State-Space Doubling for Robust Denoising

• State-space doubling is a concept that expands the signal representation to improve separation of noise by exploiting structural redundancy in data ensembles.
• It integrates RMT-based eigenvalue analysis with deep learning frameworks that decouple explicit and implicit noise, optimizing denoising for various imaging modalities.
• Experimental results show significant improvements in PSNR, SSIM, and spectral fidelity, underscoring the method’s practical impact on high-dimensional signal recovery.

ReDeNoise refers to two distinct algorithmic frameworks, both targeting robust signal recovery from noisy observations but operating in different domains and with distinct mathematical underpinnings. The first, introduced by Vinayak and Simmhan, utilizes random matrix theory (RMT) for statistical denoising of images based on eigenvalue outlier detection in empirical covariance matrices (Basu et al., 2010). The second, developed for hyperspectral imaging, is a multi-stage deep learning method for real noise decoupling via explicit and implicit component separation with domain-specific architectures, notably for hyperspectral image (HSI) denoising (Zhang et al., 21 Nov 2025). Both approaches leverage structural redundancy in ensembles of data to separate signal from noise, but employ fundamentally different estimation principles and loss formulations.

1. Random Matrix Theory–Based ReDeNoise

This algorithm exploits the spectral properties of empirical covariance (or correlation) matrices of ensembles of noisy signals to separate universal, noise-induced eigenmodes from data-driven, signal-bearing components. It is especially effective when multiple independent noisy realizations of the same underlying signal are available, as with repeated imaging or multi-acquisition setups.

Mathematical Framework

Given $N$ independent noisy measurements $x_i \in \mathbb{R}^M$ (e.g., rows or columns of an image), the data matrix $X \in \mathbb{R}^{M \times N}$ is constructed by stacking these observations. Each coordinate (pixel/timepoint/band) of $X$ is mean-centered and optionally standardized. The empirical correlation matrix is calculated as

$$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$

with eigenvalues $$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_M$$ and eigenvectors $v_1, \ldots, v_M$.

Marčenko–Pastur Spectral Bounds

Under the null hypothesis that observations are i.i.d. Gaussian noise of variance $\sigma^2$, the eigenvalue distribution of $C$ converges, in the asymptotic regime $M,N \rightarrow \infty$ with $x_i \in \mathbb{R}^M$0, to the Marčenko–Pastur distribution with edges

$x_i \in \mathbb{R}^M$1

Signal components cause empirical eigenvalues to exceed $x_i \in \mathbb{R}^M$2. Hence, eigenvalues above this threshold are indicative of structured, non-noise content.

Signal Extraction and Denoising

Let $x_i \in \mathbb{R}^M$3, with associated eigenvectors $x_i \in \mathbb{R}^M$4 and diagonal eigenvalue matrix $x_i \in \mathbb{R}^M$5. The denoised data is reconstructed by projection onto the identified signal subspace:

$x_i \in \mathbb{R}^M$6

The process is repeated independently for each row or column (or patch); outputs are then reassembled into a denoised image. Details on parameter selection, noise estimation (e.g., via spectrum median), safety margining ($x_i \in \mathbb{R}^M$7), and computational optimization (Lanczos, randomized SVD) are specified for scalable application (Basu et al., 2010).

2. Real Noise Decoupling for Hyperspectral Image Denoising

The second ReDeNoise algorithm addresses the denoising of real-world hyperspectral images by decomposing noise into two orthogonal components: explicitly modeled (physical, instrument-driven) and implicitly modeled (residual and unknown sources) noise. The framework customizes denoising strategies to these components using two network architectures and a staged learning procedure (Zhang et al., 21 Nov 2025).

Noise Model and Explicit/Implicit Decoupling

The noisy measurement $x_i \in \mathbb{R}^M$8 (where $x_i \in \mathbb{R}^M$9 is number of spectral bands) is modeled as

$X \in \mathbb{R}^{M \times N}$0

where $X \in \mathbb{R}^{M \times N}$1 follows a mixed Poisson–Gaussian distribution $X \in \mathbb{R}^{M \times N}$2 per band, and $X \in \mathbb{R}^{M \times N}$3 aggregates all remaining unmodeled noise.

3. Algorithmic Modules and Learning Strategy

Stage 1: Explicit Noise Pre-Training

Synthetic datasets $X \in \mathbb{R}^{M \times N}$4 are generated from clean $X \in \mathbb{R}^{M \times N}$5 and $X \in \mathbb{R}^{M \times N}$6 sampled from the explicit noise distribution, optionally including read-out and stripe patterns. The EMNet backbone (3D U-Net with residual and spectral attention blocks) is pre-trained with the Charbonnier loss:

$X \in \mathbb{R}^{M \times N}$7

to learn explicit noise removal in isolation.

Stage 2: Implicit Noise Removal via Wavelet-Guided IMNet

After freezing EMNet, the residual $X \in \mathbb{R}^{M \times N}$8 is found to concentrate in high-frequency subbands. Therefore, IMNet (another 3D U-Net) is conditioned on multi-scale wavelet guidance maps $X \in \mathbb{R}^{M \times N}$9, generated via DWT and wavelet convolutions on $X$0. IMNet outputs an intermediate denoised image $X$1 and a residual $X$2, trained to minimize:

• Charbonnier loss between $X$3 and ground truth.
• KL divergence enforcing $X$4 to follow the explicit noise statistics.

Stage 3: Joint Fine-Tuning

Both networks are unfrozen and optimized jointly using real data, with an added spectral consistency loss to enforce spectral similarity:

$X$5

The total loss is

$X$6

where typical choices are $X$7, $X$8.

A summary pseudocode of the full pipeline (pre-training, implicit training, fine-tuning) is provided in (Zhang et al., 21 Nov 2025).

4. Practical Considerations and Computational Complexity

The random matrix–based ReDeNoise incurs $X$9 complexity per strip for eigen-decomposition, with total cost $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$0 for $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$1 strips. For $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$2, working in the dual $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$3 domain and using iterative eigen-solvers (Lanczos, randomized SVD) mitigate computational and memory overheads (Basu et al., 2010). The deep noise decoupling approach, with TDSAT or HSDT backbones, involves $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$4 parameters and $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$5500–600 GFLOPs per HSI, training to convergence in $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$6 200–400 epochs on high-end GPUs, with inference times of $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$72s on $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$8 inputs (Zhang et al., 21 Nov 2025).

5. Experimental Results and Quantitative Gains

The multi-stage denoising strategy of ReDeNoise (HSI) achieves state-of-the-art performance on paired real datasets:

Dataset	Method	PSNR (dB)	SSIM	SAM (°)
RealHSI	HSDT	31.24	0.958	3.751
RealHSI	ReDeNoise(HSDT)	32.31	0.967	2.742
MEHSI	VolFormer	34.89	0.974	3.053
MEHSI	ReDeNoise(TDSAT)	36.36	0.981	2.228

Denoising not only removes stripe, read-out, and residual noise, but preserves spatial and spectral features, with correlation $$C = \frac{1}{N} X X^\top \in \mathbb{R}^{M \times M}$$9 across all bands. Error maps demonstrate $$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_M$$0–$$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_M$$1 dB improvement compared to prior approaches (Zhang et al., 21 Nov 2025).

6. Core Assumptions, Limitations, and Applicability

The RMT-based method presupposes i.i.d. Gaussian noise and sufficient sample-ensemble size $$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_M$$2 for optimal noise–signal separation, with dominant low-rank signal subspaces often observed in natural images (Basu et al., 2010). The noise decoupling framework assumes the suitability of the explicit noise model for simulating $$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_M$$3; residual $$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_M$$4 is assumed to be removable via high-frequency wavelet-guided deep learning. For images with different or more complex noise models, retraining or model adaptation may be necessary.

7. Summary

ReDeNoise encompasses both theoretically principled, RMT-based denoising for images and advanced, explicitly-decoupled deep learning for real-world hyperspectral image denoising. Both approaches leverage the statistical properties of ensembles to disentangle signal from noise but utilize distinct mathematical foundations and network architectures. Their documented efficacy is demonstrated in quantitative benchmarks, underlining their utility for real and synthetic denoising tasks across imaging domains (Basu et al., 2010, Zhang et al., 21 Nov 2025).