Patch-Based Denoising Procedure

• Patch-based denoising is a restoration method that extracts overlapping patches to exploit non-local similarities for effective noise suppression.
• It utilizes clustering and subspace modeling techniques to project patches onto low-dimensional structures, enhancing denoising performance.
• Aggregating individually denoised patches via weighted or simple averaging reconstructs images while preserving fine details and mitigating artifacts.

A patch-based denoising procedure is a class of signal and image restoration algorithms that operate by extracting overlapping local patches from observed noisy data and leveraging the redundancy, structure, and non-local relationships among those patches to suppress noise and recover the underlying signal. These methods are ubiquitous in modern image processing, medical imaging, and other domains where spatially and/or temporally coherent signals are corrupted by structured or unstructured noise. The unifying characteristic is that local neighborhoods (patches) are treated both as units of representation and as sets for non-local estimation, clustering, modeling, or aggregation.

1. Mathematical Foundations and Patch Extraction

Patch-based denoising begins with the extraction of overlapping (or non-overlapping) patches from the noisy input. For an image, this entails mapping the observed signal $y$ of size $n \times n$ into a collection of $N$ local patches, each of dimension $L \times L$ (vectorized $d=L^2$). The extraction stride governs overlap; typical values are 1 (full overlap) through $L$ (non-overlapping).

Formally, for each reference location $i$, the corresponding patch is $x_i \in \mathbb{R}^d$. This construction enables the exploitation of repeated structures, local low-rankness, or geometric priors within the patch-set. Patches may also be constructed in spatio-temporal or multi-channel domains, e.g., for volumetric, video, or color data (Wickremasinghe et al., 8 Dec 2025, Buades et al., 2022).

Patch extraction is a prerequisite for both classical model-based methods (e.g., non-local means, PCA/low-rank modeling) and modern data-driven approaches (e.g., patch-wise deep networks, patch-wise diffusion models).

2. Patch Clustering and Subspace Modeling

A central paradigm is the grouping or clustering of patches into sets that share geometric or statistical similarity. The purpose of clustering is to exploit the redundancy and self-similarity in natural data, enabling more accurate modeling and suppressing noise amplification caused by direct local estimation. Several clustering strategies have been established:

• Affine Subspace Clustering: Patches are assumed to reside near several low-dimensional affine subspaces $\{\mu_k + \mathrm{col}(U_k)\}_{k=1}^K$, where $\mu_k$ is a cluster centroid and $U_k$ is an orthonormal subspace basis (Wickremasinghe et al., 8 Dec 2025). Clustering alternates between assigning patches to subspaces and updating $(\mu_k, U_k)$ via K-means and local PCA, respectively. The denoising estimate is obtained by projecting patches onto the subspace of their cluster:

$$\hat{x}_i = \mu_{k(i)} + U_{k(i)}U_{k(i)}^T(x_i - \mu_{k(i)})$$

This approach significantly improves model fit and denoising performance over pure linear subspace models, especially for image data with nonzero mean (non-negativity constraints).

• Gaussian Mixture Models (GMM) and Bayesian Patch Priors: Patches are modeled as samples from a mixture of Gaussians with full covariance, $$p(x) = \sum_{k=1}^K \pi_k \mathcal{N}(x; \mu_k, \Sigma_k)$$, where clustering and parameter estimation can be performed jointly via Expectation-Maximization, possibly under a hyper-prior for domain adaptation (Lu, 2019, Luo et al., 2016). Cluster assignment and prior adaptation are performed by maximizing or integrating posteriors under the noisy observation model, sometimes leveraging pre-filtered (pilot) estimates.
• Patch Aggregation via Manifold Structure: Nonlinear patch manifolds are estimated by graph-based patch connectivity and subsequent embedding (e.g., via geodesic distances and double-centering to obtain a Gramian structure). The patch set clusters are effectively induced by proximity on the learned manifold (Gajamannage, 2020, Gajamannage et al., 15 Apr 2025).
• Patch Quality-Driven Clustering and Selection: Variants exist where patch groups are selected based not just on Euclidean proximity, but using clustering in patch space (e.g., GMM or K-means on UPE-corrected patches), often guided by distributions centered away from the noisy input (Lu, 2019).

3. Patch-Based Denoising Mechanisms

After clustering or grouping, a wide variety of patch-level denoising operators have been developed, tailored to noise statistics, structural priors, and computational framework:

• Least-Squares Projection (Patch Subspace Projection): After clustering, each patch is projected onto the corresponding low-dimensional affine (or linear) model as in the PSP approach (Wickremasinghe et al., 8 Dec 2025).
• Shrinkage and Group Sparsity: For models where patches are assumed to be well approximated in a learned basis (e.g., from local PCA or dictionary learning), group-sparse transforms and Wiener-type shrinkage are applied to patch coefficients, either by hard thresholding (PLR model) (Hu et al., 2015) or Bayesian shrinkage informed by local or targeted databases (Luo et al., 2014).
• Nonlocal Means and Collaborative Filtering: Nonlocal means-type methods aggregate each patch estimate by weighted averaging with non-locally similar patches, often enhanced by “second-stage” collaborative or low-rank filtering. In advanced variants, the weights or patch sets are determined via more sophisticated quality metrics or reliability estimates (e.g., UPE) (Lu, 2019).
• Nonlinear and Manifold Methods: Nonlinear filtering on patch manifolds involves projecting each patch onto the principal components of a Gramian (formed from manifold geodesics), or applying smoothing/diffusion in the reordered patch domain for artefact suppression (Gajamannage, 2020, Gajamannage et al., 15 Apr 2025, Bodduna et al., 2020).
• Convex Optimization and Overlap Regularization: Some patch-based denoisers are posed as global optimization with convex composite objectives that enforce both patch-level data fidelity, sparsity, and consistency of overlapping patches (Mirone et al., 2013).
• Poisson and Structured Noise Models: When noise is signal-dependent (e.g., Poisson), patch-based methods use local moment estimation from pilot denoisers to directly formulate MMSE/BLP predictors for patches, leveraging patch clustering to estimate the local covariance structure (Niknejad et al., 2018).
• Deep Learning and Patch-Wise Self-Supervised Training: Modern frameworks apply patch-level training, either in a self-supervised blind-spot manner or as patch-wise diffusion models, with aggregation strategies (e.g., patch-based windowed ensemble, local quality-based merging) for the full denoised image (Nguyen et al., 19 Oct 2025, Xia et al., 2022, Zhu et al., 2024).

4. Patch Aggregation and Image Reconstruction

The final denoised image is reconstructed from the set of individually denoised patches. In nearly all patch-based methods, patches are extracted with substantial overlap, meaning each pixel is covered by multiple denoised patch predictions. Aggregation is typically executed as follows:

• Simple Averaging: Each pixel is assigned the average of all overlapping predictions covering that pixel, ensuring a minimum-variance unbiased estimator under Gaussian errors (Wickremasinghe et al., 8 Dec 2025).
• Weighted Averaging: Weights can be inversely proportional to local reconstruction error or quality metrics, leveraging patch reliability or group confidence (Wickremasinghe et al., 8 Dec 2025, Lu, 2019, Nguyen et al., 19 Oct 2025).
• Shepard Weights: In manifold-type aggregation, weights decay exponentially with the pixel distance between the patch center and the pixel being reconstructed, effectively smoothing the aggregated estimate and reducing seams (Gajamannage, 2020, Gajamannage et al., 15 Apr 2025).
• Artifact Mitigation: Patch aggregation with smoothing of noise estimates (e.g., via DDIM-style averaging in diffusion models) prevents checkerboard artifacts and maintains global consistency (Özdenizci et al., 2022).
• Quality-Driven Adaptive Merging (Ensemble Methods): In self-supervised ensembles, for each patch location, the model with the best local quality metric (e.g., contrast-to-noise ratio, texture preservation) is selected and its patch is merged via weighted or direct assignment (Nguyen et al., 19 Oct 2025).

5. Parameter Selection, Computational Considerations, and Applicability

Patch-based denoising performance is acutely sensitive to method parameters. Key hyperparameters include:

Parameter	Typical Range / Setting	Role
Patch size $L$	$5 \leq L \leq 11$	Trade-off: detail vs. noise suppression
# Clusters $K$	$10 \leq K \leq 64$	Controls granularity of local models
Subspace dim $r$	$5 \leq r \leq 10$	Retained energy / local complexity
# Neighbors $m$	$30 \leq m \leq 245$	Sample redundancy for denoising
Overlap stride $s$	$1$ (full), $L$ (none)	Redundancy and run-time
RSVD/Gramian components $L$	$50\leq L\leq 100$	Detail preservation in manifold

Parameter selection can be guided by “elbow” curves of within-cluster error, energy fraction retained in PCA, or cross-validation on PSNR/SSIM. Overfitting is possible with overly high $K$ or $r$, leading to noise modeling; underfitting with small values results in loss of detail. For diffusion and ensemble models, patch stride and window size trade off quality and computational burden (Nguyen et al., 19 Oct 2025).

Computationally, classic methods scale as $O(n^2 d^2 K)$ for clustering, with $O(n^6)$ for naive manifold methods (reduced to $O(n^4)$ via randomized SVD) (Gajamannage et al., 15 Apr 2025). Patch matching and sorting dominates run-time in NLM/BM3D-type workflows, while nearest-neighbor graph construction and eigen-analysis are bottlenecks for Gramian approaches. Deep patchwise denoisers are bounded by GPU memory and patch-level batch size (Zhu et al., 2024).

Patch-based denoising excels in modalities with high redundancy (natural images, repetitive textures, biomedical volumes), with applications in low-dose medical imaging, HDR fusion, remote sensing, microscopy, and signal denoising (Nguyen et al., 19 Oct 2025, Buades et al., 2022, Izadi et al., 2020). Specialized adaptations exist for Poisson noise, salt-and-pepper noise, and blind/mixed noise distributions (Niknejad et al., 2018, Fu et al., 2018, Fantasia et al., 20 Mar 2025).

6. Advantages, Limitations, and Performance Comparisons

Patch-based denoising has substantively advanced the state of the art by leveraging nonlocal redundancy and expressive local models:

• Advantages:
  • Structured redundancy capture through patch grouping yields superior noise suppression without oversmoothing.
  • Subspace, group-sparse, or manifold models adapt to a broad spectrum of image structure.
  • Patch-based frameworks are modular—aggregation, clustering, model fitting, and denoising mechanisms can be independently advanced and tuned (Wickremasinghe et al., 8 Dec 2025, Gajamannage et al., 15 Apr 2025).
  • Deep patch-based diffusion and self-supervised models allow for tractable, memory-efficient deployment, including on very large images (Özdenizci et al., 2022, Nguyen et al., 19 Oct 2025).
• Limitations:
  • Patch matching and clustering can overfit to noise unless robust (UPE), Bayesian, or pilot-based estimation is employed (Lu, 2019).
  • Large-scale nearest neighbor and manifold-based frameworks are computationally intensive; even with RSVD, these scale as $O(n^4)$, which can be prohibitive at megapixel resolutions (Gajamannage et al., 15 Apr 2025).
  • Patch boundaries and aggregation schemes must be carefully designed to mitigate block artifacts or seam formation.
  • For highly nonstationary data or images with little repetition, patch grouping becomes less effective.

Quantitatively, methods such as affine-subspace clustering with least-squares projection, Gramian manifold approaches, and patch-ensemble deep denoisers regularly outperform classical local and even some deep global denoisers in PSNR/SSIM and visual quality, with typical gains of 0.5–1 dB in PSNR on standard benchmarks and more marked improvements in edge/texture preservation (Wickremasinghe et al., 8 Dec 2025, Gajamannage et al., 15 Apr 2025, Nguyen et al., 19 Oct 2025).

7. Recent Extensions and Emerging Directions

Recent work has advanced the patch-based denoising paradigm along several axes:

• Self-Supervised Deep Patch Denoising: Patch-extracted pseudo-clean targets, blind-spot networks, and patch-ensemble selection strategies for settings with no ground-truth (Nguyen et al., 19 Oct 2025).
• Manifold Learning and Nonlinear Patch Models: Randomized and scalable approximations for high-dimensional patch-manifold denoising, with emphasis on remote sensing and large-scale data (Gajamannage et al., 15 Apr 2025).
• Blind/Adaptive Model Selection: Patchwise learning of local regularization weights or parameters, especially for unknown or mixed noise conditions, via supervised or hybrid deep models (Fantasia et al., 20 Mar 2025).
• Diffusion-Based Patch Denoising: Patchwise denoising diffusion probabilistic models (DDPMs), both for image restoration and inverse problems (e.g., CT sinogram inpainting), enabling distributed and memory-scalable workflows (Xia et al., 2022, Özdenizci et al., 2022).
• Specialized Patch Models: Contour-prior stencils, impulse-robust patch regression, and patch-ordering for non-Gaussian noise types such as salt-and-pepper or compound Poisson-Gaussian distributions (Fu et al., 2018, Fu et al., 2018, Izadi et al., 2020).

Ongoing research addresses robust automated hyperparameter selection, learning of low-dimensional patch representations for multimodal data, and the theoretical understanding of patch manifold geometry for denoising and beyond.

---

References:

• "Affine Subspace Models and Clustering for Patch-Based Image Denoising" (Wickremasinghe et al., 8 Dec 2025)
• "Good Similar Patches for Image Denoising" (Lu, 2019)
• "Adaptive Image Denoising by Targeted Databases" (Luo et al., 2014)
• "Image Denoising Using the Geodesics' Gramian of the Manifold Underlying Patch-Space" (Gajamannage, 2020)
• "Efficient and Robust Remote Sensing Image Denoising Using Randomized Approximation of Geodesics' Gramian on the Manifold Underlying the Patch Space" (Gajamannage et al., 15 Apr 2025)
• "Patch-Based Denoising Diffusion Probabilistic Model for Sparse-View CT Reconstruction" (Xia et al., 2022)
• "Poisson Image Denoising Using Best Linear Prediction: A Post-processing Framework" (Niknejad et al., 2018)
• "Patch-based learning of adaptive Total Variation parameter maps for blind image denoising" (Fantasia et al., 20 Mar 2025)
• "Restoring Vision in Adverse Weather Conditions with Patch-Based Denoising Diffusion Models" (Özdenizci et al., 2022)
• "Patch-based Non-Local Bayesian Networks for Blind Confocal Microscopy Denoising" (Izadi et al., 2020)
• "A salt and pepper noise image denoising method based on the generative classification" (Fu et al., 2018)
• "Image denoising with less artefacts: Novel non-linear filtering on fast patch reorderings" (Bodduna et al., 2020)
• "ECG Signal Denoising Using Multi-scale Patch Embedding and Transformers" (Zhu et al., 2024)
• "A note on patch-based low-rank minimization for fast image denoising" (Hu et al., 2015)
• "Adaptive Image Denoising by Mixture Adaptation" (Luo et al., 2016)
• "A Faster Patch Ordering Method for Image Denoising" (Munir, 2017)
• "Joint denoising and HDR for RAW video sequences" (Buades et al., 2022)
• "A Convex Functional for Image Denoising based on Patches with Constrained Overlaps and its vectorial application to Low Dose Differential Phase Tomography" (Mirone et al., 2013)
• "SDPA++: A General Framework for Self-Supervised Denoising with Patch Aggregation" (Nguyen et al., 19 Oct 2025)
• "Patch-based Contour Prior Image Denoising for Salt and Pepper Noise" (Fu et al., 2018)