acceptodds
Under review as a conference paper at ICLR 2027

Angular Spectrum Propagation Modeling for Lightweight Image Dehazing

Abstract

Most image dehazing methods operate in the pixel-intensity domain, leaving frequency-domain spectral distortions implicitly handled. We introduce Angular Spectrum Propagation Modeling (ASPM), a lightweight frequency-domain front-end inspired by the angular spectrum method in physical optics. ASPM treats a hazy image as a complex-valued virtual wave field and performs learned, per-channel spectral filtering via a propagation-distance representation and phase maps. Designed as an architecture-agnostic module within dehazing, ASPM provides a physics-inspired structured prior—not a claim of literal coherent-wave physics—that supports restoration even at minimal parameter scales. Under a matched-capacity comparison, replacing ASPM's constrained angular-spectrum operator with an unconstrained learnable complex filter of equal size still loses 1.30 dB, indicating the benefit comes from this specific structural constraint rather than merely from having a complex-valued frequency-domain component. Our resulting model, ASPMDehazeNet-S, is compact (0.04M parameters) and achieves 26.26 dB PSNR on SOTS-Indoor at 4.2 ms (over 230 FPS) on an NVIDIA RTX 4090. We evaluate on synthetic benchmarks (SOTS-Indoor, SOTS-Outdoor, RS-Haze) and two real-world benchmarks (NH-Haze, Dense-Haze): a scaled variant (ASPMDehazeNet-L, 0.1M) reaches 32.97 dB on SOTS-Indoor and is the best among the compared sub-1M models on SOTS-Indoor, SOTS-Outdoor, and both real-world benchmarks, while remaining competitive on high-resolution RS-Haze. Overall, ASPM offers a favorable accuracy–efficiency trade-off for lightweight dehazing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Dehazing, Frequency-Domain Learning, Angular Spectrum Propagation, Lightweight Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LQ1eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ASPMDehazeNet, a lightweight image dehazing framework built around Angular Spectrum Propagation Modeling (ASPM). The method represents the input as a virtual complex-valued field and applies an input-conditioned frequency-domain phase operator inspired by angular-spectrum propagation. A separate FreqEnhance module estimates a haze-dependent radial amplitude gain, and the resulting spectral representation is fused with the input for image reconstruction. The authors evaluate two model configurations on synthetic and paired real-haze benchmarks, together with cross-backbone experiments and several component-level ablations.

Critical strengths

The paper presents a compact frequency-domain dehazing architecture and provides a relatively complete description of the main components, including the ASPM operator, FreqEnhance module, training objective, and cross-backbone integration. The distinction between the proposed phase correction and amplitude enhancement is reasonably clear.

The experimental section covers several synthetic and real-haze benchmarks and includes component ablations, cross-backbone experiments, and runtime measurements. The authors also explicitly acknowledge several limitations of the current comparisons, which makes the scope of some claims clearer.

Critical weaknesses

The main novelty and physical interpretation of ASPM are not sufficiently convincing. Although the method is motivated by angular-spectrum propagation, the predicted “propagation distance” d_c(I;f_x,f_y) is itself frequency-dependent and is effectively used to parameterize an input-conditioned phase mask. It is therefore unclear what meaningful constraint the angular-spectrum formulation provides beyond a generic learned frequency-domain phase operator. The fixed RGB constants [0.7890, 0.5907, 0.4241] are also treated as structural scales rather than physical wavelengths, yet their specific values are not justified or analyzed. Similarly, the particular radial gain function used in FreqEnhance appears largely empirical, with limited evidence that this form is preferable to other simple frequency-domain alternatives.

The experimental evidence does not fully support the claimed advantage of the proposed design. The key LCF comparison is reported only under a reduced 100-epoch setting, whereas the main models are trained for 300 epochs, leaving it unclear whether the observed gap persists under the full training schedule. NH-Haze and Dense-Haze are also used to motivate or validate the gain design and are then reused for real-haze evaluation, so these results are not fully independent. The quantitative comparison is further incomplete: DEA-Net (IEEE TIP 2024) and ConvIR (IEEE TPAMI 2024) are already discussed in the related-work section but omitted from the main results, while recent relevant methods such as CoA (CVPR 2025) and the closely related frequency-domain dehazing method FrDiff (ICCV 2025) are not discussed. Finally, the reproduced SLRNet result is internally inconsistent: Table 1 reports 28.44 dB on SOTS-Outdoor, whereas Sec. 4.2 reports 28.58 dB for the same re-run.

Questions for authors

What constraint does the proposed angular-spectrum parameterization actually impose over a generic input-conditioned phase mask? Since d_c(I;f_x,f_y) is itself frequency-dependent, could an arbitrary phase response be represented simply by choosing an appropriate d? A comparison against a capacity-matched dynamic phase-mask baseline would help clarify this.

How were the fixed RGB constants [0.7890, 0.5907, 0.4241] selected? Are the results sensitive to these values? Since they are not interpreted as physical wavelengths, it would be useful to show whether they provide any advantage over learned or randomly initialized channel scales.

Could the authors provide the exact architecture and parameter count of the LCF baseline? In particular, is the LCF also input-conditioned, and does it use the same predictor capacity as ASPM? Please also report the comparison after the full 300-epoch training schedule and, ideally, across multiple random seeds.

NH-Haze and Dense-Haze are used to motivate or validate the FreqEnhance gain function and are later used again for evaluation. How was information from these datasets separated from test-time evaluation? A leave-one-dataset-out experiment or evaluation on an unseen real-haze benchmark would provide stronger evidence for generalization.

Table 1 reports 28.44 dB for the reproduced SLRNet result on SOTS-Outdoor, while Sec. 4.2 states 28.58 dB. Which value is correct? Could the authors provide the corresponding training configuration and checkpoint details for the reproduced baselines?

Why are the reproduced outdoor baselines trained on the same ITS split rather than under their standard outdoor training protocol? It would be helpful to report both the controlled same-training-data comparison and the standard-protocol results, especially when discussing competitiveness with existing lightweight methods.

ASPM-L has only approximately 0.1M parameters, but requires 4.937G MACs and about 39 ms on an RTX 4090. In what sense should the model be considered lightweight beyond parameter count?

The final model uses a weighted combination of several reconstruction and image-statistic losses. Could the authors provide an ablation of the training objective? Without this, it is difficult to separate the contribution of ASPM from gains introduced by the additional loss terms.

Reviewer_ktp3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes ASPM, a lightweight frequency-domain front end for single-image dehazing. It models the hazy image as a complex wave field and introduces a physical prior to construct a lightweight dehazing model with high inference efficiency.

Critical strengths

The authors provide detailed mathematical formulations of the proposed operator, and Figure 1 directly characterizes the underlying spectral phenomenon.

Critical weaknesses

The baselines are outdated, with AOD-Net being one example. The authors should compare ASPM against more recent methods published within the last three years.

The layout of Table 1 is confusing. The authors' proposed methods should be placed at the bottom of the table. The rationale for the current ordering is unclear, and many entries are left blank.

Secondary details are given the same prominence as the main arguments, significantly disrupting the flow of the paper. Examples include the specification of the RGB constants as [0.7890, 0.5907, 0.4241] in line 161.

Replacing the CNN-based parameter estimator with ViT-ASPM appears to be an engineering choice rather than a substantive contribution. This component should be discussed, if necessary, in the supplementary material rather than emphasized in the main paper.

DehazeNet's indoor SSIM is reported as 0.872 in Table 1 but 0.867 in Table 4. This inconsistency is confusing and requires clarification.

Questions for authors

The manuscript appears incomplete, which makes it difficult to follow the work and provide a thorough review. Although the authors may have invested considerable effort in the implementation, the manuscript lacks clear prioritization of the key contributions and contains excessive, poorly organized details. This gives the impression that the paper was prepared hastily, potentially with indiscriminate reliance on AI-assisted writing. The authors should substantially revise the manuscript, improve its structure and clarity, and include the missing experiments before the work can be properly evaluated.

Reviewer_Zzu5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Angular Spectrum Propagation Modeling (ASPM), a lightweight frequency-domain module for image dehazing inspired by Fourier optics. ASPM combines an input-conditioned spectral phase operator with a learnable amplitude enhancement function to improve restoration quality under limited parameter budgets.

Critical strengths
  1. The proposed models achieve competitive restoration quality with extremely small parameter counts.
  2. The paper explores a structured, physics-inspired alternative to unconstrained frequency-domain learning.
Critical weaknesses
  1. The proposed operator is inspired by coherent wave propagation rather than atmospheric scattering physics. Although the authors acknowledge this distinction, it remains unclear why this particular analytical formulation should outperform other input-conditioned spectral filters.
  2. Despite having only approximately 0.1M parameters, ASPMDehazeNet-L requires 4.94G MACs and 39.2 ms per image on an RTX 4090. The paper lacks comprehensive runtime breakdowns, memory measurements, and resolution-scaling experiments to support its practical efficiency claims.
  3. NH-Haze and Dense-Haze are used both to motivate the spectral enhancement function and to evaluate the model. Therefore, these results do not provide fully independent validation of generalization to unseen haze distributions.
Questions for authors

See the Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.