{"id":"6bae9fd1-6935-4ddc-b612-b61462d6c6f6","arxiv_id":"2411.10682","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stage cascaded underwater image enhancement network using stage-1 outputs as stage-2 negative samples achieves competitive non-reference scores, but not higher PSNR/SSIM than single-stage alternatives.","lead":"CCL-Net enhances underwater images in two stages, first correcting color casts then removing haze, with contrastive learning that uses progressively better negative samples at each stage. It reports top non-reference metric scores on some benchmarks, but its own ablation shows the two-stage model scores lower on PSNR and SSIM than a single-stage variant, so the claimed superiority is not fully supported.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central two-stage claim is contradicted by the paper's own ablation: on UIEB-T90, removing CC-Net raises PSNR from 20.18 to 22.01 and SSIM from 0.866 to 0.905, and the post hoc switch to non-reference metrics does not repair this.","rationale":"The reader's REJECT verdict is supported by the evidence. The strongest load-bearing concern is not merely that PSNR/SSIM are low; it is that the paper's own controlled ablation isolates the proposed components and shows they reduce reference-based fidelity. Removing the color-correction stage improves UIEB-T90 PSNR by 1.8 dB and SSIM by 0.039, and removing contrastive learning also improves both metrics. This directly undercuts the claimed incremental improvement from cascaded design. The paper then changes the evaluation protocol post hoc, leaning on non-reference metrics without error bars and on author-selected subsets, and explicitly asks the reader to combine metrics with subjective visual judgment. That may be a reasonable position in the UIE community, but it is not adequate support for a strong superiority claim in a quantitative comparison table. A proper user study could rehabilitate the visual-quality claim, and the failure cases in Sec. IV-F already concede artifacts on low-light and color-consistency cases. The contribution is a reasonable engineering attempt, but the central claim, as evidenced, is not established. Hence the reader's reject verdict should remain unchanged.","tokens_in":22206,"tokens_out":3958,"duration_ms":41194,"concrete_test":"Conduct a forced-choice paired-comparison user study on the full UIEB-T90 and UIEB-C60 sets, comparing CCL-Net with its own w/o CC-Net and w/o CL ablations, using at least 20 independent raters and reporting per-image preference counts with confidence intervals. If the full model is not preferred over w/o CC-Net at a statistically significant margin (95% CI excluding 50%), the central two-stage claim fails; if it is preferred, the reference-metric discrepancy still needs explanation but the visual claim would be supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the cascaded two-stage architecture with cascaded contrastive learning 'incrementally improve[s] the lower and upper bound' and achieves superior performance (Abstract; Sec. I). The load-bearing evidence is Table III. On UIEB-T90, the full CCL-Net obtains PSNR 20.181 and SSIM 0.866, while removing CC-Net (i.e., training HR-Net directly on raw images) gives PSNR 22.006 and SSIM 0.905; removing contrastive loss gives 21.724/0.891; and replacing the stage-2 negative with the raw image (RAN) gives 21.845/0.890. Thus every ablated variant that the paper argues is worse actually scores higher on the same reference-based metrics used in Table I. The paper's response, declaring PSNR/SSIM unreliable and relying instead on UIQM/UCIQE and subjective visual judgment, does not repair the argument: Table II reports author-selected subsets (SQUID-T16, RUIE-T78), reports no variances or significance tests, and the UIQM advantages are small (e.g., 2.936 vs 2.837 on EUVP; 2.350 vs 2.206 on SQUID). Moreover, the 'lower/upper bound' claim is never directly measured: no intermediate-stage quality bounds are reported, and the contrastive losses in Eqs. 11 and 23 do not by themselves guarantee monotone improvement, especially since Eq. 23 explicitly pushes the final output away from the color-corrected image. Given the paper's own numbers, the central claim of incremental gains from the two-stage cascade and cascaded negatives is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CCL-Net, a two-stage underwater image enhancement network in which a color correction stage (CC-Net) is followed by a haze removal stage (HR-Net), with each stage trained under a hybrid loss that includes a contrastive loss. The raw image serves as the negative sample in the stage-1 contrastive loss, and the stage-1 output serves as the negative sample in the stage-2 contrastive loss. The authors claim that this cascaded design incrementally improves both the lower and upper bounds of the enhancement result and achieves superior performance over state-of-the-art methods on several benchmarks, including UIEB-T90, UIEB-C60, EUVP-T515, SQUID-T16, and RUIE-T78.","tokens_in":22550,"tokens_out":2345,"duration_ms":23665,"significance":"The conceptual idea of using higher-quality intermediate outputs as negative samples in a cascaded contrastive learning framework is interesting and could, in principle, be a useful training strategy for UIE. The paper also provides a large set of experiments, visual comparisons, and an ablation study, and it promises to release source code. However, the central claim of progressive improvement is not supported by the paper's own quantitative results: the full model is worse on the reference-based metrics PSNR/SSIM on UIEB-T90 than every ablated variant that the paper argues is inferior, and the justification for dismissing PSNR/SSIM is inconsistent with their use in the main comparison. As a result, the evidence does not substantiate the claimed advantage of the two-stage cascade or of the specific contrastive negative sampling strategy.","major_comments":[{"comment":"The central claim of incremental improvement is directly contradicted by the paper's own ablation results. On UIEB-T90, the full CCL-Net obtains PSNR 20.181 and SSIM 0.866, while removing CC-Net (w/o CC-Net) gives PSNR 22.006 and SSIM 0.905, removing the contrastive loss (w/o CL) gives 21.724/0.891, and replacing the stage-2 negative with the raw image (RAN) gives 21.845/0.890. Thus every ablated variant that the paper argues is worse actually scores higher on the same reference-based metrics used in Table I to compare with other methods. The paper's response that PSNR/SSIM are unreliable for underwater images is inconsistent, because Table I uses these very metrics to argue that CCL-Net outperforms traditional methods, and the paper does not provide a principled reason why PSNR/SSIM should be trusted for comparisons against other deep learning methods but not for comparisons among its own ablated variants.","section":"Table III vs. Table I, Section IV-D and IV-E"},{"comment":"The paper repeatedly claims that the cascaded contrastive losses 'incrementally improve the lower and upper bound of the enhancement result' and that using the stage-1 output as the stage-2 negative yields 'a better lower bound.' This is never directly measured: no intermediate-stage quality bounds are reported, and no theoretical argument is given. The loss in Eq. (23) explicitly pushes the final output away from the stage-1 output in VGG feature space, so the training objective itself defines 'better' in that learned feature space. The paper does not show that minimizing this loss improves any downstream quality metric or guarantees monotone improvement of image quality. Without such evidence, the 'better lower bound' claim is a design choice, not an empirical finding.","section":"Eqs. (11), (23) and the 'lower/upper bound' claim in Sections I and III"},{"comment":"The generalization claim rests on non-reference metrics computed on author-selected subsets (SQUID-T16, RUIE-T78) with no variances, confidence intervals, or significance tests. The reported UIQM differences are often small (e.g., 2.936 vs. 2.837 on EUVP-T515; 3.168 vs. 3.237 on RUIE-T78 for TACL) and the UCIQE rankings vary across datasets (1st on RUIE-T78 but outside the top three on SQUID-T16). The paper itself argues in Section IV-D that UCIQE 'cannot truly measure human perception' and that non-reference metrics should be combined with subjective judgments, yet the quantitative claim of superiority is based on exactly these metrics. The evidence is therefore insufficient to support the stated conclusion that CCL-Net 'outperforms all single-stage framework deep learning methods except TACL.'","section":"Table II, Section IV-D"}],"minor_comments":[{"comment":"There is a grammatical error in the phrase 'an haze removal sub-network'; it should be 'a haze removal sub-network.'","section":"Section I, Introduction"},{"comment":"The scaling factor s is set to 100 in Eq. (11) but to 1 in Eq. (23), and the choice is described only as 'empirically set.' The paper should report the sensitivity of the results to s, since this is a free hyperparameter that could affect the relative weighting of the contrastive loss.","section":"Eqs. (11) and (23)"},{"comment":"The paper uses UCIQE as a primary metric in Tables I and II but also cites [56] to argue that UCIQE 'cannot truly measure human perception.' This is confusing: if UCIQE is unreliable, the paper should not rank its method on UCIQE without qualification; if it is reliable, then the low rankings on several datasets should be taken seriously.","section":"Section IV-D"},{"comment":"The caption states that 'the numbers on the top-left corner of each image refer to its PSNR/SSIM values,' but in the compiled text these numbers are not visible in the figure. Either the figure should be regenerated with legible numbers or the caption should be clarified.","section":"Figure 5 caption"},{"comment":"The description of the training procedure says the second stage is trained using the 800 enhanced images generated by the first stage, but the ablation 'w/o CC-Net' trains HR-Net directly on raw images. The paper should clarify whether all other training settings (learning rate, epoch, loss weights, etc.) were kept identical for the ablation models to ensure a fair comparison.","section":"Section IV-A, Implementation Details"}],"recommendation":"reject","confidential_remarks":"This is a clear reject for me. The paper's own ablation, Table III, shows that the full model is substantially worse in PSNR/SSIM than the version without the proposed first stage (22.006 vs. 20.181 PSNR on UIEB-T90). The authors' decision to discard PSNR/SSIM for the ablation but use them in the main comparison is not logically consistent. The 'lower/upper bound' claim is asserted but never measured or proven. Even if the method had merit, the manuscript would need a major rewrite with new evidence to support the central claim; as it stands, the contradiction is load-bearing and unfixable within the current scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper has a genuinely novel idea—using the first-stage output as the negative sample for the second-stage contrastive loss—but the reported evidence contradicts the central claim that the cascade improves results. On UIEB-T90, removing CC-Net raises PSNR from 20.18 to 22.01 and SSIM from 0.866 to 0.905. The authors then declare PSNR/SSIM unreliable and lean on non-reference metrics. That is a post hoc switch, and it doesn't hold up without a proper user study.\n\nWhat's actually new: the cascaded negative strategy. Prior contrastive UIE work (TACL, HCLR-Net) uses the raw input as the negative. Using a better-quality intermediate as the negative is a reasonable idea, and the two-stage Lab-color-correction + haze-removal decomposition is sensible. The network is lightweight (0.55M params) and the paper includes a thorough comparison, honest failure cases, and a proper ablation table. Credit where due: the ablation table is exactly what makes the contradiction visible.\n\nThe soft spot is proportionally large. The central claim—incremental improvement of lower/upper bounds—is not measured. No intermediate quality bounds are reported. The contrastive loss in Eq. 23 explicitly pushes the final output away from the color-corrected image, so the training objective itself defines 'better.' The UIQM/UCIQE wins are small, without error bars, and on SQUID and RUIE the test subsets are author-selected. The authors openly acknowledge the PSNR/SSIM deficit and argue those metrics are unreliable, but that argument needs independent support, e.g., a user study. Citation practice looks fine; the relevant prior work is there.\n\nRecommendation: this deserves referee time, not a desk reject. The idea is worth exploring, but the paper as submitted overclaims. A serious referee should demand full-dataset evaluations, error bars, and a real subjective study before the claims can be accepted. I'd send it to review with a clear expectation of major revision, and if the authors can't produce that evidence, reject.","headline":"New twist (stage-1 output as stage-2 negative) undercut by own ablation; PSNR/SSIM deficit is waved away with a post hoc metric switch.","tokens_in":23214,"tokens_out":3314,"would_cite":false,"duration_ms":31992,"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":"This paper claims that underwater images are best enhanced in two stages—color correction in Lab space, then haze removal—with the first stage's output serving as the negative sample in a second contrastive loss, so each stage…","keywords":["underwater image enhancement","cascaded contrastive learning","two-stage network","color correction","haze removal","Lab color space","contrastive loss","UIEB benchmark"],"falsifier":"Run a paired-preference study on UIEB-T90 comparing the full CCL-Net against its own 'w/o CC-Net' ablation: if human viewers do not prefer the two-stage color-then-haze output over the single-stage haze output, the central claim that the cascade improves visual quality would be refuted.","tokens_in":21911,"feed_emoji":"🌊","tokens_out":6355,"duration_ms":55800,"temperature":0.7,"pith_summary":"This paper argues that underwater image enhancement should not be left to one network. It builds two cascaded subnetworks—a color correction network working in Lab color space and a haze removal network that receives the color-corrected image—and trains each with a contrastive loss that pulls the output toward a clear reference and pushes it away from a negative sample. The design's distinctive move is that the second stage's negative sample is the first stage's own output, so the network is trained to make the final image visibly better than the intermediate one. The authors report that this two-stage cascade beats single-stage deep learning methods on non-reference quality metrics (UIQM, UCIQE) and in visual comparisons across five real underwater benchmarks, while acknowledging that reference-based PSNR and SSIM scores are not the best.","feed_headline":"Cascaded contrastive learning lifts underwater image quality step by step","feed_subtitle":"Stage one fixes color cast; stage two treats that fix as a negative example to force clearer, higher-contrast output.","key_machinery":"The machinery is a two-stage cascade: CC-Net operates on the $a,b$ channels of the Lab color space, learns a residual color difference $\\Delta(I_{ab})$, and leaves the luminance channel $I_L$ untouched; HR-Net is an attention-based multi-scale fusion network using Selective Kernel Feature Fusion blocks to remove haze. The load-bearing training mechanism is the cascaded contrastive loss: in stage one, the raw image is the negative sample, and in stage two the color-corrected stage-one output $I^{CC}$ becomes the negative sample, with both contrastive losses comparing VGG-19 feature distances in Eqs. (11) and (23) via ratios $\\|E_i(I^{out})-E_i(I^{REF})\\|_1 / \\|E_i(I^{out})-E_i(I^{neg})\\|_1$. This forces the output closer to the reference and farther from the intermediate result, which is what the paper claims incrementally improves the lower and upper bounds of enhancement.","core_discovery":"The central claim is that a divide-and-conquer cascade with cascaded contrastive learning progressively lifts the quality of underwater images: CC-Net removes color cast while leaving luminance untouched, and HR-Net then removes haze from the color-corrected image. Each stage is constrained by a hybrid loss: an explicit color or SSIM loss plus a contrastive loss built on VGG-19 features, where the negative sample for stage two is the stage-one output rather than the raw underwater image. The authors contend that using a better-quality negative sample raises the lower bound of enhancement, so the final result is guaranteed to improve on the intermediate color-corrected image, and they support this with visual comparisons and UIQM/UCIQE scores on UIEB-T90, UIEB-C60, EUVP-T515, SQUID-T16, and RUIE-T78.","pith_inferences":["The cascaded-negative recipe is a general training strategy: any restoration task with two ordered degradations (for instance low light followed by haze) could reuse the same idea of making the first stage's output the negative sample for the second stage.","The paper's ablation suggests the color-then-haze order is chosen for visual quality, not reference fidelity: removing CC-Net raises UIEB-T90 PSNR from 20.18 to 22.01 and SSIM from 0.866 to 0.905, so a deployment that optimizes PSNR/SSIM might prefer the single-stage haze network.","A natural test of the lower-bound argument is to vary the negative-sample curriculum—for example, using the raw image, the color-corrected image, or an exponential moving average of intermediate outputs as negatives—and measure whether UIQM/UCIQE and human preference move monotonically with negative-sample quality."],"forward_implications":["Dividing underwater degradation into a color-correction stage and a haze-removal stage, each with its own tailored loss, is more effective than a single network trying to balance several losses at once.","Using the stage-one color-corrected output as the negative sample in the second contrastive loss ensures the final image is trained to be better than the intermediate one, raising the lower bound of enhancement.","The full model achieves the highest UIQM among all compared methods on UIEB-T90 and on EUVP-T515 and SQUID-T16, and is always in the top four for UCIQE across the four generalization datasets, supporting the claim of robustness across diverse real underwater scenes.","With 0.55M parameters and 0.06s running time, the two-stage network is comparable to or faster than several single-stage deep models, so the cascade does not impose a prohibitive computational cost."],"supporting_citations":[{"why":"Supplies the 800 paired raw-reference training images and the UIEB-T90 and UIEB-C60 test sets used for training and evaluation.","marker":"[20]"},{"why":"Provides the contrastive-loss formulation for low-level restoration, using degraded and clear images as negative and positive samples.","marker":"[29]"},{"why":"Supplies the pre-trained VGG-19 network whose 1st, 3rd, 5th, 9th, and 13th layer features define both contrastive losses.","marker":"[32]"},{"why":"Supplies the Feature Attention Block used in CC-Net for spatial and channel attention.","marker":"[45]"},{"why":"Supplies the Selective Kernel Feature Fusion module used in HR-Net for multi-scale feature integration.","marker":"[46]"},{"why":"Supplies the SSIM loss used in the haze removal stage and the SSIM metric used for quantitative evaluation.","marker":"[47]"},{"why":"A prior deep underwater enhancement method used both as a comparison baseline and as the source of the SSIM-loss idea.","marker":"[19]"},{"why":"A strong single-stage probabilistic comparison method whose uncertainty-driven approach is contrasted with the proposed cascade.","marker":"[36]"},{"why":"A twin adversarial contrastive learning method used as the key contrastive-learning baseline in the underwater setting.","marker":"[44]"}],"fun_headline_variants":["Two-stage cascade boosts underwater clarity via contrastive learning","Cascaded contrastive net removes color cast then haze stepwise","Progressive underwater enhancement with cascaded contrastive loss","Color-correct then de-haze: contrastive learning guides each step","Stepwise underwater image improvement via cascaded contrastive learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that color cast and haze are separable degradations, and that making the first stage's color-corrected output the second stage's negative sample improves the final image rather than merely steering it away from the reference.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage cascade boosts underwater clarity via contrastive learning","Cascaded contrastive net removes color cast then haze stepwise","Progressive underwater enhancement with cascaded contrastive loss","Color-correct then de-haze: contrastive learning guides each step","Stepwise underwater image improvement via cascaded contrastive learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000567,"raw_usage":{"total_tokens":2724,"prompt_tokens":1025,"completion_tokens":1699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":1615}},"tokens_in":641,"tokens_out":1699,"duration_ms":13620,"temperature":1.0,"reasoning_tokens":1615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:27:12.559950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a paired-preference study on UIEB-T90 comparing the full CCL-Net against its own 'w/o CC-Net' ablation: if human viewers do not prefer the two-stage color-then-haze output over the single-stage haze output, the central claim that the cascade improves visual quality would be refuted.","supporting_citations":[{"cited_title":"FFA-Net: Feature fusion attention network for single image dehazing,","cited_arxiv_id":null,"evidence_quote":"Supplies the Feature Attention Block used in CC-Net for spatial and channel attention."},{"cited_title":"Learning enriched features for fast image restoration and enhancement,","cited_arxiv_id":null,"evidence_quote":"Supplies the Selective Kernel Feature Fusion module used in HR-Net for multi-scale feature integration."}],"review_version":1}