Robust and adaptive subspace learning for fast hyperspectral image denoising

Hyperspectral image (HSI) denoising can be implemented in a low-dimensional spectral subspace to utilize the high correlations across different bands and reduce the imposed computational burden, and subspace learning is a key strategy for achieving improved denoising performance. In the previously developed subspace-based HSI denoising methods, the subspace learning task is usually seriously deteriorated by noise or suffers from high computational costs, and subspace dimensionality determination has not been carefully and specifically studied for subspace denoising. In this paper, we propose a highly efficient HSI denoising method by developing a robust and adaptive subspace learning algorithm. Specifically, we introduce a semisequentially truncated higher-order singular value decomposition scheme for jointly performing basic estimation and learning a robust subspace basis, and the subspace dimensionality is adaptively detected by designing an appropriate information-theoretic criterion. Then we conduct noise reduction by applying a bandwise denoiser in the subspace since various subspace bands generally have significantly different noise levels. The applied bandwise denoiser can be a single-band deep learning method, which do not require hyperspectral training data when the hyperspectral training data are few in number and more difficult to obtain than single-band images. The proposed HSI denoising method is very fast and effective. Experimental results demonstrate that the proposed method can achieve state-of-the-art denoising performance.