Patchwise NN Detection in Image Denoising

• Patchwise NN detection is a method in Non-Local Means denoising that compares overlapping patches using Euclidean and statistical metrics to improve noise reduction.
• The Statistical Nearest Neighbor (SNN) criterion adjusts standard patch matching by targeting patches with distances near 2σ², effectively reducing bias from noise correlations.
• This approach enhances denoising quality in moderate-to-high noise scenarios while maintaining computational efficiency, making it practical for real-world image processing.

Patchwise nearest-neighbor (NN) detection is central to the Non-Local Means (NLM) denoising paradigm, in which overlapped patches from a noisy image are compared and aggregated for noise removal. The canonical approach uses the Euclidean (ℓ²) distance to identify K nearest neighbors, but recent analysis has demonstrated inherent bias associated with this method. Statistical nearest neighbor (SNN) selection has been introduced to mitigate this bias by leveraging statistical properties of the noise, leading in practice to superior denoising performance, particularly when the noise is significant or only a modest number of neighbors can be processed efficiently (Frosio et al., 2017).

1. Patchwise Distance Metrics in NLM

In NLM denoising, each reference patch $\mu_n = [\mu_n^0, ..., \mu_n^{P-1}]$ of size $P$ is compared against candidate patches $\gamma = [\gamma^0, ..., \gamma^{P-1}]$ using the (normalized) squared Euclidean distance:

$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$

This metric governs both NN/SNN selection and the computation of reconstruction weights:

$$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$

where $\sigma^2$ is the known noise variance and $h$ is a filtering parameter. The denoised patch is the normalized weighted average:

$$\widehat{\mu}(\mu_n) = \frac{\sum_k w_{\mu_n, \gamma_k} \gamma_k}{\sum_k w_{\mu_n, \gamma_k}}$$

Standard practice limits the sum to the $K$ nearest patches to reduce computational complexity.

2. Bias in Standard Nearest-Neighbor Selection

Restricting reconstruction to the $K$ nearest neighbors introduces a fundamental bias. For the toy case of $P$0 (scalar patches) under i.i.d. Gaussian noise ($P$1, $P$2), the expectation of the unweighted $P$3-NN average is:

$P$4

where $P$5, $P$6, $P$7 is the half-width of the interval containing the $P$8 nearest samples, and $P$9, $\gamma = [\gamma^0, ..., \gamma^{P-1}]$0 are the standard normal PDF and CDF. For $\gamma = [\gamma^0, ..., \gamma^{P-1}]$1, this expression demonstrates a systematic bias, as the output is pulled toward the noisy observation $\gamma = [\gamma^0, ..., \gamma^{P-1}]$2. This effect manifests in noise-to-noise matching and yields colored, structured residuals, especially visible in flat image regions. Mean-squared error is dominated by this squared bias term (Frosio et al., 2017).

3. Statistical Nearest Neighbor (SNN) Criterion

The SNN approach modifies neighbor selection to be statistically aware. For independent noisy patches, the expected squared distance is $\gamma = [\gamma^0, ..., \gamma^{P-1}]$3. Standard NN selection (minimizing $\gamma = [\gamma^0, ..., \gamma^{P-1}]$4) disproportionately matches patches sharing the same noise realization. Instead, SNN identifies patches with distances close to this statistical expectation, thereby seeking neighbors with "orthogonal" noise. The SNN selection rule introduces an offset parameter $\gamma = [\gamma^0, ..., \gamma^{P-1}]$5, and the score for a candidate patch is:

$\gamma = [\gamma^0, ..., \gamma^{P-1}]$6

The $\gamma = [\gamma^0, ..., \gamma^{P-1}]$7 patches with lowest SNN-score are selected. When $\gamma = [\gamma^0, ..., \gamma^{P-1}]$8 this reduces to the standard NN criterion; $\gamma = [\gamma^0, ..., \gamma^{P-1}]$9 targets the expected inter-patch noise distance, thereby reducing bias.

Pseudo-code—SNN Neighbor Detection:

$h$0 SNN assumes additive zero-mean Gaussian noise, known variance, and that candidate patches are independent noisy replicas of the signal (Frosio et al., 2017).

4. Comparative Evaluation: Standard NN versus SNN

Empirical evaluation was performed on the Kodak dataset (24 color images) with various Gaussian noise levels ($$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$0) and also on colored noise. Metrics included PSNR, SSIM, MSSSIM, GMSD, FSIM, and FSIM_C.

Configuration	PSNR (dB)	FSIM_C
NLM$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$1	31.18	0.9468
NLM$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$2	29.21	0.9633
NLM$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$3 (SNN)	30.45	0.9621
NLM$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$4 (col noise)	29.42	0.8854
NLM$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$5 (SNN, col)	31.04	0.8778
BM3D-CFA (col noise)	31.66	0.9183

Best SNN PSNR typically occurs at $$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$6, while best FSIM_C is reached at $$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$7–$$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$8. On real images (NVIDIA Shield ISO 1200), NLM with SNN ($$\delta^2(\mu_n, \gamma) = \frac{1}{P} \sum_{i=0}^{P-1} (\mu_n^i - \gamma^i)^2$$9, $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$0) achieves PSNR=24.55, FSIM_C=0.9921; the more computationally demanding BM3D-CFA achieves PSNR=25.26, FSIM_C=0.9941. These results indicate that SNN can substantially reduce bias and residual artifacts with low neighbor counts, almost matching best-in-class methods for moderate-to-high noise (Frosio et al., 2017).

5. Computational Cost and Implementation Considerations

Both NN and SNN incur a per-patch cost of $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$1 for computing all squared distances, and $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$2 (or $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$3 with selection algorithms) for identifying the $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$4 highest-scoring neighbors. The weighting and averaging stage is $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$5. Compared to standard NN, SNN only adds a single subtraction and absolute value per candidate. The total asymptotic computational cost thus differs negligibly. When using very few neighbors (e.g., $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$6 vs. full search $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$7), SNN enables a $$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$8–$$w_{\mu_n, \gamma} = \exp\left\{-\frac{\max[0,\, \delta^2(\mu_n, \gamma) - 2\sigma^2]}{h^2} \right\}$$9 speedup; transition from NN to SNN is nearly cost-neutral (Frosio et al., 2017).

6. Practical Recommendations and Observed Effects

SNN is most beneficial for small $\sigma^2$0 (e.g., $\sigma^2$1) and moderate-to-high noise ($\sigma^2$2), where standard NN bias is maximal. SNN with $\sigma^2$3 permits practical trade-offs: as $\sigma^2$4, PSNR and SSIM improve (less structured noise), while perceptual metrics like FSIM and GMSD peak at $\sigma^2$5–$\sigma^2$6. With limited computational budgets, SNN yields cleaner flat regions than NN with comparable $\sigma^2$7, approaching the quality of BM3D-CFA for colored/demosaiced noise at significantly reduced cost. When large numbers of neighbors are affordable, standard NLM can denoise flat regions effectively but tends to over-smooth details; SNN with small $\sigma^2$8 combines smoothing of flat regions and preservation of detail (Frosio et al., 2017).

7. Summary and Implications

Classical patchwise nearest-neighbor detection in NLM using ℓ² distance is not unbiased: its tendency to match noise realizations introduces residual structured noise in the output. The SNN criterion resolves this by selecting patches at the expected inter-patch noise distance ($\sigma^2$9), nearly eliminating bias. SNN imposes minimal computational overhead and improves denoising quality when neighbor count is limited or the noise is colored, offering a practical alternative for advanced image processing pipelines (Frosio et al., 2017).