{"id":"800d03df-d820-4e00-bb5d-0c5615efa2a0","arxiv_id":"2412.18459","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UIR-PolyKernel achieves state-of-the-art underwater image restoration using a lightweight CNN with polymorphic large kernels and hybrid frequency/spatial attention.","lead":"This paper builds a lightweight pure CNN, UIR-PolyKernel, that restores underwater images by mixing very large square and strip-shaped convolutions with spatial and frequency domain attention. It reports top scores on standard underwater benchmarks while using far fewer parameters than competing Transformer models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark SOTA claim is not yet established: paired-test metrics use algorithm-generated references, and UCIQE is both in the training loss and reported as a success metric; independent validation is required.","rationale":"I read the paper as a competent architecture study: the CSC, LKA, and HDA modules are clearly described, the ablations isolate each component, and the parameter/MAC comparison is useful. The internal logic of the method is consistent, and I would not challenge the claim that a carefully designed CNN can be competitive with Transformers on the reported benchmarks. However, the load-bearing step for the stronger SOTA claim is the validity of the evaluation. The authors themselves state in Section II-D that the reference targets are not true ground truth, yet all paired-test metrics are computed against those targets. That is a known limitation of real-world underwater datasets, but it means the quantitative results establish fidelity to algorithm-generated references, not restoration of true scene radiance. In addition, UCIQE appears in the training loss and is then reported as a success metric; this is a circularity that inflates one of the four reported metrics. The reader's verdict of CONDITIONAL is appropriate, and my concern reinforces it rather than moving it to a different category. I partially agree with the reader because they identified the reference-target issue as the weakest assumption, whereas I see the UCIQE in-loss/report circularity as the more directly testable flaw; both point to the same need for independent evaluation before accepting the abstract's SOTA claim.","tokens_in":9898,"tokens_out":5325,"duration_ms":53250,"concrete_test":"Retrain UIR-PolyKernel with the UCIQE term removed (w3=0) under the same protocol, and re-evaluate PSNR/SSIM/LPIPS/UCIQE on the three paired test sets; then also evaluate Ours, Semi-UIR, and URSCT on a synthetic underwater test set with true clean ground truth (e.g., physically rendered scattering images with known clean scenes). If the model no longer ranks first on UCIQE after removing the UCIQE loss, or if it does not rank first on PSNR/SSIM/LPIPS against true clean scenes, the SOTA claim in the abstract should be downgraded to \"competitive on existing reference-based benchmarks.\"","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that UIR-PolyKernel is state-of-the-art quantitatively and qualitatively. The quantitative evidence in Table I, however, is generated under a protocol that the paper itself qualifies in Section II-D: the reference targets \"do not represent the true ground truth\" and are instead high-quality outputs of existing methods. PSNR/SSIM/LPIPS therefore measure proximity to a particular set of algorithm-produced references, not correctness of restoration; a method that memorizes the style of those references can score well without restoring true scene radiance. This concern is sharpened by the loss in Eq. (3): w3=0.01 optimizes 1−UCIQE, and UCIQE is then reported as one of the four success metrics in Table I, so the claimed UCIQE advantage (e.g., 0.458 on UIEB vs 0.432 for URSCT) is at least partly an artifact of training toward the evaluation metric. No error bars are reported, and on UIEB PSNR the method is second (23.52 vs 23.64 for Semi-UIR), so the \"large margin\" statement is not robust. The architecture itself is coherent and the ablations are sensible; the problem is that the empirical support for the headline claim is tied to an evaluation protocol that can reward matching algorithm-generated targets rather than true underwater scene restoration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UIR-PolyKernel, a lightweight pure-CNN architecture for underwater image restoration. The method combines polymorphic large-kernel convolutions (Composite Shape Convolution module with 31x1, 1x31, 31x31, and 1x1 depth-wise convolutions at the bottleneck), Large Kernel Attention modules at intermediate scales, and a Hybrid Domain Attention module that fuses frequency-domain pixel attention with spatial-domain channel attention. The model has about 1.84M parameters and 13.67 GMACs. Experiments on UIEB, EUVP, LSUI, and RUIE report PSNR, SSIM, LPIPS, and UCIQE, with all deep-learning baselines retrained on the same training data. Ablations show each component contributes. The central claim is state-of-the-art quantitative and qualitative performance relative to existing methods, including Transformers.","tokens_in":10199,"tokens_out":5969,"duration_ms":48691,"significance":"If the claims are reliable, the paper makes a useful practical point: a carefully designed lightweight CNN can compete with heavier Transformer-based underwater restoration models. The architecture is coherent, the ablation study is systematic, all baselines are retrained under a common protocol, and source code is provided. However, the evaluation protocol has serious limitations that directly affect the headline claim. The paper itself states that the reference targets are not true ground truth, and UCIQE appears both as a training loss term and as a reported success metric. These issues, combined with the lack of error bars and a specific counterexample in the UIEB PSNR column, undermine the 'state-of-the-art' assertion. The underlying architectural ideas are plausible and worth further validation, but the current evidence is not sufficient to support the central claim.","major_comments":[{"comment":"The loss function in Eq. (3) includes the term w3 * Lu, where Lu = 1 - UCIQE and w3 = 0.01. This makes UCIQE a training objective. Table I then reports UCIQE as one of the four success metrics. Because the method is trained to increase UCIQE, the reported UCIQE advantage over competing methods (e.g., 0.458 on UIEB vs. 0.432 for URSCT) is at least partly an artifact of direct optimization rather than an independent quality measure. The paper itself cautions in Section III-C that UCIQE is 'not absolutely reliable,' yet still presents it as a headline result. The authors should either remove UCIQE from the training loss or not report UCIQE as a comparative benchmark metric.","section":"Section II-D (Eq. 3) and Table I"},{"comment":"Section II-D explicitly states that 'the reference target, while consisting of high-quality results from state-of-the-art methods, does not represent the true ground truth.' Under this protocol, PSNR, SSIM, and LPIPS in Table I measure similarity to algorithm-generated reference outputs, not to true clean scene radiance. A model could achieve high scores by memorizing the statistics of these references without genuinely restoring underwater scenes. This is load-bearing for the 'state-of-the-art' claim, especially because all deep-learning baselines are retrained on the same references. The authors should provide independent validation on data with true ground truth (e.g., synthetic underwater scenes with known clean images) or include a human perceptual study to support the qualitative superiority asserted in Section III-C.","section":"Section II-D and Table I"},{"comment":"On the UIEB test set, the method achieves PSNR 23.52 dB, which is second to Semi-UIR's 23.64 dB. Section III-C states that 'In all other results, our method leads other approaches by a large margin,' but this is not true for a primary metric on the primary dataset. Moreover, no error bars, confidence intervals, or repeated runs are reported for any metric, so the margins on other datasets (e.g., EUVP PSNR 26.42 vs. 25.74 for URSCT) are not established as statistically significant. The authors should qualify the state-of-the-art claim and provide variance estimates or significance tests.","section":"Table I, UIEB PSNR"},{"comment":"The ablation results show incremental improvements from adding each module (e.g., SSIM changes from 0.888 to 0.891 to 0.893 when HDA components are added). Without repeated runs or statistical testing, it is unclear whether these differences exceed training noise. The conclusion that each component contributes to 'superior performance' is therefore not robust. The authors should report means and standard deviations over multiple random seeds, or at least perform significance tests on the differences.","section":"Table II, Ablation Studies"}],"minor_comments":[{"comment":"The text says the FDPA block has 'two 1x1 convolution branches,' but Eq. (2) uses the same notation Conv1 twice. Please clarify whether the two branches have separate learnable weights or share them.","section":"Section II-C, Eq. (2)"},{"comment":"The text states 'we limit our UCIQE comparison to the deep learning methods,' yet Table I lists UCIQE values for all methods, including traditional ones. Please reconcile the statement with the table.","section":"Section III-C"},{"comment":"The phrase 'underwater and on-air images' appears to contain a typo; likely 'in-air' or 'atmospheric' images are intended.","section":"Abstract"},{"comment":"The UCIQE metric is mentioned without a citation. Please add a reference to the original UCIQE paper.","section":"Section III-A"},{"comment":"There are several spelling errors in affiliations ('Univeristy' appears twice). Please proofread the author information.","section":"Author affiliations"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a coherent and lightweight architecture, and the ablation study is a strength. The primary concern is the evaluation protocol: the use of non-ground-truth references and the inclusion of UCIQE in the loss weaken the SOTA claim. I recommend major revision, with the expectation that the authors add independent validation (e.g., synthetic ground-truth benchmarks or a user study), stop reporting UCIQE as a success metric, and provide error bars. The paper may be suitable for the journal after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look if you work in underwater image restoration. The architecture is a genuine combination, not a paradigm shift: polymorphic depth-wise convolutions (31x1, 1x31, 31x31) at the bottleneck, LKA modules at medium scales, and a hybrid frequency/spatial attention block at full resolution. The ablations are clean and consistent, each module adds a measurable gain, and the baselines were retrained under the same data, which is more than many papers in this area do. At 1.84M params and 13.67 GMACs it is genuinely light, and it beats heavier Transformer-based models on most metrics on the paired test sets. That is a real result for practitioners who care about deployment cost. The soft spots are real but not fatal. The paper itself says in Section II-D that the reference targets are not true ground truth, only outputs of existing methods. So the PSNR/SSIM/LPIPS numbers measure how well the model reproduces a particular set of algorithm-generated references, not actual restoration quality. That alone should temper the state-of-the-art claim. Second, UCIQE appears both in the loss (Eq. 3, w3=0.01) and as a headline success metric; the reported UCIQE advantage is partly a fitted quantity. The weight is small, so this is a minor-to-moderate issue, but it should be acknowledged or the metric dropped from the loss. Third, there are no error bars or repeated runs. On UIEB PSNR the method is second (23.52 vs 23.64 for Semi-UIR), so the large margin language is overstated. The paper is honest about the reference limitation, which counts in its favor. The math is straightforward and the code is public. The main weakness is that the empirical evidence underdetermines the grand claim. It is a good engineering contribution, not a breakthrough. Who should read it: anyone building lightweight underwater enhancement models, or benchmarking CNNs against Transformers in low-level vision. It deserves a serious referee, desk rejection would be wrong. I would recommend conditional acceptance, with the authors asked to either remove UCIQE from the loss, report results on a held-out set with true references where possible, or at minimum re-frame the SOTA claim to best among lightweight CNNs and provide error bars.","headline":"A solid, lightweight CNN architecture for underwater restoration with a real design contribution, but the SOTA claim is weakened by the evaluation protocol: algorithm-generated references and UCIQE in the training loss.","tokens_in":744,"tokens_out":752,"would_cite":false,"duration_ms":21567,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A carefully designed pure CNN can match transformer-based models on underwater image restoration, the paper claims.","keywords":["underwater image restoration","large kernel CNN","composite shape convolution","hybrid domain attention","frequency domain attention","lightweight architecture","image enhancement"],"falsifier":"Take a set of underwater scenes photographed with known color-calibration targets at measured depths, use the measured target values as true references, and recompute PSNR, SSIM, and LPIPS for UIR-PolyKernel and the transformer baselines; if the margin over the baselines disappears or reverses, the reported superiority is an artifact of algorithm-generated references.","tokens_in":9734,"feed_emoji":"🌊","tokens_out":5405,"duration_ms":42959,"temperature":0.7,"pith_summary":"The paper claims that a carefully built pure CNN, UIR-PolyKernel, matches or beats heavier transformer-based models on underwater image restoration benchmarks while using a fraction of the compute. The two load-bearing design choices are polymorphic large-kernel convolutions, which capture long-range and direction-dependent degradation cheaply, and hybrid frequency-space attention, which preserves fine detail. On paired test sets from UIEB, EUVP, and LSUI and unpaired RUIE data, the model reports leading SSIM, LPIPS, and UCIQE scores, with a few exceptions such as second place to Semi-UIR on UIEB PSNR. A reader should care because it suggests that pure CNN architectures are not a compromise for this task, and that restoration quality can be obtained at 1.84M parameters and 13.67G MACs.","feed_headline":"A 1.84M-parameter CNN tops underwater restoration benchmarks","feed_subtitle":"Author reports the pure CNN with polymorphic large kernels rivals heavier transformers at lower cost.","key_machinery":"The central mechanism is the Composite Shape Convolution (CSC) module: a parallel composition of depthwise strip, square, and pointwise convolutions with kernels 31x1, 1x31, 31x31, and 1x1, followed by a 1x1 pointwise convolution for inter-channel mixing. It is placed at the 64x64 bottleneck of a U-Net, where a single 31x31 depthwise convolution over the whole feature map would be costly; the polymorphic shapes reduce that cost while preserving both isotropic and anisotropic long-range context. Around it, Large Kernel Attention (LKA) modules simulate roughly 19x19 receptive fields for medium-range relations, and the Hybrid Domain Attention (HDA) module uses a Fast Fourier Transform branch so that attention can act on frequency components that have no obvious spatial-domain counterpart.","core_discovery":"On its own terms, the paper's discovery is that a purely convolutional encoder-bottleneck-decoder, UIR-PolyKernel, can reach state-of-the-art underwater restoration results. At the bottleneck, the Composite Shape Convolution module runs four parallel depthwise convolutions (31x1, 1x31, 31x31, and 1x1) on quarter-resolution features to model global context and anisotropic scattering; at two intermediate scales, Large Kernel Attention modules extend receptive fields; at full resolution, the Hybrid Domain Attention module applies an FFT branch, pixel attention, and channel attention. Trained on 4800 real underwater images with a loss combining SmoothL1, SSIM, and UCIQE, the model achieves the best or near-best PSNR, SSIM, LPIPS, and UCIQE across four benchmarks, including a 26.42 PSNR on EUVP, while the ablation study attributes the largest single gain to the frequency-domain FDPA block.","pith_inferences":["Because the benchmark references are themselves algorithm-generated outputs, the stated PSNR, SSIM, and LPIPS gains may measure fidelity to those machine targets rather than to true clean scenes; a physically grounded reference benchmark could change the ranking.","The same polymorphic-kernel and frequency-spatial attention recipe should transfer to other degradation tasks such as dehazing, deraining, or low-light enhancement, where anisotropic degradation and subtle frequency cues also dominate.","A straightforward extension would be a human perceptual study on the restored images; if human preference tracks the reported metrics, the method's practical value is confirmed despite the imperfect references."],"forward_implications":["UIR-PolyKernel's per-image cost of 13.67 G MACs and 1.84M parameters makes near-state-of-the-art restoration feasible on edge devices.","The design suggests that large, heterogeneous kernel shapes can replace self-attention for long-range dependencies in low-level vision.","Frequency-domain pixel attention can expose degradation cues invisible in the spatial domain, improving robustness across different water types.","The composite loss that includes UCIQE nudges outputs toward perceptually good underwater quality beyond pixel fidelity."],"supporting_citations":[{"why":"Supplies the extremely large kernel convolution design that the CSC module adapts to lower cost.","marker":"[43]"},{"why":"Provides the strip-shaped separable kernel attention idea used to replace costly square kernels.","marker":"[44]"},{"why":"Defines the large kernel attention module pattern that the LKA modules follow.","marker":"[41]"},{"why":"Supplies the paired UIEB training and test underwater images and their reference targets.","marker":"[45]"},{"why":"Supplies the paired EUVP training and test images used in the experiments.","marker":"[46]"},{"why":"Introduces the LSUI dataset and the U-Transformer baseline that the paper compares against and exceeds.","marker":"[15]"},{"why":"Is the transformer baseline URSCT, a main competitor on all three paired test sets.","marker":"[16]"},{"why":"Is the semi-supervised baseline Semi-UIR, the only method that beats the paper's PSNR on UIEB.","marker":"[17]"},{"why":"Supplies the unpaired RUIE test set used to evaluate generalization to unseen water types.","marker":"[47]"}],"fun_headline_variants":["Lightweight pure CNN tops underwater restoration benchmarks","Tiny CNN with large polymorphic kernels wins underwater fix","1.84M-param CNN outshines heavy transformers underwater","Polymorphic large-kernel CNN: SOTA underwater at low cost","Underwater restoration: small CNN, big kernels, top results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported quality scores assume the reference images used for training and evaluation are trustworthy stand-ins for true clean underwater scenes, even though those references are themselves outputs of other algorithms rather than measured ground truth.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight pure CNN tops underwater restoration benchmarks","Tiny CNN with large polymorphic kernels wins underwater fix","1.84M-param CNN outshines heavy transformers underwater","Polymorphic large-kernel CNN: SOTA underwater at low cost","Underwater restoration: small CNN, big kernels, top results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1512,"prompt_tokens":1014,"completion_tokens":498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":414}},"tokens_in":630,"tokens_out":498,"duration_ms":4669,"temperature":1.0,"reasoning_tokens":414,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:42:06.610410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of underwater scenes photographed with known color-calibration targets at measured depths, use the measured target values as true references, and recompute PSNR, SSIM, and LPIPS for UIR-PolyKernel and the transformer baselines; if the margin over the baselines disappears or reverses, the reported superiority is an artifact of algorithm-generated references.","supporting_citations":[{"cited_title":"Omni-kernel network for image restoration,","cited_arxiv_id":null,"evidence_quote":"Supplies the extremely large kernel convolution design that the CSC module adapts to lower cost."},{"cited_title":"Large separable kernel attention: Rethinking the large kernel attention design in cnn,","cited_arxiv_id":null,"evidence_quote":"Provides the strip-shaped separable kernel attention idea used to replace costly square kernels."},{"cited_title":"Visual attention network,","cited_arxiv_id":null,"evidence_quote":"Defines the large kernel attention module pattern that the LKA modules follow."},{"cited_title":"An underwater image enhancement benchmark dataset and beyond,","cited_arxiv_id":null,"evidence_quote":"Supplies the paired UIEB training and test underwater images and their reference targets."},{"cited_title":"Fast underwater image enhancement for improved visual perception,","cited_arxiv_id":null,"evidence_quote":"Supplies the paired EUVP training and test images used in the experiments."},{"cited_title":"U-shape transformer for underwater image enhancement,","cited_arxiv_id":null,"evidence_quote":"Introduces the LSUI dataset and the U-Transformer baseline that the paper compares against and exceeds."},{"cited_title":"Re- inforced swin-convs transformer for simultaneous underwater sensing scene image enhancement and super-resolution,","cited_arxiv_id":null,"evidence_quote":"Is the transformer baseline URSCT, a main competitor on all three paired test sets."},{"cited_title":"Contrastive semi- supervised learning for underwater image restoration via reliable bank,","cited_arxiv_id":null,"evidence_quote":"Is the semi-supervised baseline Semi-UIR, the only method that beats the paper's PSNR on UIEB."},{"cited_title":"Real-world underwater enhancement: Challenges, benchmarks, and solutions under natural light,","cited_arxiv_id":null,"evidence_quote":"Supplies the unpaired RUIE test set used to evaluate generalization to unseen water types."}],"review_version":1}