acceptodds
Under review as a conference paper at ICLR 2027

DORIAN: Zero-Shot Directionally Structured Noise Removal via Graph Neural Priors in Fourier Space

Abstract

Various scientific imaging modalities, for instance, light-sheet fluorescence microscopy, atomic force microscopy, and remote sensing, suffer from degradations that persist along a dominant orientation over image-scale distances. We refer to them as Directionally Persistent Structured Noise (DPSN). Three characteristics render DPSN difficult to remove: it is non-local in the image domain, its spectrum overlaps with that of the underlying image, and its appearance varies across modalities in which clean references are rarely available. Motivated by its compact representation in the Fourier domain, we propose DORIAN, a zero-shot framework that poses DPSN removal as a Fourier-domain spectral reconstruction problem. Specifically, we construct a polar graph over Fourier coefficients, on which a graph neural prior reconstructs corrupted coefficients from spectrally related uncorrupted ones. The graph prior is optimized on each degraded observation alone, requiring neither external training data nor clean targets. Experiments on natural images, microscopy, and remote sensing data show that DORIAN removes DPSN of diverse physical origins and appearances within a single framework.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image restoration, structured noise removal, zero-shot learning, Fourier domain, graph neural networks, test-time optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qo7aOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a zero-shot denoising method focusing on directionally persistent structured noise (DPSN) removal. This approach transforms noisy image into Fourier domain and uses the prior knowledge of the noise direction to locate the wedge region of possible corruption, and then uses GNN to reconstruct and recover the frequency coefficients. Additionally, this paper proposes to use directional gradient readout, anisotropic resampling, and residual transfer to preserve image structures and details. The experiments show that this approach achieves improvements on synthetic stripe noise. The method is also evaluated on real data through qualitative comparisons and the absolute correlation metric.

Critical strengths
  1. The problem of stripe noise removal has practical significance, and the method also uses a relatively reasonable structural prior. Transforming directionally structured noise to the Fourier domain for targeted restoration is intuitive.
  2. The synthetic experiments are relatively comprehensive and the performance on synthetic data is also good.
Critical weaknesses
  1. The real contributions of this paper are not clearly presented. Specifically, the first contribution presented in the Introduction largely overlaps with previous work DeStripe, which has proposed to use Fourier-domain GNN to recover frequency coefficients for stripe noise without clean data. Although this paper added some extensions and discussed the differences in the appendix, it doesn’t sufficiently clarify the limitations of DeStripe and how the proposed designs address them, and whether these extensions are novel. Therefore, the contribution, motivation and novelty of this paper are unclear.
  2. Experimental results have not sufficiently demonstrated the superiority of this method. For synthetic data, PSNR is only calculated on corrupted regions, and the full-image PSNR seems to be ignored. For real data, although some qualitative results show some improvements, the advantages are still not obvious and sufficient. The quantitative results mainly rely on absolute correlation which cannot directly reflect the stripe noise removal or the preservation of image structures. Additional no-reference metrics for stripe removal or other evaluation metrics suitable for scientific images would help.
Questions for authors
  1. What’s the major limitation of DeStripe and how does this paper address it? Could you clarify the main new contributions compared with DeStripe and how these changes improve the results?
  2. Could you report full-image PSNR for synthetic data and provide more quantitative results on real data?
  3. According to Eq.2, all non-DC coefficients within radius of 32 are marked as corrupted. But in Sec 3.2, the reference nodes should be uncorrupted and have a radial bandwidth of 5. Therefore, for corrupted nodes with a radius of at most 27, the reference set seems to be empty? How are these nodes handled in the implementation?
Reviewer_X34TOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed an image recovery scheme for images corrupted with directionally persistent structured noise (DPSN). The key idea is to construct a mask prior--a wedge in the polar Fourier coordinate (Section 3.1), which is optimized per degraded image. Processing is done via a graph neural net--a directed attributed graph with alternating FC and ReLU layers (Section 3.2). Then the readout is decomposed into directional gradients--along and orthgonal to DPSN direction, also via an MLP with alternating FC and ReLU layers (Section 3.3). The framework (called DORIAN) is optimized via a loss function composed of three terms: fidelity term processed by guilded filter at coarse resolution, a TV term, and an Hessian term (Section 3.4).

Critical strengths
  1. DPSN is a known hard problem, and the description of the problem, its inherent technical challenges, as well as discussion of previous works such as wavelet and contourlet looks thorough and comprehensible. Approaching the problem from a self-supervised scenario is indeed challenging, which is appreciated.

  2. Experimental results show modest gain over SOTA.

Critical weaknesses
  1. The overall scheme--composed of i) a defined wedge-shaped mask in 2D Fourier coordinate (equation (2), which BTW, should have an intersection instead of a union operator), ii) a directed graph propagation via a graph neural net (GNN) based on a directed attributed graph alternating between FC and ReLU layers, iii) anisotrophic spectral construction also alternating between FC and ReLU layers--- feel very hand-crafted and ad-hoc. Why an attributed graph (with hand-crafted connections and edge weights) for polar Fourier domain processing? The zero-shot optimization objective in equation (7), composed of three terms, is defined in a manner completely unrelated to the 2D Fourier domain mask wedge idea--the central thesis of the paper. The first term is the difference between the reconstructed image X^\hat{X} and observation YY, but processed by a guilded filter and compared at a coarse scale. Why guilded filter? Why coarse scale? The Second term is a total variation (TV) term computed along and orthogonal to the DPSN dominate direction. TV is a 30+ year old image prior (which the authors themselves criticized in Section 2 as being hand-crafted), while the emphasis of the paper is in the spectral domain. It looks out of place. The last term is a Hessian term, defined using operators ∇⊥,⊥\nabla_{\perp, \perp}, ∇∥,⊥\nabla_{\parallel, \perp} and ∇∥,∥\nabla_{\parallel, \parallel}. This came out of no where. Why is this important? The combination of these three terms look disconnected and under-motiviated.

  2. The combined use of graph neural net (GNN) with frequencies defined in the polar Fourier coordinates seems incompatible. If 2D / polar coordinates are defined as the data kernel, why would the processing engine not be defined in regular 2D / polar Fourier domain? If GNN is used, then the natual frequencies should be defined in the graph Fourier domain via eigen-decomposition of a graph shift operator like the graph Laplacian matrix. This is particularly jarring, because ultimately the chosen optimization objective in Section 3.4 has no graph frequency consideration at all, so why GNN?

Questions for authors
  1. Why is GNN chosen as a processing engine for a given wedge mask in the polar Fourier domain, the main thesis of the paper? How is the chosen attributed graph, with specified connectivity and edge weights, optimal in information propagation?

  2. Given GNN is used for processing, why is graph spectrum not used (defined as eigen-decomposition of graph shift operator like the graph Laplacian, common in the graph signal processing (GSP) literature), especially given the optimization objective has no spectral notion? For example, if you want to propagate information from the known region to the masked-out region, it's bascially a low-pass filter on the defined graph. Why is this not even considered?

Reviewer_BZtzOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript presents a zero-shot framework that removes directionally persistent structured noise by treating it as a Fourier-domain reconstruction problem, solved with a graph neural prior.

Critical strengths

The problem matters for scientific imaging, and the results across several modalities are promising.

Critical weaknesses

The first concern is methodological novelty. Fourier-domain removal of structured noise is well established, from early destriping of AFM images to anomaly detection in the Fourier domain for remote sensing. Of particular relevance is DeStripe, which already applies a self-supervised graph neural network to Fourier coefficients to remove stripes in light-sheet microscopy. A recent variational method also handles light-sheet, FIB-SEM and remote-sensing data within one framework. The authors should state precisely how the polar graph and the zero-shot prior go beyond these approaches, and should compare with them directly.

The second concern is the spectral assumptions. The method recovers corrupted coefficients from spectrally related, uncorrupted ones. If this relation is defined by adjacency in polar coordinates, the dependencies may be weak for complex or strongly oriented textures.

Moreover, the authors themselves acknowledge that the noise spectrum overlaps with that of the image. Replacing the affected coefficients is therefore likely to remove genuine image content as well. The manuscript should explain how corrupted coefficients are identified, how real directional structures are preserved, and how much information is lost under severe contamination.

The third concern is whether the graph prior is necessary and whether its assumptions hold in every modality. Optimising on each image separately is considerably more expensive than classical spectral filtering.

Questions for authors

See weakness part

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.