MultiFLARE Reconstruction

• MultiFLARE Reconstruction is a framework that restores flare-affected imaging data by modeling overlapping flare sources and estimating their spatial and parametric structures.
• It integrates methodologies such as Bayesian sequential Monte Carlo for solar imaging and hierarchical deep neural networks for lens flare removal.
• The approach is validated through quantitative metrics like PSNR, SSIM, and precise localization measures, ensuring high fidelity across astrophysics, photography, and 3D applications.

MultiFLARE Reconstruction encompasses a spectrum of methods for inferring and restoring flare-affected imaging data, with applications in solar astrophysics (multi-flare X-ray imaging), computational photography (lens flare removal), and 3D scene reconstruction. These approaches address the problem of disentangling overlapping flare sources, accurately estimating their spatial and parametric structure, and recovering uncorrupted image content in scenarios where flare artifacts introduce significant ambiguity. Notable frameworks include parametric Bayesian inference pipelines for solar flare localization (Sciacchitano et al., 2018), hierarchical deep learning architectures for lens flare separation (Huang et al., 4 Aug 2025, Jiang et al., 2024), and neural radiance field formulations for multi-view flare disentanglement (Matta et al., 2024). The following sections detail the mathematical models, optimization methods, and evaluation paradigms characterizing current MultiFLARE Reconstruction research.

1. Mathematical Modeling and Problem Formulation

Across domains, MultiFLARE Reconstruction is predicated on explicit forward models that describe the combination of underlying sources (flares) and their observable manifestations. For solar X-ray telescopes (e.g., RHESSI), the photon flux map $F(x, y)$ is decomposed as a sum of $N$ discrete geometric objects, each parameterized by type ($T_i$; e.g., circle, ellipse, loop) and shape/location parameters $\theta_i$ (Sciacchitano et al., 2018). The measurement model for visibilities is:

$$\nu_j = V_j(F) + \epsilon_j,\quad V_j(F) = \sum_{i=1}^N V_j(\theta_i; T_i),\quad \epsilon_j \sim \mathcal{N}(0, \sigma_j^2)$$

For lens flare in RGB images, the formulation is additive:

$I = I_0 \oplus F$

where $I$ is the observed image, $I_0$ the clean background, $F$ the flare component, and $\oplus$ denotes pixelwise addition with clipping to $[0,1]$ (Huang et al., 4 Aug 2025). In neural 3D rendering, the observed color is modeled as

$$C^{\text{obs}}(\mathbf{r}) = C^{\text{scene}}(\mathbf{r}) + C^{\text{flare}}(\mathbf{r}),$$

with view-dependent $C^{\text{flare}}$ explicitly modeled and masked (Matta et al., 2024).

Prior distributions (e.g., truncated Poisson for $N$, categorical for $T_i$) and likelihoods (Gaussian, cross-entropy) are defined over these latent variable spaces. Losses in learning-based models combine pixelwise error, perceptual differences (e.g., VGG/AlexNet feature distances), and structural similarity (SSIM).

2. Inference and Optimization Methodologies

Bayesian Sequential Monte Carlo for Multi-flare Parametric Imaging

In solar imaging, reconstruction proceeds via Adaptive Sequential Monte Carlo (ASMC), targeting the joint posterior of number, types, and parameters of sources:

$$p(N, \theta_{1:N} | d) \propto p(d | N, \theta_{1:N})\, p(N) \prod_{i=1}^N p(T_i) p(\theta_i|T_i)$$

The ASMC explores this posterior by tempering the likelihood over annealing steps $\gamma_i$, adaptively set so the effective sample size (ESS) remains stable. Birth/death, type-change, split/merge, and local parameter-update moves (via reversible-jump MCMC) enable transitions on the variable-dimensional state space. Rao–Blackwellization analytically marginalizes linear flux parameters to improve efficiency (Sciacchitano et al., 2018).

Deep Neural Architectures

DeflareMamba (Huang et al., 4 Aug 2025) employs a hierarchical U-shaped encoder–decoder, integrating Local-enhanced Residue State-Space Blocks (L-RSSB) and Hierarchical Residue State-Space Groups (H-RSSG). The local detail is preserved by depth-wise convolutional branches, while long-range dependencies are propagated by windowed state-space 2D recurrences. Hierarchical strided sampling in the decoder (H-RSSB) addresses sequence length limitations intrinsic to state-space models.

MFDNet (Jiang et al., 2024) decomposes images into low and high-frequency components using a Laplacian pyramid. The Low-Frequency Flare Perception Module (LFFPM) removes flare on the coarse band with a hybrid Transformer–CNN, while the Hierarchical Fusion Reconstruction Module (HFRM) fuses denoised low-frequency content back with high-frequency residuals for final image synthesis.

Multi-view Neural Flare Removal

GN-FR (Matta et al., 2024) extends radiance field inference by incorporating a Flare-occupancy Mask Generator (FMG), which segments flare-affected regions in each observation using a PSPNet backbone. The View Sampler (VS) and Point Sampler (PS) utilize these masks to exclude flare-corrupted rays, enabling view-transformer aggregation of only flare-free features during NeRF-based color prediction. Training leverages an unsupervised masking loss, enforcing fidelity exclusively on unflared pixels.

3. Algorithmic Pipelines

Representative pipelines comprise the following generic stages:

Stage	Example (RHESSI/solar)	Example (Vision/photography)
Data acquisition	Measured visibilities (Fourier domain)	RGB images (single or multi-view)
Source decomposition/model	Geometric parametric (C/E/L)	Pyramid, SSM, Transformer-CNN fusion
Optimization/inference	ASMC with RJ-MCMC moves	SGD on composite loss, SSM recurrence
Uncertainty/credibility	Posterior samples, credible intervals	Ablation, perceptual metrics
Output	Object list + uncertainty bands	Deflared image, flare component

In the ASMC pipeline, particles represent candidate sets of sources; resampling and MCMC transitions enable convergence to posterior modes. In DeflareMamba and MFDNet, forward passes through the U-Net or pyramid-based architecture enable end-to-end mapping from flare-corrupted to restored images via staged multi-scale operations. GN-FR requires preprocessing with FMG, carefully curated view selection, pointwise masking, and transformer-based fusion per ray (Matta et al., 2024).

4. Quantitative and Qualitative Evaluation

Evaluations benchmark both restoration fidelity and computational efficiency. Solar flare models report location errors of $\lesssim1$ arcsec and size errors $<10\%$ for low-noise synthetic and real RHESSI visibilities, with correct recovery of source multiplicity (mode of posterior $p(N)$ matches ground truth) (Sciacchitano et al., 2018). Posterior uncertainties (e.g., $95\%$ intervals in position/flux) are explicitly reported.

In photographic applications, metrics include PSNR, SSIM, and LPIPS on standardized datasets such as Flare7K and Flare7K++ (Huang et al., 4 Aug 2025, Jiang et al., 2024), with DeflareMamba achieving PSNR=27.78 dB, SSIM=0.899, and MFDNet achieving PSNR=26.98 dB, SSIM=0.895 on real-world sets. Efficiency is captured by GMACs, parameter counts, and inference speed: MFDNet requires 18.3 GMACs and 6.3M parameters for 512×512 inputs (Jiang et al., 2024). In GN-FR, multi-view PSNR gains ($\sim 4$ dB over single-view GNT) are realized on novel-view flare removal at the scene level (Matta et al., 2024).

Ablations indicate criticality of hierarchical scanning, local/global fusion, and view/mask selection for state-of-the-art performance.

5. Domain Extensions and Practical Considerations

The MultiFLARE paradigm generalizes across modalities and instruments. In solar imaging, ASMC adapts to other sparse Fourier-sampling X-ray telescopes (e.g., STIX) by altering only the analytic forward operator $V_j(\theta)$ (Sciacchitano et al., 2018). Particle-filter analogs enable real-time multi-energy or spectral sequence monitoring by propagating posteriors between time bins.

For vision tasks, MultiFLARE models support rigorous pipeline construction for high-fidelity content restoration, robust to diverse flare morphologies (streaks, rings, haze). DeflareMamba demonstrates downstream gains for object detection and cross-modal alignment, with documented increases in COCO mAP and CLIP/BLIP metrics when operating on flare-affected benchmarks (Huang et al., 4 Aug 2025).

NeRF-based approaches (GN-FR) establish unsupervised, generalizable solutions where ground truth is never directly accessible, leveraging multi-view context for robust flare removal. Dataset curation, especially collection and annotation of view-dependent flare-occupancy masks, is essential for high-quality learning (Matta et al., 2024).

6. Limitations and Prospects

Physical limitations persist where flare artifacts occlude all candidate observations, precluding recovery of true content even with multi-view reasoning (Matta et al., 2024). For sequential Monte Carlo, computational cost scales superlinearly with the number of particles and visibilities, though marginalization techniques and acceleration via GPU or parallel chains partly mitigate bottlenecks (Sciacchitano et al., 2018).

Future work may incorporate physically based flare optics into radiance field models, enable end-to-end co-training of segmentation and reconstruction, and extend to zero-shot unseen flare patterns via meta-learning. MultiFLARE Reconstruction continues to evolve as an intersection of probabilistic modeling, deep learning, and computational imaging, enabling principled recovery in flare-challenged scientific and photography data streams.