- 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.