{"id":"726fed65-1b6d-4b46-aea1-30c8af9025f9","arxiv_id":"2411.12530","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A contourlet-plus-CLIP prompt framework with a spectral fidelity loss is reported to improve infrared image super-resolution and downstream perception, though the evidence has arithmetic and baseline-fairness problems.","lead":"This paper proposes an infrared-specific super-resolution network that combines Contourlet high-frequency subbands, a Fourier-spectrum fidelity loss, and CLIP text prompts to reconstruct thermal images and improve downstream detection and segmentation. It reports top PSNR/SSIM and detection/segmentation scores on M3FD, TNO, and RoadScene benchmarks, but the quantitative comparisons contain inconsistencies and may be unfair to baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's PSNR and MSE are mutually inconsistent (e.g., Ours x2 Set5: 48.637 dB vs MSE 4.470 implies 41.63 dB), so the SOTA quantitative claim lacks reliable support.","rationale":"After reading the paper in good faith, the central claim is that the proposed method achieves SOTA infrared SR quality and improves downstream perception. For that to be true, the quantitative evidence in Table 1 must be reliable. The reported PSNR and MSE values are mathematically incompatible under any standard definition, which is a verifiable, internal inconsistency rather than a matter of interpretation. I checked multiple rows: the implied PSNR from MSE is 6-7 dB lower than the reported PSNR. This is far beyond rounding or evaluation-protocol differences. It suggests a bug in one of the metric computations or a transcription error, and makes the SOTA claim unsupported. The downstream results (Tables 2-3) lack statistical significance, and the baseline-retraining protocol is ambiguous, but those are secondary once the primary SR numbers are known to be incoherent. The paper does provide code and a detailed architecture, and the ablations are internally consistent, so there is no evidence of fraud; the concern is strictly about the correctness of the empirical validation. I agree with the reader's REJECT verdict, though I weigh the metric inconsistency more heavily than the baseline-fairness point. A single deterministic arithmetic check can settle whether my concern lands; if the authors can show a consistent definition that reconciles the numbers (e.g., a nonstandard PSNR formula), the verdict could be revisited.","tokens_in":23688,"tokens_out":6995,"duration_ms":61960,"concrete_test":"Analytical check: For every row of Table 1, compute PSNR_from_MSE = 10*log10(255^2/MSE) and compare to the reported PSNR. If the maximum absolute deviation exceeds 0.5 dB across all rows, the table is internally inconsistent under the standard 8-bit definition; also test MAX=1 (normalized) and MAX=1023 to see if any single definition reconciles all rows. These formulas are deterministic and require no code. If the table fails this check, reproduce the released code's evaluation on M3FD Set5 to verify which metric (or both) is misreported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports PSNR and MSE for every method and dataset. For 8-bit images, PSNR = 10*log10(255^2/MSE). Plugging in the reported values yields systematic contradictions: for Ours x2 Set5, 10*log10(65025/4.470) = 41.63 dB, not 48.637 dB; for SwinIR x2 Set5, 10*log10(65025/5.121) = 41.04 dB, not 46.880 dB. The gap persists across all rows and scales (x2 rows: ~7 dB; x4 rows: ~6 dB). No constant MAX (255, 1023, 65535) or color-plane convention explains the discrepancy: if the MSE were computed on [0,1]-normalized images, values >1 would be impossible; if computed on a different bit depth, the implied PSNR would shift uniformly, not vary by scale. Thus at least one of the two metrics is not computed as stated. Because the paper's central claim of state-of-the-art PSNR/SSIM is based on this table, the quantitative comparison is invalid as reported. The inconsistency also propagates to the downstream experiments: Tables 2 and 3 use the same reconstructed images, and their gains (e.g., mAP 0.408 vs 0.388, mIoU 57.24 vs 56.58) are reported without error bars or significance tests, so they could reflect noise. The reader's concern about baseline retraining is legitimate and unresolved (§4.2 never states whether the SR baselines were retrained on infrared data; §4.3 retrains only the detector), but the arithmetic inconsistency alone is sufficient to reject the headline claim. The architecture and ablations (Tables 4-6) are plausible, but they validate components against a flawed primary metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an infrared-specific super-resolution framework ('CoRPLE') combining a Contourlet Refinement Gate with global-local interactive attention, a Spectral Fidelity Loss defined on normalized log-magnitude Fourier spectra, and a two-stage CLIP prompt-learning degradation loss. The authors report state-of-the-art PSNR/SSIM on M3FD benchmark sets and improved YOLOv5 detection and SegFormer segmentation when their SR outputs are used as inputs, with ablation studies of the attention blocks, contourlet levels, prompt choices, and loss components.","tokens_in":24109,"tokens_out":11880,"duration_ms":110937,"significance":"If the evaluation were reliable, the work would be a useful empirical contribution to infrared SR: it targets a real domain gap between RGB and infrared super-resolution, introduces a concrete frequency-domain regularization in Eq. (9), and extends the authors' ECCV work with a gating mechanism and downstream perception evaluation. The public code link is a positive reproducibility feature. However, the central quantitative comparison is currently undermined by an unresolved arithmetic inconsistency in Table 1 and by an underspecified baseline training protocol, so the stated significance is not established in this version.","major_comments":[{"comment":"The PSNR and MSE columns are mutually inconsistent across essentially every row. For 8-bit images, PSNR = 10 log10(255^2/MSE). In the row 'Ours x4 Set5', PSNR 40.784 dB corresponds to MSE ≈ 5.43, but the table lists MSE 31.217 (which itself corresponds to PSNR ≈ 33.2 dB). In 'Ours x2 Set5', PSNR 48.637 dB corresponds to MSE ≈ 0.89, while the listed MSE 4.470 implies PSNR ≈ 41.6 dB. No pixel-range convention (255, 1023, 65535, or [0,1] normalization) can reconcile all rows, because the discrepancy changes with scale and method. Since the headline state-of-the-art claim is drawn directly from this table, the quantitative evaluation must be recomputed and carefully re-reported before any conclusions can be drawn.","section":"§4.2, Table 1"},{"comment":"The paper compares with nine SR methods but does not state whether any of them was retrained on the M3FD infrared training set (182 images). Section 4.1 gives implementation details only for the proposed network, while Sections 4.3–4.4 mention retraining only for YOLOv5 and SegFormer. If the SR baselines are used with weights pretrained on RGB data, the reported gains could be due to domain adaptation rather than to the proposed architecture or loss. The authors must state the training protocol for every baseline and, if those baselines were not retrained, provide experiments in which all methods are trained on the same infrared data or justify the pretrained-weight comparison.","section":"§4.1–4.2, baseline training protocol"},{"comment":"The downstream detection and segmentation results are reported as single deterministic numbers with no error bars, multiple runs, or statistical tests. Improvements such as mAP 0.408 vs 0.388 (Ours x2) and mIoU 57.24 vs 56.58 (Ours x2) are presented as 'significant enhancements,' but the paper provides no information about run-to-run variability, so these gaps cannot be distinguished from noise. Moreover, these experiments use the same SR reconstructions as Table 1 and therefore inherit its unresolved metric problem. The authors should provide variance estimates and significance testing, and should resolve the asymmetry between the documented training protocol for downstream networks and the undocumented training protocol for the SR baselines.","section":"§4.3–4.4, Tables 2 and 3"}],"minor_comments":[{"comment":"The narrative that spectral-distribution preservation is independently 'discovered' should be softened: Eq. (9) explicitly enforces matching of normalized log-magnitude Fourier spectra as part of the training loss. The ablations show that including this penalty helps, which is a legitimate empirical finding, but the text should not present the matching as an observed outcome rather than as the explicit objective.","section":"§3.6, Eq. (9) and §4.2"},{"comment":"The sentence about falling behind HAT-L and EDT-T at the ×4 scale appears inconsistent with Table 3, where Ours has the highest mIoU (56.89) versus HAT-L (55.63) and EDT-T (55.67); if the intended comparison is for a specific category, that category should be named.","section":"§4.4, text after Table 3"},{"comment":"The caption for Fig. 6 appears twice in the text; the duplicate should be removed.","section":"Figure 6"},{"comment":"The total loss is written as a plain sum of four terms with no weighting coefficients; the paper should state whether all weights are set to 1 and whether they were tuned.","section":"§3.6, Eq. (10)"},{"comment":"Entries such as '0.1910.509' and '0.408x4' are missing delimiters and should be reformatted for readability.","section":"Table 2, Ours x2 row"},{"comment":"The notation in Eq. (6) uses I_i in the denominator while Eq. (5) uses I; the indexing and the role of the paired images should be clarified.","section":"§3.6, Eqs. (5)–(6)"}],"recommendation":"reject","confidential_remarks":"This is a borderline case between major revision and rejection. The PSNR/MSE contradiction is pervasive and unexplained, and the baseline training protocol is undocumented, so the current version's central claims are not supported. I would be willing to look at a resubmission with corrected, complete experimental reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the infrared SR paper. Bottom line: the architecture is plausible and the domain-specific motivation is reasonable, but the headline quantitative claim is not supported as written. Table 1 has a systematic internal inconsistency between PSNR and MSE — for example, Ours x2 Set5 reports PSNR 48.637 and MSE 4.470, but 10*log10(255^2/4.470) is about 41.63 dB, a 7 dB gap. The same pattern holds across nearly every row and both scales. No constant bit depth or normalization convention explains it, so at least one of the two metrics is not computed as stated. The core SOTA comparison rests on that table.\n\nWhat is genuinely new: the Spectral Fidelity Loss (a frequency-domain MSE regularization), the global-local interactive attention block, and the downstream detection/segmentation evaluation. The paper honestly identifies itself as an extension of the authors' ECCV 2024 contourlet/prompt work, so the novelty is incremental but real. The ablations in Tables 4-6 are internally consistent and suggest the contourlet gate and prompt learning do something. The code is public.\n\nSoft spots beyond the metric problem: the paper never states whether the SR baselines were retrained on infrared data. §4.2 lists methods but not the training protocol; §4.3 retrains only YOLOv5. If the baselines are evaluated with RGB-pretrained weights, the comparison is unfair and the reported gains could reflect domain shift. The downstream tables have no error bars or significance tests, and the gains are small (mAP 0.408 vs 0.388). The spectral fidelity loss is also partly definitional — it directly minimizes spectrum distance, so \"preserving spectral distribution\" is enforced by construction rather than independently discovered. That is acceptable as a loss, but it should not be advertised as an empirical finding.\n\nWho this is for: researchers working on infrared-specific SR. The idea of contourlet-domain gating plus spectral regularization is worth exploring. But as it stands, the central claim of state-of-the-art performance should not be taken at face value.\n\nRecommendation: send to peer review, but expect heavy revision. The metric inconsistency requires the authors to recompute or correct the entire table, and the baseline protocol needs to be stated explicitly. After that, the paper could become a modest but solid contribution. I would not cite it in its current form.","headline":"Plausible infrared SR architecture, but Table 1's PSNR/MSE inconsistency and unclear baseline retraining make the SOTA claim unsupported as written.","tokens_in":24667,"tokens_out":1813,"would_cite":false,"duration_ms":17765,"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 infrared super-resolution should preserve the thermal spectrum distribution, and that a Contourlet refinement gate plus a Spectral Fidelity Loss achieves state-of-the-art PSNR/SSIM while also improving downstream…","keywords":["infrared image super-resolution","Contourlet transform","spectral fidelity loss","two-stage prompt learning","thermal spectrum distribution","downstream detection","semantic segmentation","CLIP-guided degradation loss"],"falsifier":"Retrain all nine comparison SR models on the identical M3FD infrared training set with the same bicubic degradation, then recompute the PSNR/SSIM tables and the detection/segmentation evaluations; if the baselines close the gaps or surpass the proposed method, the claimed advantage of the Contourlet gate and spectral fidelity loss is not supported.","tokens_in":23470,"feed_emoji":"🌡️","tokens_out":5142,"duration_ms":42091,"temperature":0.7,"pith_summary":"Infrared images carry less high-frequency detail and follow different degradation patterns than visible-light images, yet most super-resolution models are built and trained for RGB. This paper argues that applying such models to infrared images distorts the spectral distribution and harms downstream machine perception. To fix that, the authors propose a framework with a Contourlet Refinement Gate that restores high-frequency subbands from multi-scale and multi-directional infrared decomposition, and a Spectral Fidelity Loss that matches the normalized Fourier magnitude spectrum of the super-resolved image to the ground truth. They also add a two-stage CLIP prompt-learning loss so the network is guided toward infrared HR characteristics rather than toward generic RGB texture. The reported experiments show improvements in PSNR/SSIM on M3FD, and better YOLOv5 detection and SegFormer segmentation, compared with nine existing SR methods.","feed_headline":"Infrared-aware SR beats RGB models by guarding thermal spectra","feed_subtitle":"A Contourlet gate and spectral fidelity loss keep thermal distributions intact, lifting SR quality and downstream detection and…","key_machinery":"The Contourlet Refinement Gate (CRG) is the central module: it applies the Contourlet Transform—a multi-scale, multi-directional decomposition via Laplacian pyramid and directional filter banks—to deep features, enhancing the high-pass subbands that infrared images lack, then fuses them back through a residual connection with a Global-Local Interactive Attention block. The other load-bearing piece is the Spectral Fidelity Loss, which takes 2D DFT magnitude spectra of SR and HR images, compresses and normalizes them, and penalizes their mean-squared difference, thereby constraining the frequency distribution. A two-stage CLIP prompt-learning loss supplies semantic guidance by first refining positive/negative text prompts in the CLIP latent space and then locking the text encoder to drive the degradation loss.","core_discovery":"The central claim is that preserving infrared spectral distribution fidelity, not just pixel fidelity, is the key to infrared image super-resolution. The paper's discovery, on its own terms, is that regularizing the high- and low-frequency thermal spectrum during reconstruction—via the Spectral Fidelity Loss—combined with extracting modal-specific features through a Contourlet-based gate, lets a single SR model outperform RGB-oriented transformers and diffusion methods while also improving detection and segmentation. The authors assert they are the first to emphasize infrared spectral distribution fidelity for learning the infrared LR-HR mapping.","pith_inferences":["A testable extension would be to apply the same spectral fidelity loss and contourlet gating to visible-light SR to see whether the gains are specific to infrared or generalize to any domain with band-limited spectra.","The paper does not state whether the baseline SR models were retrained on the infrared training set; if they were used with pretrained RGB weights, the reported margins likely conflate domain adaptation with super-resolution quality, and a fair comparison would retrain all baselines on the same M3FD training data.","The spectral fidelity loss is defined on the magnitude spectrum only, discarding phase; since phase carries much of the structural information, an implicit consequence is that the loss mainly enforces power-spectrum shape, and combining it with a phase-sensitive term might further improve alignment with HR structure.","The prompt-learning stage depends on CLIP's semantic alignment, which is built from natural images; its applicability to thermal imagery may degrade for categories or scenes that are underrepresented in CLIP's training data."],"forward_implications":["RGB-oriented SR methods, when applied to infrared images, distort the thermal spectrum distribution; the paper's results imply that infrared SR should be treated as a modal-specific task rather than a drop-in application of visible-light models.","Regularizing the Fourier magnitude spectrum during reconstruction preserves both high- and low-frequency infrared components, and this correlates with better PSNR/SSIM.","Super-resolving infrared images with spectral fidelity also improves YOLOv5 detection mAP and SegFormer segmentation mIoU, meaning SR quality and machine perception move together.","The two-stage prompt-learning strategy, which first refines positive/negative prompts in the CLIP latent space and then locks the text encoder, is a transferable way to inject infrared-aware semantic guidance into the SR loss.","Ablation results show that both the Laplacian pyramid and directional filter bank branches of the Contourlet transform contribute, and that four decomposition levels give the best reconstruction."],"supporting_citations":[{"why":"Supplies the Contourlet Transform used in the CRG module to decompose images into multi-scale, multi-directional subbands.","marker":"Do and Vetterli (2005)"},{"why":"The authors' previous work that first showed restoring high-frequency subbands improves infrared SR, and that this paper extends with prompt learning and spectral fidelity loss.","marker":"Li et al (2024)"},{"why":"SwinIR, a transformer-based RGB SR baseline the paper compares against and argues is mismatched to infrared characteristics.","marker":"Liang et al (2021)"},{"why":"M3FD dataset used for training, validation, and the object detection evaluation.","marker":"Liu et al (2022)"},{"why":"FMB benchmark used for the semantic segmentation evaluation.","marker":"Liu et al (2023)"},{"why":"CLIP model whose latent space the two-stage prompt learning uses to align SR images with positive/negative text prompts.","marker":"Radford et al (2021)"},{"why":"SegFormer, the segmentation model retrained and used to evaluate downstream segmentation performance.","marker":"Xie et al (2021)"},{"why":"TNO dataset, part of the evaluation mixture for infrared SR.","marker":"Toet (2017)"},{"why":"RoadScene dataset, part of the evaluation mixture for infrared SR.","marker":"Xu et al (2020)"}],"fun_headline_variants":["Gate keeps thermal spectra intact for sharper IR super-resolution","Thermal-spectrum guard improves infrared SR for machine vision","Contourlet gate preserves thermal fidelity, outdoes RGB SR models","Spectral fidelity loss lets IR super-resolution beat RGB-only methods","Thermal spectrum guard boosts IR SR and downstream detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the baseline SR models were fairly adapted to infrared data, but the paper only states it retrained the YOLOv5 detector (and SegFormer), never the SR baselines; if the baselines used pretrained RGB weights, the reported superiority could reflect domain shift rather than better super-resolution.","fun_headline_variants_meta":{"raw":{"variants":["Gate keeps thermal spectra intact for sharper IR super-resolution","Thermal-spectrum guard improves infrared SR for machine vision","Contourlet gate preserves thermal fidelity, outdoes RGB SR models","Spectral fidelity loss lets IR super-resolution beat RGB-only methods","Thermal spectrum guard boosts IR SR and downstream detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2771,"prompt_tokens":902,"completion_tokens":1869,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":1787}},"tokens_in":518,"tokens_out":1869,"duration_ms":11167,"temperature":1.0,"reasoning_tokens":1787,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:25:05.807418+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain all nine comparison SR models on the identical M3FD infrared training set with the same bicubic degradation, then recompute the PSNR/SSIM tables and the detection/segmentation evaluations; if the baselines close the gaps or surpass the proposed method, the claimed advantage of the Contourlet gate and spectral fidelity loss is not supported.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5802--5811","cited_arxiv_id":null,"evidence_quote":"M3FD dataset used for training, validation, and the object detection evaluation."},{"cited_title":"In: Proceedings of the International Conference on Machine Learning, pp 8748--8763","cited_arxiv_id":null,"evidence_quote":"CLIP model whose latent space the two-stage prompt learning uses to align SR images with positive/negative text prompts."},{"cited_title":"Advances in Neural Information Processing Systems 34:12,077--12,090","cited_arxiv_id":null,"evidence_quote":"SegFormer, the segmentation model retrained and used to evaluate downstream segmentation performance."},{"cited_title":"Data in Brief 15:249--251","cited_arxiv_id":null,"evidence_quote":"TNO dataset, part of the evaluation mixture for infrared SR."}],"review_version":1}