{"id":"721d2a54-b90e-449d-afdf-f864ca854013","arxiv_id":"2504.19198","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SS-UIE, a spatial-spectral dual-domain network combining Mamba-style scanning and FFT filtering, achieves state-of-the-art underwater image enhancement with lower computational cost.","lead":"This paper presents a new neural network for enhancing underwater images that adapts its processing to different image regions and color bands. The reported results beat eight previous methods on three benchmarks while using fewer parameters and computations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim hinges on unverified baseline comparison; no error bars or evaluation scripts are provided, and the SWSA 'linear complexity' claim contradicts the paper's own Appendix C.","rationale":"The reader's weakest assumption identified the fairness and completeness of the baseline comparison as the key risk to the SOTA performance claim, and I concur. Without evaluation scripts, seeds, or error bars, the reported margins cannot be validated; this is the most load-bearing concern because it directly supports the central claim. The paper does provide a public code link and ablations, which are positive evidence, and the complexity error is a real but secondary issue that can be corrected without changing the measured efficiency or the performance ranking. Therefore, the appropriate disposition remains conditional acceptance pending the release of a reproducible evaluation protocol and a correction to the complexity statement. My concern aligns with the reader's, so no verdict change is warranted.","tokens_in":15701,"tokens_out":7988,"duration_ms":75665,"concrete_test":"Retrieve the released code (github.com/LintaoPeng/SS-UIE) and the official implementations of the eight baselines. Retrain all learning-based methods on the same UIEB and LSUI splits used in the paper, with the settings stated in §A.1 (256×256, Adam, 600 epochs, batch size 10, and the same data augmentation), running at least three random seeds per method. Report mean and standard deviation of PSNR/SSIM on the fixed test sets. If the 95% confidence intervals of the baselines overlap with SS-UIE, or if the ranking changes across seeds, the claimed SOTA margin is not established. Separately, compute the asymptotic complexity of the SWSA branch to confirm that it is O(N log N) rather than O(N), as the paper's own Appendix C states.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SS-UIE outperforms eight SOTA methods while being cheaper depends on the fairness and reproducibility of the comparisons in Table 1. The paper states that provided source codes and identical settings were used, but it does not report per-baseline training configurations, random seeds, or any statistical variance. The reported margins are small in some cases (e.g., +0.33 dB PSNR over WF-Diff on UIEB; +1.22 dB over DM-Water on LSUI) and are given as single runs. Without confidence intervals or a specified protocol, a third party cannot verify that the gains are due to the method rather than to under-trained baselines or accidental test-set differences. Additionally, the paper repeatedly claims 'linear complexity' for both MCSS and SWSA (Abstract, §1, §2), but Appendix C explicitly derives that FFT/IFFT is O(N log N). This internal inconsistency weakens the efficiency claim, though it does not by itself overturn the measured FLOPs in Fig. 1. The load-bearing issue is that the performance claim is not independently checkable from the manuscript as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SS-UIE, an underwater image enhancement network built from Spatial-Spectral blocks (SS-blocks). Each SS-block combines a Multi-scale Cycle Selective Scan (MCSS) module for spatial modeling and a Spectral-Wise Self-Attention (SWSA) module based on FFT for spectral modeling, connected in parallel. A Frequency-Wise Loss (FWL) is added to emphasize high-frequency reconstruction. The authors report state-of-the-art results on UIEB, LSUI, and U45 datasets while claiming lower parameter and FLOP counts than eight compared methods. The method section includes architectural details, the ablation study supports each component, and code is made available.","tokens_in":15938,"tokens_out":4492,"duration_ms":43926,"significance":"If the reported results hold, the paper contributes a practically useful dual-domain architecture for underwater image enhancement, with a frequency-domain loss that improves high-frequency detail recovery. The efficiency comparison is a valuable part of the contribution: the proposed network reportedly achieves competitive quality with lower parameter and FLOP counts. However, the paper's central efficiency claim is weakened by an internal inconsistency about the complexity of the FFT-based SWSA module, and the state-of-the-art claim rests on single-run comparisons without statistical verification. These issues are fixable and do not by themselves invalidate the proposed method.","major_comments":[{"comment":"The paper repeatedly claims that SWSA has 'linear complexity' (Abstract, §1, §2, Method), but SWSA uses 2D FFT and IFFT as defined in Eqs. 5–7, and Appendix C explicitly derives that FFT reduces DFT complexity to O(N log N). The SWSA branch therefore has O(N log N) complexity, not O(N). This is a load-bearing inconsistency because 'linear complexity' is presented as a central advantage of the design. The claim should be corrected throughout to 'quasi-linear' or 'O(N log N)' for SWSA, with MCSS's O(N) complexity stated separately.","section":"Abstract, §1, §2, Method (Eqs. 5–7), Appendix C"},{"comment":"The state-of-the-art claim is supported only by single-run comparisons against eight baselines. The paper states that provided source codes and identical settings were used, but does not report random seeds, number of runs, confidence intervals, or statistical significance. Some differences are small (e.g., +0.33 dB PSNR over WF-Diff on UIEB), so without variance estimates a third party cannot determine whether the reported gains are due to the method or to accidental training/selection differences. Please provide mean±std over multiple runs or release the full evaluation scripts and per-baseline configurations for verification.","section":"Experiments, Table 1, §Experiment Setup"},{"comment":"The ablation results are presented as single numbers for each configuration, and the checkmark-based table does not unambiguously identify which row corresponds to 'BL+SS2D', 'Serial', and 'Parallel' as described in the text. The conclusion that the parallel design outperforms the serial design rests on this table, so the configuration labels must be made explicit and, ideally, supported by multiple runs to rule out optimization noise.","section":"Ablation Study, Table 2"},{"comment":"The definition of the dynamic weight θ is vague: the text states it is 'linearly related to the distance d(u,v)' and then says it 'changes linearly with the absolute value of the kth channel frequency distance sqrt(|d|)'. The functional form of θ is not written explicitly, and it is unclear whether θ is normalized or how it affects the gradient. Please give the exact formula for θ and explain how it adapts across channels and frequencies.","section":"Method, Frequency-Wise Loss (Eqs. 12–13)"}],"minor_comments":[{"comment":"The text refers to 'Tab. 3' and 'Tab. 4' when discussing the quantitative comparison and the ablation study, but those tables are actually labeled Table 1 and Table 2 in the manuscript. Please renumber or fix the cross-references.","section":"References to tables in full text"},{"comment":"The input size in Table 4 says 'Conv(16,1,3,1,1)' produces 'Output Size (256,256,3)', but the final convolution output has 3 channels; please confirm whether the reconstruction module outputs 3 channels directly and whether the preceding upsample and conv blocks are correctly described.","section":"Appendix A and Table 3/4"},{"comment":"The proof is a standard derivation of DFT/FFT, and the text contains 'our SF-block' which appears to be a typo for 'SS-block'. Consider condensing this proof or moving it to supplementary material.","section":"Appendix C"},{"comment":"There are several typos: 'Enhancemnet' in the Related Work heading, 'dual domian loss' in Eq. 15, 'SS-Blcok' in the Ablation Study, and inconsistent punctuation (e.g., 'Mamba' vs. 'Mamba,'). These should be corrected.","section":"Typos throughout"},{"comment":"The code link is provided, but no commit hash or version is given, and the evaluation scripts for the baseline comparisons are not included. To support reproducibility, please provide a specific release or commit and document the exact commands used to produce Table 1.","section":"Code release and reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical contribution to underwater image enhancement, and the proposed architecture is sound in principle. The main concerns are the complexity overstatement and the lack of statistical rigor in the comparisons; both are addressable through revision. The authors' use of their own previous work (U-Shape) as a baseline is not circular because the evaluation is on public datasets and against independent methods. I would encourage the editor to request a revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two quick things to know. First, this is a genuinely useful engineering paper: the parallel MCSS (a multi-scale cycle selective scan) and SWSA (an FFT-based spectral filter) with the frequency-wise loss is a sensible new combination for underwater image enhancement, and the ablations make a coherent case for each piece. Second, the paper's own complexity claim is wrong, and the experimental reporting is thinner than I would like — but neither problem sinks the core result.\n\nWhat is actually new: the specific architecture SS-UIE, stacking SS-blocks with dense connections, plus a frequency-weighted loss. The components are borrowed from VMamba and global-filter networks, but the parallel spatial-spectral design and its application to degraded underwater images is new. The paper gives a public code link, uses standard public datasets (UIEB, LSUI, U45), and reports PSNR/SSIM plus FID/LPIPS. The ablation study breaks down each module and even compares parallel vs serial arrangements, which is more than many papers in this space do. The consistent SOTA numbers across three datasets are a real result, even if the margins are modest.\n\nThe soft spots, in order of seriousness. First, the repeated 'linear complexity' claim for SWSA is simply wrong: the SWSA branch uses FFT/IFFT, and Appendix C correctly derives O(N log N). That is an internal contradiction the authors should fix by saying 'near-linear' or 'log-linear' or by reporting the actual FLOPs honestly. Second, there are no error bars, no multiple seeds, and no description of per-baseline hyperparameters, so a third party cannot verify that the +0.33 dB over WF-Diff on UIEB reflects the method rather than training luck. The stress-test note has this right. Third, the frequency-wise loss formula is under-specified: the dynamic weight θ is described verbally but not given as an equation. That is a minor but real reproducibility gap.\n\nNone of this invalidates the paper. The architecture works, the ablations support the design, and the code is public. It would benefit from a revision that fixes the complexity language, adds statistical rigor if possible, and pins down the loss formula. I would send it to peer review — a competent referee will ask for those changes, but this is not a desk-reject.\n\nFor you: if you work on low-level vision or efficient attention mechanisms, this is worth reading as a recent data point. I'd bring it to the reading group mainly to talk about how often the 'linear' claim gets waved through.","headline":"Solid new architecture for underwater enhancement with a real SOTA result, but the 'linear complexity' claim contradicts its own appendix and the experimental reporting lacks statistical rigor.","tokens_in":16429,"tokens_out":3562,"would_cite":true,"duration_ms":32945,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SS-UIE claims state-of-the-art underwater image enhancement by adapting to degradation in both space and frequency, at lower computational cost.","keywords":["underwater image enhancement","spatial-spectral adaptive learning","selective state space models","multi-scale cycle selective scan","spectral-wise self-attention","frequency-wise loss","FFT global filter","UIEB LSUI U45 benchmarks"],"falsifier":"A reader could settle the claim by rerunning all eight comparison methods from official code on the same UIEB and LSUI splits with several seeds and checking whether SS-UIE's PSNR, SSIM, LPIPS, and FID margins persist. Separately, timing SS-UIE at 256, 512, and 1024 pixel inputs would show whether the FFT branch's $O(N \\log N)$ growth undercuts the linear-complexity claim.","tokens_in":15527,"feed_emoji":"🌊","tokens_out":7309,"duration_ms":68364,"temperature":0.7,"pith_summary":"Underwater images degrade unevenly: some regions are hazier than others, and some color channels are absorbed more strongly. The paper tries to show that an enhancer built to adapt to both kinds of unevenness—spatial and spectral—outperforms methods that apply uniform processing. Its SS-UIE network runs a selective state-space scan over image regions in parallel with an FFT-based learnable filter over frequency bands, and adds a loss that emphasizes frequencies the model gets wrong. The paper reports higher PSNR, SSIM, LPIPS, and FID scores than eight prior methods on the UIEB, LSUI, and U45 benchmarks, while using fewer parameters and FLOPs. If this holds, cheaper and sharper underwater enhancement could become practical for real-time use.","feed_headline":"Underwater enhancer wins on three benchmarks with fewer FLOPs","feed_subtitle":"SS-UIE pairs a selective spatial scan with an FFT spectral filter to adapt to uneven water degradation.","key_machinery":"The load-bearing object is the SS-block: a parallel combination of a Multi-scale Cycle Selective Scan (MCSS) module and a Spectral-Wise Self-Attention (SWSA) module, followed by a $1\\times1$ convolution residual. MCSS unfolds input features into patch sequences along several cyclic traversal paths and passes each sequence through an S6 state-space block, giving a global spatial receptive field. SWSA applies a 2D FFT, multiplies the spectrum by a learnable filter $K$ of the same shape, and transforms back with IFFT, giving global inter-spectral mixing. The paper argues that running these two branches in parallel lets the network capture severe-attenuation regions and bands simultaneously. The Frequency-Wise Loss (FWL) computes a per-channel spectral distance $d_k(u,v)$, reweights it by a factor $\\theta_k(u,v)$ that grows with reconstruction error at that frequency, and adds it to an $L_1$ term, which is what directs extra capacity to hard high-frequency content.","core_discovery":"The paper's central claim is that underwater image enhancement improves when the network allocates its capacity according to degradation level, rather than treating all pixels and color bands equally. SS-UIE pairs a Multi-scale Cycle Selective Scan (MCSS), which gives a global spatial receptive field at claimed linear cost, with a Spectral-Wise Self-Attention (SWSA) module that multiplies the FFT of features by a learnable global filter, running both in parallel inside an SS-block. The block is designed to model, for each spatial region and each spectral band, how strongly the water has attenuated the signal, and to strengthen the network's response where attenuation is severe. A Frequency-Wise Loss (FWL) then reweights residual errors by how far the predicted spectrum is from the ground truth, pushing the model to recover high-frequency detail. The paper reports that this architecture surpasses eight prior enhancement methods on UIEB, LSUI, and U45 while requiring cheaper computational and memory costs.","pith_inferences":["A strong test of the paper's mechanism would be to visualize the learned filter $K$ and the MCSS scan weighting per region: if the network truly adapts to degradation levels, these should concentrate on the haziest patches and the most attenuated color channels. The paper does not show such maps.","The same dual-domain design may generalize to other inverse problems with wavelength-dependent attenuation, such as dehazing or imaging through scattering media, but the paper only evaluates underwater benchmarks.","Because FWL's weight $\\theta$ is tied to absolute frequency error, it could function as a generic finishing loss for any image restoration model; a cheap experiment is to swap FWL into a Swin-based baseline and measure PSNR on LSUI.","The reported 'linear complexity' of the spectral branch is an idealization: a strict scaling analysis would count FFT cost as $O(N \\log N)$, so the efficiency advantage should be verified by wall-clock tests at larger resolutions."],"forward_implications":["If the reported margins hold, underwater image enhancement can be performed at lower computational and memory cost than current transformer and diffusion baselines, making on-device or real-time deployment plausible.","The parallel MCSS/SWSA block gives a reusable template for restoration tasks with non-uniform degradation, such as haze, low light, or turbid-medium imaging, where region-specific and frequency-specific adaptation should help.","FWL can be added to any existing underwater enhancement network as a plug-in training loss with no added inference cost; it should improve high-frequency fidelity whenever the model underfits texture.","A degradation-level-adaptive design may reduce the need for explicit physical parameter estimation, such as depth or transmission maps, in underwater enhancement pipelines."],"supporting_citations":[{"why":"Supplies the S6 selective state-space block that MCSS adapts to vision with multi-scale cyclic scans.","marker":"(Gu and Dao 2023)"},{"why":"Provides the original 2D selective scan (SS2D) used as the ablation baseline that MCSS is designed to improve.","marker":"(Liu et al. 2024)"},{"why":"Provides the FFT machinery behind the SWSA spectral filter.","marker":"(Nussbaumer and Nussbaumer 1982)"},{"why":"F-Principle result that neural nets fit low frequencies first, motivating the frequency-wise loss.","marker":"(Xu, Zhang, and Xiao 2019)"},{"why":"Defines the UIEB benchmark and is one of the two full-reference training/test datasets.","marker":"(Li et al. 2020)"},{"why":"Defines the LSUI dataset, the other full-reference benchmark, and the U-Shape baseline.","marker":"(Peng, Zhu, and Bian 2023)"},{"why":"Provides the U45 non-reference generalization test set.","marker":"(Li, Li, and Wang 2019)"},{"why":"Swin transformer block is the baseline replaced in the ablation study.","marker":"(Liu et al. 2021)"},{"why":"WF-Diff is the strongest prior baseline that SS-UIE must outperform for the claim to stand.","marker":"(Zhao et al. 2024)"}],"fun_headline_variants":["Adaptive dual-domain learning sharpens underwater images with less compute","Spatial-spectral attention adapts to uneven underwater degradation","Frequency-aware loss and dual-domain scanning lift underwater detail","SS-UIE: degradation-aware spatial-spectral block for clear underwater","Low-cost dual-domain adaptation sharpens murky underwater scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline result rests on the eight comparison methods being run fairly with their official code and identical settings on the same splits; if any baseline was under-tuned, the reported margins could shrink. Relatedly, counting the FFT-based branch as linear is not exact, since a 2D FFT costs $O(N \\log N)$, so the efficiency claim depends on treating that growth as negligible.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive dual-domain learning sharpens underwater images with less compute","Spatial-spectral attention adapts to uneven underwater degradation","Frequency-aware loss and dual-domain scanning lift underwater detail","SS-UIE: degradation-aware spatial-spectral block for clear underwater","Low-cost dual-domain adaptation sharpens murky underwater scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002026,"raw_usage":{"total_tokens":7913,"prompt_tokens":981,"completion_tokens":6932,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":6847}},"tokens_in":597,"tokens_out":6932,"duration_ms":43819,"temperature":1.0,"reasoning_tokens":6847,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:58:43.195278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the claim by rerunning all eight comparison methods from official code on the same UIEB and LSUI splits with several seeds and checking whether SS-UIE's PSNR, SSIM, LPIPS, and FID margins persist. Separately, timing SS-UIE at 256, 512, and 1024 pixel inputs would show whether the FFT branch's $O(N \\log N)$ growth undercuts the linear-complexity claim.","supporting_citations":[{"cited_title":"J.; and Nussbaumer, H","cited_arxiv_id":null,"evidence_quote":"Provides the FFT machinery behind the SWSA spectral filter."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"F-Principle result that neural nets fit low frequencies first, motivating the frequency-wise loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the LSUI dataset, the other full-reference benchmark, and the U-Shape baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"WF-Diff is the strongest prior baseline that SS-UIE must outperform for the claim to stand."}],"review_version":1}