSR) algorithm which uses polynomial model to describe the relationship between noise and signal were introduced for comparison. Compared with signal and noise results by OSP and SBSR, both signal and noise extracted by DOBSP on the simulated image are closer to the original simulated signal and noise, and the noise image obtained by DOBSP on the Hyperion image has fewer textures.