LF-BVN: Blind-View Network for Self-Supervised Light Field Denoising

Conference: CVPR 2026
Paper: CVF Open Access
Code: https://github.com/shuozh/LF-BVN (Code availability as stated in the paper)
Area: Image Restoration / Light Field Denoising / Self-supervised
Keywords: Light Field Denoising, Self-supervised, Blind-view, Geometric Invariant Mask, Cross-view Consistency

TL;DR

The "blind-spot" ideology from single-image denoising is extended to "blind-view" for light fields (LF). By masking a subset of views and reconstructing them using the multi-view consistency of the remaining views, the network is trained without clean images. A Geometric Invariant Mask (GIM) enables a single weight-shared network to denoise all views. The method achieves or surpasses supervised counterparts on synthetic, real-world, and microscopic light fields.

Background & Motivation

Background: Light field (LF) cameras record spatial and angular information in a single exposure. However, devices like camera arrays, Lytro Illum, and Raytrix often produce noisy LFs due to hardware and lighting limitations, which hampers downstream tasks like view synthesis, super-resolution, and depth estimation. Existing methods are categorized into traditional methods (LFBM5D, Hyperfan4D), which rely on handcrafted priors and often cause artifacts, and data-driven methods (APA, MSPNet, SRDNet), which require noise-clean pairs but struggle with generalization.

Limitations of Prior Work: Accurately aligned noise-clean LF pairs are nearly impossible to obtain in real-world scenarios. Existing self-supervised attempts (VCDNet, V2V3D) derived from Light Field Microscopy (LFM) suffer from two issues: heavy dependence on the Point Spread Function (PSF), which is absent in standard LF imaging, and the use of multiple independent networks to denoise different views, leading to visible inconsistencies and artifacts.

Key Challenge: Single-image blind-spot denoising relies on spatial neighborhoods to infer target pixels, often mistaking noise for texture in flat areas or real edges for noise. Light fields provide more reliable supervision via multi-view consistency: projections of the same 3D point across views are photometrically consistent when noise-free, whereas noise is highly unlikely to be consistent across views. The challenge lies in extending the blind-spot concept to "blind-view" without training separate networks for every view while maintaining consistency.

Goal: (1) Design a fixed mask compatible with LF structures to allow a single network to denoise all views; (2) Enforce cross-view consistency during multi-branch denoising; (3) Extract reliable depth cues from noisy LFs to guide the denoising process.

Key Insight: The authors observe that LFs possess "rotational invariance." By synchronously rotating the entire LF in both spatial and angular domains, the geometric (disparity) relationships between views remain unchanged (following [28]). This implies that applying the same mask to rotated inputs preserves the geometry observed by the network, allowing a single weight-shared network to handle all rotation branches.

Core Idea: Replace "blind-spot" with "blind-view"—masking several views to force the network to reconstruct them from other views. Multi-view consistency serves as a free self-supervision signal. Combining GIM with weight-sharing collapses the "one network per view" requirement into "one network for four rotation branches."

Method

Overall Architecture

LF-BVN is a self-supervised light field denoising framework. It takes a noisy LF (central 7×7 views) as input and outputs a denoised LF without needing clean references. The architecture uses a four-branch, weight-shared structure. First, a Geometric Invariant Mask (GIM) masks a fixed 25% of the views. The LF is then rotated by 0°/90°/180°/270° and fed into the branches. All branches share the parameters \(f_\theta\). Inside each branch, the masked LF is processed by a denoising module \(f_{\theta_1}\) to reconstruct a "latent representation volume" encoding the scene geometry, followed by a rendering decoder \(f_{\theta_2}\). Finally, outputs are inverse-rotated and fused. A refocused depth estimation module extracts depth probabilities from the noisy LF for guidance, while the reconstruction consistency loss \(\ell_{rc}\) ensures the latent representations across branches remain consistent.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Noisy LF<br/>Central 7×7 Views"] --> B["Geometric Invariant Mask (GIM)<br/>Fixed 25% view masking"]
    B -->|"Spatial-Angular Rotation 0/90/180/270°"| C["Four-branch Denoising Module<br/>Weight-shared → Latent representation"]
    G["Refocused Depth Module<br/>PSV→Focal Stack→Depth Prob."] -->|"Depth Guidance ℓdepth"| C
    C -->|"Cross-branch Consistency ℓrc"| D["Rendering Decoder<br/>Inverse Rotation + Fusion"]
    D --> E["Denoised LF"]

Key Designs

1. Geometric Invariant Mask (GIM): Denoising All Views with One Network Blind-view denoising is difficult because while pixels in single images are statistically similar, LF views depend on cross-view geometry, which varies by view position. Changing the masking pattern usually requires retraining. Based on the disparity formula \(D_{j\to i}(x_i)=(a_j-a_i)\cdot\frac{B\cdot f}{Z(x_i)}\), the authors argue that for a fixed scene \(Z\), disparity depends only on relative angular positions. Therefore, a fixed angular mask is necessary to maintain geometric consistency. GIM leverages the property that geometry is invariant under synchronous spatial-angular rotation. By rotating the LF through 4 branches, the fixed mask "rotates" to cover all views, while the network always sees the same geometric relationships. GIM satisfies three principles: fixed angular pattern, uniform blind-view distribution, and a 25% masking ratio (optimal for 90° rotations across 4 branches).

2. Latent Representation Volume + Reconstruction Consistency Loss \(\ell_{rc}\) Denoising views separately naturally leads to photometric inconsistency. The authors introduce a view-independent LF representation—the Latent Representation Volume (following [29], shape \(4C\times H\times W\)). Since all branches observe the same scene, they should learn the same latent representation. Consistency is enforced via: $\(\ell_{rc}=\sum_{r}\left\|R^{-r}_{x}\big(f_{\theta_1}(R^{r}_{ax}(L)\odot M)\big)-f_{\theta_1}(L\odot M)\right\|_1,\quad r\in\{90,180,270\}\)$ By explicitly aligning geometric representations, cross-view photometric consistency is significantly improved, avoiding artifacts found in methods like V2V3D. The blind-view loss \(\ell_{blind}=\|f_\theta(M\odot L)-(1-M)\odot L\|_2^2\) measures the error only on masked views.

3. Refocused Depth Estimation Module + Depth Loss \(\ell_{depth}\) Accurate blind-view denoising requires finding 3D correspondences, which necessitates depth. Typical self-supervised depth estimation relies on photometric consistency, which is unreliable under noise. The authors shift the masked LF to the central view across depth planes to build a Plane Sweep Volume (PSV), then generate a focal stack: $\(FS=\frac{1}{|M|}\sum_{u=1}^{U}\sum_{v=1}^{V}PSV_{uv}\)$ Since the central view is masked, the refocused image does not contain the original central view's information. Thus, the noisy central view can safely supervise the depth module \(f_\sigma\) without learning an identity mapping: $\(\ell_{depth}=\Big\|\sum_{d}^{D}[f_\sigma(PSV)\odot FS]_d-\tilde I_c\Big\|_1\)$ The resulting depth probabilities weigh the PSV to guide the latent representation. Total loss: \(\ell_{total}=\ell_{blind}+\alpha\ell_{rc}+\beta\ell_{depth}\) with \(\alpha=0.3, \beta=0.1\).

Loss & Training

Implemented in PyTorch with Adam (\(\beta_1=0.9, \beta_2=0.999\)), batch size 20, fixed learning rate \(10^{-4}\). Trained for \(10^5\) iterations on an NVIDIA A5000. Training utilizes central 7×7 views cropped into 48×48 patches. Evaluation metrics: PSNR / SSIM.

Key Experimental Results

Main Results

Trained on 16 scenes from the HCI dataset and evaluated on 8 scenes, with zero-shot testing on HCIold (5 scenes) and DLFD (30 scenes). Compared against single-image blind-spot (B2U, ZSBlind), supervised LF (APA, DRLF, SRDNet, MSPNet), and self-supervised LF (V2V3D).

Synthetic data (sRGB, PSNR↑/SSIM↑):

Noise	Dataset	Ours	SRDNet (Supervised SOTA)	V2V3D (Self-supervised)
σ=20	HCI	37.86 / 0.960	38.17 / 0.925	36.21 / 0.909
σ=20	HCIold	37.72 / 0.953	36.78 / 0.937	35.97 / 0.896
σ=20	DLFD	36.04 / 0.937	35.44 / 0.901	33.89 / 0.891
σ=50	HCI	34.51 / 0.911	34.73 / 0.901	33.37 / 0.901
σ=50	DLFD	33.19 / 0.883	32.27 / 0.884	33.04 / 0.883

Ours outperforms self-supervised methods and most supervised methods. Supervised SRDNet drops significant performance on HCIold/DLFD due to training distribution shifts, whereas Ours demonstrates superior generalization.

Unseen noise types (trained on Gaussian σ=20, tested on Poisson λ=30 / Uniform [−50,50]):

Noise	Dataset	Ours	V2V3D	SRDNet
Poisson	HCI	36.09 / 0.932	34.17 / 0.889	30.64 / 0.763
Poisson	DLFD	34.51 / 0.917	33.85 / 0.867	29.26 / 0.726
Uniform	HCI	36.46 / 0.943	33.33 / 0.885	29.96 / 0.670
Uniform	DLFD	35.13 / 0.917	32.78 / 0.898	27.83 / 0.659

Supervised methods fail on unseen noise (SRDNet drops to 27–30 dB), while Ours maintains robustness, proving the architecture learns intrinsic geometry rather than overfitting noise.

Ablation Study

Conducted on HCI with Gaussian σ=20.

Mask Ratios & Branches (Table 5):

Config	Mask Ratio	PSNR	SSIM	FLOPs(T)
4-Branch (90°)	25%	37.86	0.960	5.0
4-Branch (90°)	50%	36.91	0.942	5.0
4-Branch (90°)	75%	36.10	0.926	5.0
2-Branch (180°)	50%	35.97	0.925	2.5

Component Breakdown (Table 6):

Config	PSNR	SSIM	Description
Base (Random Mask)	30.56	0.812	Cannot learn cross-view relations
+ GIM	36.47	0.915	Adapts blind-spot to LF, +5.91 dB
+ LRV	37.27	0.930	Collaborative denoising
+ \(\ell_{depth}\)	37.55	0.954	Depth guidance
+ \(\ell_{rc}\) (Full)	37.86	0.960	Consistency constraints

Key Findings

• Replacing GIM with random masking leads to a crash (30.56 dB), proving GIM is the critical component (+5.91 dB).
• 25% masking with 4 branches is the optimal trade-off between complexity and accuracy. Larger masking ratios reduce available context and lower performance.
• Effectiveness on Light Field Microscopy (LFM): Using PSF rendering, the method preserves more details compared to V2V3D.

Highlights & Insights

• Blind-spot to Blind-view Transition: Single-image methods are limited by spatial continuity assumptions. LF multi-view signals provide stronger supervision because noise is rarely consistent across geometries.
• GIM + Weight Sharing: Utilizing rotational invariance to collapse a multi-network problem into a single 4-branch network is an elegant design that saves parameters and promotes consistency.
• Pseudo-Supervision for Depth: Using the noisy central view to supervise the refocusing module (which excludes the central view's own info) is a clever self-supervised signal for extracting geometry from noise.

Limitations & Future Work

• The framework is strictly bound to light field geometry (disparity/refocusing) and rotational invariance, making it hard to transfer to monocular or sparse multi-view setups.
• The 4-branch architecture requires four forward passes, introducing higher inference overhead compared to single-branch methods.
• Geometric consistency might break in non-regular angular layouts or highly non-Lambertian scenes.

Related Work & Insights

• vs. Single-image Blind-spot: Methods like B2U often blur textures or fail at edges; the blind-view approach leverages multi-view consistency to recover details in weak-texture areas.
• vs. V2V3D: V2V3D uses two independent networks and relies on PSF; LF-BVN uses weight-sharing and consistency losses, resulting in fewer artifacts and broader applicability to standard LFs.
• vs. Supervised Methods: LF-BVN avoids the need for clean data and shows significantly better generalization once the noise distribution changes.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Elegant extension of blind-spot to blind-view with geometric invariance.
• Experimental Thoroughness: ⭐⭐⭐⭐ Extensive datasets and noise types; downstream depth evaluation is a plus.
• Writing Quality: ⭐⭐⭐⭐ Clear theoretical derivation, though some details are in supplementary materials.
• Value: ⭐⭐⭐⭐ First self-supervised framework applicable to both standard LF and LFM with high generalization.

Related Papers

• [CVPR 2026] TM-BSN: Triangular-Masked Blind-Spot Network for Real-World Self-Supervised Image Denoising
• [CVPR 2026] Self-Diffusion Driven Blind Imaging
• [CVPR 2026] Next-Scale Prediction: A Self-Supervised Approach for Real-World Image Denoising
• [CVPR 2026] Convexity-Aware Noise Calibration: A Self-Supervised Framework for Noise-Level-Unknown Image Denoising
• [CVPR 2026] SelfHVD: Self-Supervised Handheld Video Deblurring