A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion

Kai Wang; Yiming Gong

REVIEW 4 major objections 5 minor 25 references

A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single resolution-conditioned network, CondLAformer, fuses orthogonal line-scan microscopy images at near-specialist quality across all slit widths and generalizes smoothly to unseen intermediate configurations.

desk verdict A promising conditioning recipe and a coherent internal ablation, but the headline numbers and the specialist comparison need major correction before the central claim is believable. read the letter →

arxiv 2608.03107 v1 pith:CK72F4OL submitted 2026-08-04 cs.CV

Yiming Gong , Kai Wang This is my paper · ORCID

classification cs.CV

keywords imagefusionline-scanningmicroscopylinearattentionconditionalrestorationFiLMconditioninganisotropicresolutionrankenhanced

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that one trained network can replace a library of per-configuration models for fusing orthogonal line-scan microscopy images. The key idea is to condition the network on the resolving-power ratio $\rho = 0.727/s$, a continuous physical parameter derived from a separable point-spread-function model, using FiLM and an adaptive attention mechanism. If true, a microscopy facility with variable slit settings would need a single model rather than dozens of specialists, and the model would handle arbitrary slit values, including ones never seen in training. The paper's central claim is that conditioning resolves a catastrophic ambiguity: unconditioned multi-slit training collapses to 24.3 dB PSNR while per-slit specialists lose 4-9 dB outside their training setting, whereas the conditioned network maintains high fidelity across the full range.

What carries the argument

The central identity is the resolving-power ratio $\rho = \Delta_{\text{high}}/\Delta_{\text{low}} = 0.727/s$, derived from a separable PSF model whose slit-broadened profile has FWHM scaling linearly with slit number (fitted constant $\alpha = 9.83$ px). The argument runs on three coupled mechanisms: FiLM modulates LayerNorm outputs with per-level scale and shift parameters generated from $\rho$, letting one network index the full family of restoration strategies; Adaptive RELA replaces the fixed rank-enhancement convolution with three parallel dilated depthwise convolutions (3 px, 13 px, 25 px receptive fields) blended by $\rho$-conditioned softmax weights, and multiplies the linear attention query by a learnable temperature $\tau(r) \geq 0.5$ to control selectivity; and a physics-grounded simulation pipeline generates dense training triplets at 12 slit values from the separable PSF, creating a dataset where conditioning variables are known exactly.

What would settle it

Measure the actual 2-D PSF of a line-scanning confocal microscope across the slit range $s \in [8,52]$, compute its rank-1 energy ratio (does separability hold?) and the FWHM-versus-slit slope (is it linear?), then run CondLAformer on real orthogonal line scans; if the 34-40 dB PSNR or the smooth interpolation curve fails on real data, the simulated-PSF assumption is the cause.

Watch

Extended reading notes

Core claim

The paper claims that feeding the resolution ratio $\rho \in (0,1]$, computed from slit width as $\rho = 0.727/s$, into a linear-attention transformer through Feature-wise Linear Modulation (FiLM) turns multi-configuration image fusion from an ill-posed, configuration-ambiguous problem into a well-behaved continuous family of restoration tasks. The architecture, CondLAformer, augments Rank Enhanced Linear Attention (RELA) with two physics-motivated changes: multi-scale dilated depthwise convolutions whose blend weights are conditioned on $\rho$, and a learned attention temperature $\tau(r) \geq 0.5$ that sharpens attention under severe anisotropy. On a simulated dataset spanning 15 slit configurations, the full model achieves 34-40 dB PSNR across configurations, outperforming both unconditioned multi-slit training (collapse to 24.3 dB) and per-slit specialists (which lose 4-9 dB off their training configuration), and interpolates smoothly to unseen slit values. Ablations attribute +11.5 dB to FiLM conditioning, a further gain to global linear attention capturing long-range directional correspondences, and up to +2 dB to the adaptive temperature in the near-isotropic regime.

Load-bearing premise

The entire training and evaluation pipeline uses simulated point-spread functions that are assumed to be separable and to widen linearly with slit number, and the paper states that validation on experimentally acquired orthogonal line scans is still future work.

Editorial extensions

If this is right

A single CondLAformer model can replace the per-configuration model library for line-scanning fusion across the practical slit range $s \in [8,52]$, including interpolated settings never seen during training.
Conditioning on a continuous physical parameter (here $\rho$) resolves the multi-configuration ambiguity that collapses naive multi-task training by 14 dB, turning the network into a smooth family of restoration functions.
The adaptive attention temperature provides a real gain (up to 2 dB) specifically where anisotropy is subtle and complementary signals are weak, which is the regime where fixed attention mechanisms fail.
The learned behavior suggests a separation of concerns: depthwise convolutions provide a fixed local prior, FiLM acts as a soft bypass between near-identity and aggressive fusion, and the attention mechanism itself implements the $\rho$-adaptive receptive field.
The framework and its ablations establish that global linear attention with explicit conditioning is sufficient to recover long-range directional correspondences in line-scan fusion at linear computational cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The FiLM-on-ratio design likely transfers to other imaging restoration tasks with continuously variable acquisition parameters (e.g., defocus blur, noise level, PSF width), where the same ambiguity-collapse failure would occur without conditioning.
The learned 'soft bypass' behavior, where the near-isotropic configuration triggers the strongest FiLM modulation, suggests the network is effectively learning a continuous gate between fusion and pass-through; a testable prediction is that a hard-residual formulation would concentrate the gating at the same layer (Enc2).
If real experimental validation confirms the separability and linear-FWHM assumptions, the same conditioning scheme could be applied to 3-D volumetric acquisition by replacing the scalar $\rho$ with a vector of axis-specific resolution parameters, as the paper hints; evaluating that extension on real volumetric data would be the natural next step.
The reported 34-40 dB figures come from simulated data; a quick, cheap falsification would be to train the same architecture on real measured PSF kernels (the companion study's measured set) and compare the per-slit PSNR curve, which would isolate the domain-gap contribution.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is worth a look. The core idea is a single network, conditioned continuously on the resolving-power ratio rho via FiLM, that fuses orthogonal line scans across a range of slit widths. That is a sensible and well-motivated direction, and the architecture choices are grounded in the physics. The separable PSF analysis is the strongest part: the alpha=9.83 px fit, the rank-1 approximation capturing >99.7% of the energy, and the 48.3 dB reconstruction of measured kernels are concrete and reproducible. The ablations are also internally coherent: FiLM resolves the multi-slit ambiguity, and the adaptive temperature helps most in the small-slit regime. This is a real step toward a unified model in microscopy restoration.

But the paper overclaims in ways that matter. The abstract says 34-40 dB across configurations, but Table 2 caps at 34.5 dB with a 32.6 dB average. That is not a small discrepancy; it is the headline result. The qualitative section's 40-44 dB figures look like per-image examples, not the benchmark.

The bigger problem is Table 5. The stress-test note is right: the specialist comparison mixes PSF models and training budgets. The specialists were trained on measured PSFs, while CondLAformer was trained on Gaussian-approximated PSFs, and the table's note about test data using the specialists' measured PSFs puts CondLAformer in a different domain. On top of that, the specialists' at-training PSNRs are 23-26 dB, far below the companion study's reported 31.6 dB at the same rho. That strongly suggests the specialists are undertrained or evaluated under incompatible conditions. The '4-9 dB loss' and the unified model's superior transfer are not established by this table.

Also, the entire evaluation is simulation-based, generated using the same separable PSF model whose scaling was fitted to measured kernels. Section 5.4 admits experimental validation is future work. That is fine as a limitation, but it means the 34-40 dB claim is a simulator claim, not a microscopy claim, and the 48.3 dB figure is calibration, not external validation. No error bars or seeds are reported, and no code or data are released.

Who is this for? People working on conditional restoration and microscopy fusion will find the conditioning recipe and the PSF analysis useful. The paper deserves a serious referee, but it needs major revision: rescope the claims to the simulated setting, fix Table 5 with matched PSF models and training budgets, and report variance. If the specialist comparison is corrected and the claims are aligned with Table 2, this could be a solid contribution. I would not cite it in its current form.

Referee Report

4 major / 5 minor

Summary. The paper proposes a unified, resolution-conditioned deep-learning framework (CondLAformer) for fusing two orthogonal line-scanning microscopy images into an isotropic estimate. The network is conditioned on the resolving-power ratio rho = 0.727/s via FiLM layers, and extends Rank Enhanced Linear Attention with ratio-conditioned multi-scale depthwise convolutions and a learnable attention temperature. Training data are generated from a separable PSF model with a fitted linear FWHM scaling constant alpha = 9.83 px, and the model is evaluated on simulated test data at 12 trained and 6 interpolated slit configurations. The authors report that unconditioned multi-slit training collapses to about 24.3 dB PSNR, while their conditioned model achieves 31-35 dB in Table 2, with qualitative examples showing larger gains; they also argue that per-slit specialists lose 4-9 dB outside their training configuration.

Significance. If the central claims held, this would be a practically valuable result: a single trained network replacing a library of per-configuration models for orthogonal line-scan fusion, with smooth interpolation over unseen slit widths. The architectural ideas are well motivated: FiLM conditioning on a continuous physical parameter is a sensible way to handle configuration ambiguity, and the adaptive RELA mechanism is a plausible extension of linear attention. The paper also includes a useful dense multi-configuration benchmark and a clear ablation hierarchy (FiLM, global linear attention, adaptive temperature). These are genuine strengths. However, the quantitative evidence as presented does not yet support the headline claims: the reported 34-40 dB range is not reproducible from Table 2, the specialist comparison in Table 5 is confounded by inconsistent PSF-model and training-budget conditions, and the entire evaluation is internal to the simulator whose validity for real line-scanning data is explicitly deferred to future work. The contribution is therefore potentially valuable but currently overstated.

major comments (4)

[§4.7, Table 5] The central specialist-vs-unified comparison is confounded. Section 4.5 states that the single-configuration specialists in companion work [7] were trained on real measured PSFs, while CondLAformer was trained on Gaussian-approximated PSFs. Table 5 then says the evaluation uses test data generated with the 'exact same PSF' used during specialist training, i.e., measured .mat PSFs. If all rows of Table 5 use measured-PSF test data, CondLAformer is evaluated out-of-domain; if the specialist rows and the CondLAformer rows use different PSF models, the comparison is apples-to-oranges. In addition, the specialist training-slit values (26.1 dB at s=14, 25.1 at s=26, 23.1 at s=52) are far below the companion study's reported single-slit specialist performance, such as 31.6 dB at rho=0.052, which corresponds to s=14. This suggests the specialists in Table 5 are undertrained or evaluated under incompatible conditions. The claimed '4-9 dB loss' and the conclusion that CondLAformer maintains >35 dB across configurations are therefore not established by this table.
[Abstract; §4.2, Table 2] The headline '34-40 dB PSNR across configurations' is not supported by the paper's own main table. Table 2 reports a maximum of 34.5 dB (at s=8) and an average of 32.6 dB for CondLAformer with Adaptive RELA. Section 4.2 also states that CondLAformer v2 achieves 35.8 dB over the unconditioned DualUnet, but Table 2 lists 31.2 dB average and 32.9 dB maximum for the FiLM-only variant. The qualitative text in §4.3 mentions '>40 dB' and '44+ dB' results that are not reconciled with the per-slit PSNR numbers. Since the abstract's numerical range is a central claim, the authors should either report the Table 2 values in the abstract or describe the precise experimental protocol under which 34-40 dB was obtained.
[§3.1, §5.4] The evaluation is entirely internal to the separable PSF simulator. All training triplets are generated with the same model whose alpha = 9.83 px was fitted from six measured PSF kernels, and all test inputs, including the unseen-configuration generalization test, are generated from that same simulator. The 48.3 dB figure is the reconstruction accuracy of the separable approximation against the measured 2-D PSF, not an external task benchmark. Section 5.4 explicitly concedes that validation on experimentally acquired orthogonal line scans is future work. Consequently, the claims about smooth generalization to unseen configurations and about 34-40 dB performance should be stated as simulator-conditional, and the paper should either provide out-of-simulator validation or substantially soften the practical-deployment framing.
[§4.2, Tables 2-5] No error bars, confidence intervals, or repeated-seed results are reported anywhere in the main tables. Several central comparisons rest on small differences, such as the roughly 1-2 dB advantage of Adaptive RELA over fixed RELA at small slits, and the large claimed specialist gap in Table 5. Because single-run PSNR values can vary with initialization and data sampling, the authors should report means and standard deviations over at least three training seeds for the main comparisons, or otherwise justify why seed dependence is negligible.

minor comments (5)

[§4.6, Table 4] The 'w/o Multi-slit (per-slit)' row reports '26.1/25.1/23.1' without a column header or explanation; clarify whether these are training-configuration PSNRs, cross-configuration PSNRs, and which slit values they correspond to.
[§4.4, §4.7] The number of configurations is inconsistent: Table 3 refers to a 'controlled 15-slit dataset,' §4.7 mentions '16 configurations,' and Table 2 evaluates 18 slit values (12 trained and 6 interpolated). Please reconcile these counts.
[Eq. (2)] The constant 0.727 in rho(s) = 0.727/s is introduced without derivation; specify the wavelength, numerical aperture, and pixel size used, and show how the confocal FWHM in pixels follows from the alpha = 9.83 px fit so that Eq. (2) and Eq. (10) can be checked independently.
[§5.1, Eq. (20)] The finding that the multi-scale DWConv blend weights are essentially rho-invariant (w_med approximately 0.58 across configurations) appears to undercut the stated purpose of ratio-conditioned blending in Eq. (20); explain what the ratio-conditioned softmax contributes if the learned weights do not vary with rho.
[Abstract] There are formatting and typographical issues, including 'aunified,resolution-conditionedimage fusion framework' missing spaces; a careful proofread is needed.

Circularity Check

2 steps flagged · score 4.0 of 10

Fitted PSF simulator is both the training and test bed; the 48.3 dB 'verification' is a fit residual, so the 34-40 dB headline is in-simulator, while the specialist comparison is additionally inconsistent.

fitted input called prediction [Sec. 3.1 (PSF calibration); Sec. 3.6 (training data); Sec. 4.1-4.2 (test data and results)]
"We determined α precisely through SVD analysis of the measured 2-D PSF kernels at six configurations (s ∈ {1,3,6,14,26,52}): a linear fit yields α = 9.83 pixels ... enabling synthesis of training data at arbitrary slit values through simple 1-D profile rescaling (48.3 dB reconstruction accuracy vs. the measured 2-D PSF)."

The 48.3 dB 'reconstruction accuracy' is computed against the same six measured kernels used for the linear fit of α, so it is a training-fit residual presented as verification. All training triplets (Sec. 3.6) and the 18-slit test set (Sec. 4.1) are then generated with this same fitted separable PSF model. Consequently the headline '34-40 dB PSNR across configurations' and the smooth interpolation claim are performance measures on the model's own fitted forward process; they do not predict performance on independent measured image data, which Sec. 5.4 concedes is future work.
self citation load bearing [Sec. 5.2 Step 0; Sec. 2.2]
"In companion work [7], we compared Attn-DualUnet against an identical Base-DualUnet lacking directional attention (XDA/YDA replaced by plain concatenation). The attention benefit grows monotonically with anisotropy severity ... We therefore adopt Attn-DualUnet as the baseline CNN architecture for all subsequent experiments."

The premise that directional attention encodes a physical prior, and hence the choice of Attn-DualUnet as the baseline for all comparisons, rests on [7], an under-review WACV paper by the same authors. This is not machine-checked or externally reproduced, so it is self-citation used as load-bearing motivation. It does not, however, force the main CondLAformer results: the FiLM/Adaptive-RELA ablations are trained in-house, so the central architectural claim retains independent content.

full rationale

The paper's core architectural derivation is not circular: the ablations in Table 4 are trained and evaluated in-house on the same simulated data, and the FiLM and Adaptive-RELA improvements are genuine empirical outcomes within that simulator. The main circularity is that the simulator itself is the fitted object: α = 9.83 px is fit to six measured PSF kernels, the same separable model generates all training and test triplets, and the 48.3 dB 'verification' is a fit residual rather than independent validation. Thus the 34-40 dB headline and smooth interpolation claims are in-simulator results, as Sec. 5.4 concedes that real experimental validation is future work. Self-citation of companion work [7] provides the baseline-architecture premise, but the central CondLAformer comparisons are not reduced to that citation. A separate correctness risk, noted but not counted as formal circularity, is the Table 5 specialist comparison: Sec. 4.5 states that direct numerical comparison is 'not applicable' because specialists were trained on measured PSFs while CondLAformer was trained on Gaussian-approximated PSFs, yet Sec. 4.7 then uses those specialist numbers as the key comparison; the specialists' at-training values (26.1 dB at s=14) also fall far below the companion study's 31.6 dB at the same rho, suggesting undertraining or incompatible evaluation conditions. Overall, partial circularity is present because the fitted PSF model is both the calibration input and the benchmark generator, so the external-facing performance claims are not yet supported by independent data.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central performance claims depend on two calibrated constants (alpha and 0.727), on the separability and linear-scaling assumptions of the PSF model, and on the assumption that simulation-based training transfers to real microscopes. No new physical entities are introduced; the architecture components are learned network parameters rather than fitted physical constants.

free parameters (2)

alpha (PSF FWHM scaling constant) = 9.83 px
Obtained by linear fit to SVD analysis of measured 2-D PSF kernels at six slit values. It sets FWHM_low approximately equal to alpha times s and therefore controls the simulated blur width used for all training and test data (Sec. 3.1).
Constant 0.727 in rho(s) = 0.727/s = 0.727
Determined from the confocal PSF width at NA=0.8 together with alpha. It converts slit number into the resolving-power ratio that conditions the network, so the physics-grounded conditioning rests on this calibration (Eq. 2, Sec. 3.1).

assumptions (5)

domain assumption The line-scanning PSF is separable: hx = h_high(x) h_low(y), hy = h_low(x) h_high(y), with a rank-1 approximation capturing more than 99.7% of the PSF energy (Sec. 3.1).
All training data generation and the claimed 48.3 dB PSF reconstruction accuracy depend on this separability. The paper verifies it on measured PSF kernels, but not on end-to-end image data.
domain assumption The slit-broadened PSF width scales linearly with the slit parameter: FWHM_low approximately equal to alpha times s, with alpha = 9.83 px (Eq. 10, Sec. 3.1).
The linear scaling is fitted from six measured configurations and is used to synthesize training data at all 12 slit values and 6 interpolated values.
domain assumption The ground truth for fusion is O convolved with h_point, and the two orthogonal scans O convolved with hx and O convolved with hy contain enough complementary information to recover it (Eqs. 3-4 and 13).
The paper gives no information-theoretic or experimental proof that two orthogonal line scans are sufficient for near-isotropic recovery across the tested slit range.
domain assumption Simulated training data from Bio-TISR, Bio-LFSR, 3D RCAN, CARE, BioSR+, plus synthetic structures, are representative of real line-scanning microscopy images and degradations (Sec. 3.6).
The paper acknowledges in Sec. 5.4 that generalization to entirely unseen specimen types and to experimental scans remains unvalidated.
standard math Standard linear-attention kernel reordering, RELA, FiLM, U-Net, and AdamW training behave as described in the cited literature (Eqs. 5-6, 14-16).
These are taken from prior work and are not verified by the paper beyond its own experiments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion." pith.science (2026). https://pith.science/paper/CK72F4OL

@misc{pith2026260803107,
  author       = {Pith},
  title        = {Pith review of: A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CK72F4OL}},
  note         = {Machine review of arXiv:2608.03107}
}

read the original abstract

Laser line-scanning microscopy enables fast volumetric imaging but produces anisotropic lateral resolution. Orthogonal line scans provide complementary directional information that can recover near-isotropic resolution, yet existing deep-learning methods require a separate model for each optical configuration. We present a unified, resolution-conditioned fusion framework based on Rank Enhanced Linear Attention (RELA). Feature-wise Linear Modulation (FiLM) conditions the network continuously on the resolving-power ratio, enabling one model to adapt across slit widths. We further introduce Adaptive RELA, which replaces fixed-kernel rank enhancement with ratio-conditioned multi-scale depthwise convolutions and uses a learnable attention temperature to adjust selectivity with degradation severity. Training data spanning multiple slit configurations are generated using a physics-grounded separable point-spread-function model verified against measured optical data at 48.3 dB accuracy. The resulting model achieves 34-40 dB PSNR across configurations, whereas unconditioned multi-slit training collapses to 24.3 dB and per-slit specialists lose 4-9 dB outside their training setting. It also generalizes smoothly to unseen intermediate configurations without interpolation artifacts. Ablations show that FiLM resolves configuration ambiguity, global linear attention captures long-range directional correspondences, and adaptive temperature yields an additional 2 dB in the challenging near-isotropic regime, where complementary signals are weak.

Figures

Figures reproduced from arXiv: 2608.03107 by the authors.

**Figure 1.** Motivation: directional attention is essential under severe anisotropy. Base-DualUnet (no directional attention) vs Attn-DualUnet [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Architecture overview of CondLAformer v3. (a) The overall U-Net topology with four encoder-decoder levels, skip connections, and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Quantitative performance summary. (a) Per-configuration PSNR for CondLAformer v2 (FiLM only, blue) and v3 (Adaptive RELA, [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Visual comparison across five resolution ratios on an endoplasmic reticulum (ER) specimen. Rows: Ground Truth, Attn-DualUnet [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 5.** Figure 5: Directional attention analysis (log scale) across four slit configurations. X-only input (top row) produces vertically-spreading [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 24 canonical work pages

[7]

[Authors], Regime-aware fusion of orthogonal line scans for isotropic resolution recovery in fast fluores- cence microscopy, in: WACV, 2027, under review

work page 2027
[1]

Pawley, Handbook of Biological Confocal Mi- croscopy, 3rd Edition, Springer, 2006

J. Pawley, Handbook of Biological Confocal Mi- croscopy, 3rd Edition, Springer, 2006

work page 2006
[2]

W. C. P. Wasser, et al., Live imaging of muscles in Drosophila metamorphosis, Methods 96 (2016) 103– 117. 11

work page 2016
[3]

C. J. R. Sheppard, X. Q. Mao, Confocal microscopes with slit apertures, Journal of Modern Optics 35 (1988) 1169–1185

work page 1988
[4]

Fiolka, A

R. Fiolka, A. Stemmer, Y. Belyaev, Virtual slit scan- ning microscopy, Histochemistry and Cell Biology 128 (2007) 499–505

work page 2007
[5]

Swoger, P

J. Swoger, P. Verveer, K. Greger, J. Huisken, E. H. K. Stelzer, Multi-view image fusion improves resolution in three-dimensional microscopy, Optics Express 15 (2007) 8029–8042

work page 2007
[6]

Y. Xue, J. Park, I. S. Lee, Multiline orthogonal scan- ning temporal focusing (mostf) microscopy for scat- tering reduction in in vivo brain imaging, Scientific Reports 14 (2024) 10954

work page 2024
[8]

Y. Zhi, R. Lu, B. Wang, Q. Zhang, X. Yao, Rapid super-resolution line-scanning microscopy through virtually structured detection, Optics Letters 40 (2015) 1683–1686

work page 2015

Show all 25 references

[9]

Chen, et al., Meta-rlls-vsim: lattice light-sheet virtual structured illumination microscopy, Nature MethodsIn press (2025)

Z. Chen, et al., Meta-rlls-vsim: lattice light-sheet virtual structured illumination microscopy, Nature MethodsIn press (2025)

2025
[10]

Stathaki, Image Fusion: Algorithms and Applica- tions, Academic Press, 2011

T. Stathaki, Image Fusion: Algorithms and Applica- tions, Academic Press, 2011

2011
[11]

H. Li, X. J. Wu, DenseFuse: A fusion approach to infrared and visible images, IEEE Trans. Image Pro- cessing 28 (5) (2019) 2614–2623

2019
[12]

J. Ma, H. Xu, J. Jiang, X. Mei, X. Zhang, DDc- GAN:Adual-discriminatorconditionalgenerativead- versarial network for multi-resolution image fusion, in: IEEE Trans. Image Processing, Vol. 29, 2020, pp. 4980–4995

2020
[13]

Weigert, U

M. Weigert, U. Schmidt, T. Boothe, et al., Content- aware image restoration: pushing the limits of fluo- rescence microscopy, Nature Methods 15 (2018) 1090– 1097

2018
[14]

J. Chen, H. Sasaki, H. Lai, et al., Three-dimensional residual channel attention networks denoise and sharpen fluorescence microscopy image volumes, Na- ture Methods 18 (2021) 678–687

2021
[15]

Z. Liu, A. Liu, Y. Yang, C. Fang, Fusion method for cell refractive-index microscopy images and bright- field microscopy images based on convolutional neural networks, Laser & Optoelectronics Progress 58 (2021) 2217001

2021
[16]

Liang, J

J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, R. Timofte, SwinIR: Image restoration using Swin transformer, in: ICCVW, 2021, pp. 1833–1844

2021
[17]

5728–5739

S.W.Zamir, A.Arora, S.Khan, etal., Restormer: Ef- ficient transformer for high-resolution image restora- tion, in: CVPR, 2022, pp. 5728–5739

2022
[18]

Katharopoulos, A

A. Katharopoulos, A. Vyas, N. Pappas, F. Fleuret, Transformers are RNNs: Fast autoregressive trans- formers with linear attention, in: ICML, 2020, pp. 5156–5165

2020
[19]

Y. Ai, H. Huang, T. Wu, Q. Fan, R. He, Breaking complexity barriers: High-resolution image restora- tion with rank enhanced linear attention, arXiv preprint arXiv:2505.16157 (2025)

2025 arXiv
[20]

Zhang, W

K. Zhang, W. Zuo, L. Zhang, FFDNet: Toward a fast and flexible solution for CNN-based image denoising, IEEE Trans. Image Processing 27 (9) (2018) 4608– 4622

2018
[21]

Potlapalli, S

V. Potlapalli, S. W. Zamir, S. Khan, F. S. Khan, PromptIR: Prompting for all-in-one blind image restoration, in: NeurIPS, 2023

2023
[22]

Zheng, et al., DiffUIR: Controlling diffusion models for universal image restoration, in: CVPR, 2024

Z. Zheng, et al., DiffUIR: Controlling diffusion models for universal image restoration, in: CVPR, 2024

2024
[23]

Perez, F

E. Perez, F. Strub, H. De Vries, V. Dumoulin, A. Courville, FiLM: Visual reasoning with a general conditioning layer, in: AAAI, 2018, pp. 3942–3951

2018
[24]

J. Hu, L. Shen, G. Sun, Squeeze-and-excitation net- works, in: CVPR, 2018, pp. 7132–7141

2018
[25]

analyze-then-integrate

C. Qiao, D. Li, Y. Guo, et al., Evaluation and de- velopment of deep neural networks for image super- resolution in optical microscopy, Nature Methods 18 (2021) 194–202. 12 Figure 3: Quantitative performance summary. (a) Per-configuration PSNR for CondLAformer v2 (FiLM only, b...

2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [7]

[Authors], Regime-aware fusion of orthogonal line scans for isotropic resolution recovery in fast fluores- cence microscopy, in: WACV, 2027, under review

work page 2027

[2] [1]

Pawley, Handbook of Biological Confocal Mi- croscopy, 3rd Edition, Springer, 2006

J. Pawley, Handbook of Biological Confocal Mi- croscopy, 3rd Edition, Springer, 2006

work page 2006

[3] [2]

W. C. P. Wasser, et al., Live imaging of muscles in Drosophila metamorphosis, Methods 96 (2016) 103– 117. 11

work page 2016

[4] [3]

C. J. R. Sheppard, X. Q. Mao, Confocal microscopes with slit apertures, Journal of Modern Optics 35 (1988) 1169–1185

work page 1988

[5] [4]

Fiolka, A

R. Fiolka, A. Stemmer, Y. Belyaev, Virtual slit scan- ning microscopy, Histochemistry and Cell Biology 128 (2007) 499–505

work page 2007

[6] [5]

Swoger, P

J. Swoger, P. Verveer, K. Greger, J. Huisken, E. H. K. Stelzer, Multi-view image fusion improves resolution in three-dimensional microscopy, Optics Express 15 (2007) 8029–8042

work page 2007

[7] [6]

Y. Xue, J. Park, I. S. Lee, Multiline orthogonal scan- ning temporal focusing (mostf) microscopy for scat- tering reduction in in vivo brain imaging, Scientific Reports 14 (2024) 10954

work page 2024

[8] [8]

Y. Zhi, R. Lu, B. Wang, Q. Zhang, X. Yao, Rapid super-resolution line-scanning microscopy through virtually structured detection, Optics Letters 40 (2015) 1683–1686

work page 2015