N2V2: Advanced Denoising & VK Band Modeling

• N2V2 in image processing is an advanced self-supervised denoising framework that integrates BlurPool operations and improved blind-spot replacement to nearly eliminate checkerboard artifacts.
• The method achieves measurable PSNR gains by addressing aliasing and high-frequency noise leakage through a non-residual U-Net design and strategic in-painting of masked pixels.
• N2V2 also denotes the modeling of the N₂ Vegard-Kaplan band system in Martian dayglow, enabling accurate retrieval of N₂/CO₂ mixing ratios and atmospheric dynamics.

N2V2 refers to two distinct concepts in scientific literature: (1) Noise2Void v2, an advanced self-supervised image denoising method addressing artifacts in convolutional network outputs (Höck et al., 2022), and (2) the N₂ Vegard-Kaplan (VK) band system, specifically the A³Σᵤ⁺–X¹Σg⁺ transitions in Martian dayglow, with emphasis on the A (v′=2)→X (v″) transitions and their modeling (Jain et al., 2011). Both are significant developments in their respective fields, one focusing on computational microscopy and image processing, the other on atmospheric spectroscopy and planetary aeronomy.

1. Noise2Void v2 (N2V2): Definition and Motivation

Noise2Void v2 (N2V2) is a self-supervised image denoising framework that builds upon the original Noise2Void paradigm, targeting biomedical microscopy and generic noisy imaging applications. N2V2 is designed to nearly eliminate the checkerboard artifacts that are characteristic of the vanilla Noise2Void approach. The method maintains the core blind-spot training protocol, in which randomly selected pixels are masked (the “blind spots”) and replaced with intensity values, and the network is penalized for prediction errors at those masked locations. N2V2 uses improved architectural choices and in-painting strategies that address the origins of grid artifacts due to aliasing and high-frequency noise leakage (Höck et al., 2022).

2. Sources of Artifacts in Vanilla N2V

Vanilla N2V’s checkerboard artifacts arise from two primary sources:

• Aliasing due to pooling/upsampling: Standard max-pooling layers in U-Net-style encoders, followed by up-convolutions in decoders, create aliasing that manifests as grid-like patterns when combined with blind-spot training.
• High-frequency noise leakage: Residual connections (input-to-output skip) and top-level skip connections in the U-Net allow direct transmission of noisy, high-frequency image components, which get amplified by the pixel-wise replacement strategy that can sample noise-laden pixels as replacements.

These factors lead to visually distracting, periodic artifacts (“checkerboard pattern”) in the reconstructed, denoised images.

3. Key Modifications Introduced in N2V2

N2V2 addresses the artifact sources through two classes of changes: network architecture and blind-spot pixel replacement.

3.1. Network Architecture Adjustments

• BlurPool layers: Every max-pooling layer is replaced with a BlurPool operation, consisting of a fixed low-pass (e.g., binomial) filter prior to downsampling. This suppresses frequency components above the new Nyquist limit, mitigating aliasing during upsampling.
• Non-residual U-Net: The classic U-Net residual shortcut (combining input with network output) is omitted, preventing direct propagation of unfiltered high-frequency image noise.
• Omission of top-level skip: The skip connection linking the first encoder and last decoder layers is removed, further impeding the passage of unfiltered, high-frequency features.

Combined, these architectural changes preserve multi-scale context aggregation while blocking aliasing and noise bypasses.

3.2. Improved Blind-Spot Replacement Strategies

Unlike vanilla N2V’s random neighbor replacement (which could select the noisy center pixel), N2V2 computes replacement values using the full local neighborhood (typically a 5×5 patch, excluding center):

• Mean replacement: $$R_\text{mean}(x)_i = \frac{1}{|N_i|} \sum_{j\in N_i} x_j$$
• Median replacement: $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$

These aggregated replacements are less sensitive to isolated noise spikes, ensuring smoother, more stable in-painting.

4. Formal Training Objective and Algorithmic Workflow

Let $x \in \mathbb{R}^{H \times W}$ be a noisy image and $M \in \{0,1\}^{H \times W}$ a random binary mask indicating blind spots. The modified image is

$\tilde{x} = M \odot x + (1-M) \odot R(x)$

with $R(x)$ as the neighborhood mean or median.

The U-Net $f(\cdot;\theta)$ is trained to minimize:

$$L(\theta) = \mathbb{E}_{x\sim D}\left[\sum_{i: M_i=0} (f(\tilde{x};\theta)_i - x_i)^2\right]$$

Training proceeds by:

1. Random sampling of blind-spot mask $M$
2. Computing $R(x)$ over each patch
3. Replacing blind-spot pixels accordingly
4. Minimizing squared error at masked locations
5. Backpropagating gradients and updating $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$0

Inference is straightforward: the denoiser operates on the full, unmasked input.

5. Empirical Evaluation and Performance

N2V2 was evaluated on both natural and biological image datasets with Gaussian and salt-&-pepper noise:

Dataset	Vanilla N2V (PSNR)	N2V2 (PSNR)	PSNR Gain
Mouse SP12	33.10 dB	33.34 dB	+0.24
BSD68 G70	27.70 dB	28.32 dB	+0.62
Mouse G20	34.49 dB	34.74 dB	+0.25
Convallaria_95	35.78 dB	36.39 dB	+0.61

Qualitatively, the grid artifacts prominent in vanilla N2V are almost entirely absent in N2V2 outputs, resulting in smoother, visually superior restorations across diverse settings (Höck et al., 2022).

6. Applications, Limitations, and Extensions

N2V2 functions as a drop-in replacement for standard N2V across microscopy and general imaging pipelines. No additional training data or runtime cost is required. It matches or exceeds alternative self-supervised methods (e.g., PN2V, DivNoising) on benchmarks. Potential future developments discussed include:

• Learnable blur filters for adaptive frequency suppression
• Hybrid architectures balancing residual connections for convergence with artifact suppression
• Probabilistic output layers atop the N2V2 backbone for uncertainty quantification

N2V2 is thus positioned as a practical default for blind-spot denoising in computational imaging workflows (Höck et al., 2022).

---

7. N₂ Vegard-Kaplan (VK) v=2 Bands in Martian Dayglow

Separately, “N2V2” can also refer to the A (v′=2)→X (v″) transitions of the N₂ Vegard-Kaplan (VK) system in the context of Martian dayglow emission modeling. These transitions are excited predominantly by photoelectron impact on N₂, followed by radiative and collisional relaxation.

The steady-state population for each vibrational level $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$1 of electronic state $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$2 at altitude $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$3 is governed by:

$$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$4

The direct excitation and line-of-sight intensity calculations, as well as dependencies on electron-impact cross sections, solar EUV flux, model atmospheres, and N₂/CO₂ mixing ratios, are rigorously described in (Jain et al., 2011).

To match SPICAM limb data, GCM N₂ densities must be reduced by a factor $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$5, resulting in derived N₂/CO₂ mixing ratios of 1–3% at 120–140 km and 4–7% at 170 km. Limb peak intensities at SZA = 45° for the A (v′=2) bands reach $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$6–$$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$7 kR in low solar activity; sensitivity to model parameters can induce up to $$R_\text{med}(x)_i = \mathrm{median}_{j\in N_i} \{x_j\}$$8 variation in predicted intensities.

---

N2V2 thus refers to both a decisive step in denoising algorithm development for imaging science (Höck et al., 2022) and a set of key molecular transitions central to atmospheric remote sensing models (Jain et al., 2011). In both domains, the term denotes refined techniques or features that improve the accuracy and reliability of quantitative scientific analyses.