Hyperspectral Denoising Using Asymmetric Noise Modeling Deep Image Prior

Deep image prior (DIP) is a powerful technique for restoration that leverages an untrained network as handcrafted prior. DIP can also be used hyperspectral (HSI) denoising tasks and has achieved impressive performance. Recent works further incorporate different regularization terms to enhance the performance of successfully show notable improvements. However, most DIP-based methods HSI rarely consider distribution complicated mixed noise. In this paper, we propose asymmetric Laplace noise modeling deep (ALDIP) removal. Based on observation real-world exhibits heavy-tailed properties, model each band using distribution. Furthermore, in order fully exploit spatial–spectral correlation, ALDIP-SSTV, which combines ALDIP with total variation (SSTV) term preserve more information. Experiments both synthetic data demonstrate ALDIP-SSTV outperform state-of-the-art methods.