Denoising Hyperspectral Image With Non-i.i.d. Noise Structure.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral image (HSI) denoising is crucial for improving data quality in remote sensing.
• Current denoising methods often assume independent and identically distributed (i.i.d.) noise, which is unrealistic for natural HSIs.
• This assumption leads to reduced robustness and performance in real-world applications.

Purpose of the Study:

• To address the limitations of existing HSI denoising techniques.
• To develop a novel HSI denoising model that accounts for complex, non-i.i.d. noise characteristics.
• To improve the robustness and accuracy of HSI denoising.

Main Methods:

• Proposed a novel noise modeling strategy using a non-i.i.d. mixture of Gaussians (NMoGs) assumption for HSIs.
• Integrated the NMoGs noise model into a low-rank matrix factorization (LRMF) framework, resulting in the NMoG-LRMF model.
• Employed a variational Bayes algorithm for posterior inference within the Bayesian framework.

Main Results:

• The proposed NMoG-LRMF model effectively captures the complex statistical structures of noise in natural HSIs.
• Experimental results on synthetic and real noisy HSIs demonstrate superior denoising performance compared to state-of-the-art methods.
• The method shows enhanced robustness in practical HSI denoising scenarios.

Conclusions:

• The NMoGs noise assumption provides a more accurate representation of noise in natural HSIs.
• The NMoG-LRMF model offers a significant advancement in HSI denoising, outperforming existing techniques.
• This approach enhances the reliability and utility of hyperspectral data for various applications.