Source-linked AI summary
Non-local Meets Global: An Integrated Paradigm for Hyperspectral Denoising
Wei He, Quanming Yao, Chao Li, Naoto Yokoya, Qibin Zhao
TL;DR
HSI denoising must scale to increasing spectral dimensionality while exploiting both spatial non-local similarity and global spectral low-rank structure. The paper introduces a unified iterative paradigm that learns a low-dimensional basis and reduced image, refines them through non-local denoising and iteration regularization, and reports superiority on simulated and real datasets.
Problem
Increasing spectral bands make non-local low-rank denoising substantially slower while providing little additional denoising benefit, and existing projection methods may retain noise.
Method
The method jointly learns an orthogonal spectral basis and reduced image, applies non-local low-rank denoising to the reduced image, and iteratively regularizes the basis using alternating minimization.
Results
The proposed method achieves the best results in simulated comparisons and the most appealing recovered images on real datasets.
Takeaways & Limitations
The unified spatial-spectral paradigm integrates spatial non-local similarity with global spectral low-rank structure while keeping non-local processing cost nearly unchanged as spectral bands increase.
Takeaways & Limitations
The objective is hard to optimize because it combines an orthogonal constraint on the basis with complex regularization on the reduced image.
Abstract
from arXiv · showhide
Non-local low-rank tensor approximation has been developed as a state-of-the-art method for hyperspectral image (HSI) denoising. Unfortunately, with more spectral bands for HSI, while the running time of these methods significantly increases, their denoising performance benefits little. In this paper, we claim that the HSI underlines a global spectral low-rank subspace, and the spectral subspaces of each full band patch groups should underlie this global low-rank subspace. This motivates us to propose a unified spatial-spectral paradigm for HSI denoising. As the new model is hard to optimize, we further propose an efficient algorithm for optimization, which is motivated by alternating minimization. This is done by first learning a low-dimensional projection and the related reduced image from the noisy HSI. Then, the non-local low-rank denoising and iterative regularization are developed to refine the reduced image and projection, respectively. Finally, experiments on synthetic and both real datasets demonstrate the superiority against the other state-of-the-arts HSI denoising methods.
1. Introduction
HSI denoising must address both spatial non-local similarity and global spectral low-rank structure, while existing approaches face scaling, projection-quality, and optimization challenges. The paper proposes an iterative unified paradigm and reports superior simulated and real-dataset results.
- HSI provides rich spectral information but is often corrupted by Gaussian noise, making denoising a fundamental preprocessing step.Hyperspectral imaging supports material discrimination and applications including remote sensing, medical diagnosis, face recognition, and quality control.
- As spectral bands increase, non-local full-band patch groups become larger, substantially increasing runtime while yielding limited denoising benefits.
- Spectral-only low-rank regularization lacks spatial regularization, while two-stage projection methods are sensitive to projection quality and may retain noise.
- The proposed paradigm jointly integrates spatial non-local similarity and global spectral low-rank structure by learning and iteratively updating an orthogonal basis matrix and reduced image.
- The optimization uses an efficient iterative algorithm inspired by alternating minimization because the model combines complex spectral constraints with spatial regularization.
- The method achieves the best comparison results in simulated experiments and the most appealing recovered images on real datasets.
2. Related work
Prior HSI denoising methods primarily exploit spatial non-local similarity or spectral low-rank structure, while two-stage combinations do not iteratively refine noisy subspaces. The related work motivates balancing these complementary properties.
- HSI denoising methods are commonly grouped into spatial non-local and spectral low-rank approaches, with prior knowledge needed because denoising is ill-posed.
- Non-local methods group full-band patches and apply low-rank tensor approximation, but prior work has explored many tensor models within a largely fixed pipeline.
- Some prior work reports weak spectral low-rank structure in non-local patch groups, leaving the balance between spectral low-rank modeling and spatial similarity unresolved.
- Global spectral low-rank methods exploit an intrinsic spectral subspace whose dimension is far smaller than the original spectral dimension.
- Two-stage methods first map HSI data into a low-dimensional spectral subspace and then apply spatial denoising, but they do not iteratively refine the subspace and it remains noise-corrupted.
3. The Proposed Approach
NGmeet jointly models global spectral low-rank structure and spatial non-local similarity by learning an orthogonal basis and reduced image, then iteratively refining both. Its alternating-minimization algorithm applies SVD-based spectral reduction, non-local denoising of the reduced image, and iterative regularization, while reducing the spectral cost of non-local processing.
- 3.1. Unified spatial-spectral paradigm: NGmeet represents the clean HSI as a reduced image multiplied along the spectral mode by an orthogonal basis with K ≪ B.The basis captures the common spectral subspace, while the reduced image has K spectral components.
- 3.1. Unified spatial-spectral paradigm: The paradigm jointly integrates global spectral low-rank structure with spatial non-local similarity through a non-local regularizer on the reduced image.The clean HSI is reconstructed as X = M∗×3 A∗.
- 3.2. Optimization: The objective is difficult to optimize because it combines an orthogonal constraint on A with complex regularization on M, so the paper uses alternating minimization.The algorithm approximately solves the objective through iterative updates.
- 3.2.1 Spectral low-rank denoising: SVD on each iterative input estimates the orthogonal basis Ai and reduced image M̄i, using the closed-form solution Ai = V and M̄i = fold3(US).Only an SVD of the folding matrix is required for this stage.
- 3.2.2 Spatial denoising via M: Similar patches are grouped in the reduced image, denoised by low-rank approximation, and reassembled to produce Mi and Xi = Mi ×3 Ai.The non-local regularizer requires grouping, tensor denoising, and assembly.
- 3.2.3 Iterative refinement: Iterative regularization updates the next noisy input and increases K by δ × i, enabling Ai+1 to capture more useful information with additional iterations.The basis is refined because its estimate is influenced by the noise intensity of the input image.
- 3.3. Complexity analysis: NGmeet adds O(MNB^2) SVD cost in stage A but makes stage B at least B/K times faster than comparable non-local denoising.Stage B has complexity O(Tn^2Kp^2), avoiding the full spectral dimension during non-local processing.
4. Experiments
Experiments evaluate NGmeet on simulated and real hyperspectral datasets using PSNR, SSIM, and SAM, alongside runtime, convergence, and parameter analyses. NGmeet achieves the best reported performance across nearly all simulated cases, strong visual quality on real data, and computational advantages as spectral dimensionality increases.
- Quantitative comparison: NGmeet achieves the best results in almost all simulated noise-level cases across the evaluated datasets.The evaluation uses PSNR, SSIM, and SAM; lower SAM indicates greater spectral similarity.
- Visual comparison: FastHyDe, LLRT, and NGmeet outperform the remaining comparison methods visually, while NGmeet avoids artifacts observed in FastHyDe and LLRT.The visual comparison uses CAVE-toy color images formed from bands 31, 11, and 6.
- Computational efficiency: NGmeet’s computational time is nearly unchanged as spectral number increases, while KBR and LLRT runtime increases linearly.Figure 4 compares computational time and SSIM on WDC with noise variance 100.
- Convergence: NGmeet converges to stable PSNR values very quickly at different noise levels on WDC.Convergence is assessed by plotting PSNR against iteration number.
- Real data experiments: NGmeet removes noise while preserving spectral details on Indian Pines and Urban real datasets, whereas LRTV changes color substantially and FastHyDe and LLRT retain stripes.Reference clean images are unavailable for these real-data experiments, so the comparison is visual.
- Parameter analysis: HySime initializes K reliably, and updating K improves performance while the selection of δ is robust.The reported HySime initializations are K = 7, 6, 6, 5 for noise variances 10, 30, 50, and 100, respectively.
5. Conclusion
The conclusion presents GNmeet as a unified spatial-spectral HSI denoising paradigm, solved by alternating minimization and supported by simulated and real-dataset experiments.
- GNmeet integrates spatial non-local similarity and global spectral low-rank structure through a low-dimensional orthogonal basis and reduced-image denoising.
- The proposed optimization uses an alternating minimization method with an iteration strategy to solve the GNmeet objective.
- Experiments on simulated and real datasets confirm the superiority of the proposed method.
- The unified paradigm does not require WNNM, and future work considers convolutional neural networks for non-local similarity and automated machine learning for hyperparameter tuning.
6. Appendix
The appendix derives a closed-form low-rank approximation solution and analyzes the projected noise variance; accompanying figures show PaU and WDC color images before and after denoising.
- 6. Appendix: The objective is reformulated as finding the best K-rank approximation of (Y_i)_(3), yielding a solution from its rank-K SVD.
- 6. Appendix: The rank-K SVD factors provide A_i = V and M̄_i = fold3(US) as the closed-form solution.
- 6. Appendix: The projected noise analysis establishes zero mean and derives an isotropic variance expression, var[b] = σ0I.
- 6. Appendix: The appendix concludes the proposition after the variance derivation.
- 6. Appendix: Figures 9 and 10 display PaU and WDC color images before and after denoising, using bands 80, 34, 9 and 190, 60, 27, respectively.