DPSF-Net: A Dual-Prior Spatial-Frequency Network for Real-World Remote Sensing Image Dehazing
Mei Lu, Shangliang Shao, Shanliang Yao
cs.CV · cs.AI
Quick verdict
Extends MCAF-Net with DCP, Fourier and selective-fusion modules, claiming SOTA on RRSHID; promising but internal number inconsistencies warrant scrutiny.
Summary
The paper proposes DPSF-Net, an encoder-decoder network for real-world remote sensing image dehazing built as a conservative extension of MCAF-Net. It feeds hazy RGB concatenated with dark-channel-prior (DCP) maps as a 6-channel input, replaces MCAF-Net's pointwise branch with a FourierUnit branch inside a spatial-frequency residual interaction block (SFRIB), extends its colour attention into a prior-guided feature attention module (PGFAM), and replaces its fusion module with selective kernel complementary fusion (SKCFM: bidirectional residual gating plus selective kernel fusion). Training combines L1, a structure-aware channel-weighted VGG19 perceptual loss, and a contrastive loss on two decoder scales. On the real-world RRSHID benchmark the model reports 24.63 dB PSNR, 0.7102 SSIM and 0.1510 LPIPS on average versus 24.11/0.6680/0.3921 for MCAF-Net, and stays competitive on RSID, RICE1, RICE2 and StateHaze1K-thick, with only 669.8K parameters and 18.27G FLOPs. Stepwise ablations attribute gains to each component. Main limitations: the ablation baseline value (24.66 dB) conflicts with the same model's Table 1 result (25.40 dB) on RRSHID-thick; PGFAM itself is never independently ablated; the dramatic LPIPS reductions (e.g., 0.3921 to 0.1510) far exceed the PSNR/SSIM gains and go undiscussed; there is no code release, no variance across runs, and no downstream-task evaluation.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Joint RGB-DCP input with separate embedding branches lets the physical dark-channel-prior cue guide feature learning without acting as a hard restoration constraint (Sec. III-A).
- Replacing MFIBA's pointwise branch with a FourierUnit branch (SFRIB) adds global low-frequency haze perception without increasing cost; the final model uses 18.27G FLOPs vs the baseline's 19.82G (Tables 3, 5).
- DPSF-Net achieves state-of-the-art on real-world RRSHID: 24.63 dB PSNR, 0.7102 SSIM and 0.1510 LPIPS on average, improving MCAF-Net by 0.52 dB and more than halving LPIPS (Table 1).
- Each proposed component contributes incrementally, per the stepwise ablation on RRSHID-thick raising PSNR from 24.66 to 25.63 dB (Table 3).
- The method remains competitive on four synthetic benchmarks with only 669.8K parameters (Tables 2, 5).
Strengths
Transparent component-wise attribution via stepwise and loss ablations
Table 3 isolates FourierUnit (+0.50 dB on RRSHID-thick), BRCM, SKFusion and 6-channel DCP sequentially; Table 4 separates SCWP (+0.84 dB over L1 alone) from the contrastive term (+0.41 dB), letting readers see which pieces matter.
Credible efficiency claim, including below-baseline FLOPs
Table 5: 669.8K parameters vs 558.1K for MCAF-Net, and 18.27G FLOPs below the baseline's 19.82G despite adding a Fourier branch, gating and selective fusion.
Broad evaluation with metric-disagreement honesty
Candid limitations section
Weaknesses
Internal number contradictions across tables
Disproportionate LPIPS improvements left undiscussed
PGFAM never independently ablated; DCP preparation details missing
No statistical rigor and no code release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Structured Review: DPSF-Net (arXiv 2609.06962)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DPSF-Net?
DPSF-Net is a lightweight encoder-decoder network (about 670K parameters, 18.27G FLOPs) for removing haze from remote sensing images. It extends MCAF-Net by feeding hazy RGB plus dark-channel-prior maps as a joint 6-channel input, adding a Fourier-domain branch (SFRIB) for global haze, prior-guided attention (PGFAM), and selective kernel fusion of skip features (SKCFM).
What results does DPSF-Net achieve?
On the real-world RRSHID benchmark it reports 24.63 dB PSNR, 0.7102 SSIM and 0.1510 LPIPS on average, versus 24.11/0.6680/0.3921 for MCAF-Net, and stays competitive on the RSID, RICE1, RICE2 and StateHaze1K-thick datasets. Gains over the closest baseline are modest on PSNR (0.23-0.52 dB) but large on LPIPS.
Is code or data available for DPSF-Net?
The reviewed text contains no code, model or dataset release statement. The underlying datasets (RRSHID, RSID, RICE1/2, StateHaze1K-thick) are described as public, but reproducing the numbers requires a full reimplementation from the architecture section.
What are the paper's main weaknesses?
The main reviewer concerns are internal number inconsistencies (MCAF-Net scores 25.40 dB in Tables 1/5 but 24.66 dB as the ablation baseline on the same subset), the absence of an independent ablation of the PGFAM module, dramatic LPIPS improvements that dwarf the PSNR/SSIM gains without explanation, and no variance reporting or statistical tests.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 6-9 are referenced only by caption in the extracted text; assessment of qualitative claims relies on captions, not image content.
- Some extracted table digits look anomalous (SCANet SSIM 0.0538 on RRSHID-moderate and 0.3741 on average in Table 1); these may be paper typos or PDF-extraction artifacts and could not be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.