NL3DLogTNN: An Effective Hyperspectral Image Denoising Method Combined Non-Local Self-Similarity and Low-Fibered- Rank Regularization

Abstract

Hyperspectral image denoising is an important research topic in the field of remote sensing image processing. Recently, methods based on non-local low-rank tensor approximation have gained widespread attention towing to their ability to fully exploit non-local self-similarity. However, existing non-local low-rank tensor approximation methods fall short in capturing the correlations between various modes in hyperspectral images, thus failing to achieve the optimal approximation. To solve this issue, a novel three-directional log-based tensor nuclear norm (3DLogTNN)–based non-local hyperspectral image denoising model NL3DLogTNN is proposed. The correlation between the various modes of the model was obtained by performing TNN decomposition in three directions on the extracted non-local comparable blocks, better capturing the global low-rank property of the image. To effectively solve the proposed NL3DLogTNN model, we developed an approximate alternating direction method of multipliers (ADMM)-based methodology and offered a thorough numerical convergence proof. Extensive experiments are conducted on hyperspectral image datasets with simulated noise and real-world noise, which demonstrated that the proposed NL3DLogTNN model outperforms state-of-the-art methods in terms of quantitative and visual performance evaluation.

Keywords

• Hyperspectral image denoising
• non-local self-similarity
• 3DLogTNN decomposition