Denoising Hyperspectral Images Using Interband Correlation

This paper proposes a method for self-supervised denoising of hyperspectral images (HSI). Learning-based HSI denoising has limited success because it is often difficult to collect large datasets. To address the difficulty, we proposed a self-supervised approach, in which we exploit the self-similarity that hyperspectral images inherently have among adjacent bands. Unlike other hyperspectral denoising methods based on training a neural network, we do not need any training dataset, and it is possible to restore an image by the CNN trained only by a single noisy input. We demonstrate the validity of our method through some experiments and show that our approach has better or comparable characteristics than conventional methods.