Hyperspectral Image Denoising Using 3-D Geometrical Kernel With Local Similarity Prior

The majority of existing hyperspectral (HS) image denoising methods exploit local similarity in HS images by rearranging them into the matrix or vector forms. As the typical 3-D data, the inherent spatial and spectral properties in HS images should be simultaneously explored for denoising. Therefore, a 3-D geometrical kernel (3DGK) is developed in this article to describe the local structure. The proposed method assumes that the pixel can be represented by other pixels within a 3-D block efficiently owing to the local similarity with adjacent positions. Then, the HS image is modeled by the 3-D kernel regression with L1-norm constraint, in which the local similarity is captured by the proposed 3DGK. To efficiently compute the parameters in 3DGK, geometrical structures, such as scale, shape, and orientation in the 3-D block, are estimated from the gradient information approximately. Finally, the noises are effectively removed while preserving the structures in HS images. Moreover, experimental results on simulated and real datasets demonstrate that the performance of 3DGK is better than those of the methods based on local similarity prior.