{"id":"711d9004-c978-49e4-9a8b-6e953d067c10","arxiv_id":"2502.09000","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"RTF-Net, combining residual blocks with a convolutional vision transformer, reports the highest PSNR for salt-and-pepper denoising on most tested images.","lead":"This paper introduces RTF-Net, a two-stage network that uses residual blocks to suppress salt-and-pepper noise and a convolutional vision transformer to enhance image structure. On four classic test images, it reports PSNR gains over six earlier methods, but releases no code, error bars, or ablation study.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of superiority on BSD300 rests on Table II entries (44.56/38.03/34.96 dB) that are internally implausible and lack an aggregation protocol; this is the load-bearing weak point.","rationale":"The reader's weakest_assumption correctly identifies the comparison table as the foundation of the central claim. I go one step further and point to a specific, checkable anomaly: the BSD300 values in Table II. These values are not merely under-specified; they are internally inconsistent with the paper's own Lena results and with typical behavior of denoising methods on a diverse dataset. A 44.56 dB average on BSD300 at 30% S&P noise, when the same method obtains 38.87 dB on Lena, is suspicious; if anything, Lena is an easier target. The baseline PARIGI values below the noisy-input PSNR also suggest a protocol artifact. Because the abstract's 'superior in all cases except one' depends on the BSD300 row, this is the load-bearing weak point. A single reproducibility experiment can settle it. This does not change the reader's conditional verdict, but it sharpens the condition: the authors must release code/data and verify the BSD300 numbers before the claim can be accepted.","tokens_in":7551,"tokens_out":5907,"duration_ms":58123,"concrete_test":"Obtain or reimplement RTF-Net and evaluate it on all 300 BSD300 images at p=0.3, 0.5, and 0.7 using the same inference and noise generation described for Lena (Eq. 1), averaging PSNR over the dataset and over 10 random noise seeds. If the mean differs from 44.56/38.03/34.96 by more than 1 dB, the BSD300 entries in Table II are not reproducible and the 'superior in all cases' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RTF-Net outperforms six baselines on Lena, Bridge, Pepper, and BSD300 at 30%, 50%, and 70% salt-and-pepper noise, with only one exception. The BSD300 rows in Table II are the most load-bearing part of that claim: they show RTF-Net at 44.56/38.03/34.96 dB, which is 13–15 dB above the best baseline and about 5–6 dB above RTF-Net's own results on Lena at the same noise levels. For a diverse 300-image dataset, average PSNR is typically no higher than on a single clean image such as Lena, so this internal pattern is suspicious. The paper gives no aggregation rule for BSD300 (average across images, median, or per-image?), no noise-realization seed, and no code. The baseline PARIGI values (12.04/6.01/5.42 dB) fall below the PSNR of a 30%-corrupted noisy image (~11.3 dB), indicating that the comparison protocol or transcription is not standard. If the BSD300 numbers are wrong or obtained under a different protocol, the claimed superiority is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes RTF-Net, a two-stage network for blind salt-and-pepper image denoising. The first stage, a ResNet-based Noise Suppression Network, estimates a noise map that is subtracted from the noisy input, and the second stage, a Convolutional Vision Transformer based Structure Enhancement Network, adds back structural details. The model is trained on DIV2K patches at 30%, 50%, and 70% noise and evaluated on Lena, Bridge, Pepper, and BSD300. The paper's central claim is that RTF-Net outperforms DBA, NASNLM, PARIGI, NLSF, NLSF-MLP, and NLSF-CNN in all tested configurations except Pepper at 30% noise, where NLSF-CNN is 1.29 dB better.","tokens_in":7824,"tokens_out":7446,"duration_ms":69545,"significance":"RTF-Net is a straightforward combination of two established components: residual CNN blocks for noise-map learning and CvT blocks for structural enhancement. If the reported gains were reproducible under a common protocol, the method would be a modest but useful engineering contribution to impulse-noise removal. The paper provides explicit architecture equations, training curves, and a clear noise model. However, the claimed superiority is currently supported only by single-run PSNR numbers with no ablations, no statistical analysis, no code, and no evaluation protocol. The BSD300 entries in Table II are internally implausible, and the comparison therefore does not yet establish the central claim.","major_comments":[{"comment":"The BSD300 rows are the load-bearing part of the central claim, but the paper gives no aggregation protocol. It does not state whether the reported 44.56/38.03/34.96 dB values are means over all 300 images, medians, or per-image selections, nor whether BSD300 was converted to grayscale or processed per color channel. A mean over a diverse 300-image set being 13-15 dB above the best baseline and 5-6 dB above the method's own Lena results at the same noise levels is internally implausible. The PARIGI entries (12.04/6.01/5.42 dB) fall below the PSNR of a 50% or 70% salt-and-pepper corrupted image, indicating a protocol or transcription mismatch. These rows must be recomputed and described before the claimed superiority is credible.","section":"Section III.D / Table II"},{"comment":"There is a large unexplained gap between the reported validation PSNR and the test PSNR at the same noise level. For 30% noise, Figure 7 reports a final validation PSNR of 32.689 dB, while Table II reports 38.87 dB on Lena and 44.56 dB on BSD300. The authors do not explain why the model would generalize 6-12 dB above its DIV2K validation average; this gap further undermines confidence in the Table II entries.","section":"Section III.A-C and III.D"},{"comment":"The contribution of the proposed fusion is not isolated. No ablation removes the SEN stage, the NSN stage, or replaces the CvT with a plain transformer or the ResNet with a plain CNN. The closest architectural predecessor, the authors' earlier CvT denoiser [7], is cited but not included as a baseline. Without such comparisons, the paper cannot support the claim that the specific NSN+SEN combination is responsible for the reported improvement.","section":"Section II.B / III.D"},{"comment":"All reported PSNR values are single deterministic-looking numbers with no random seeds, no repeated trials, and no error bars. Some reported margins are small, such as the 0.24 dB difference on Pepper at 50% noise (30.47 vs 30.23), and the single exception on Pepper at 30% noise could easily change with a different noise realization. The 'excels in all cases except one' claim therefore cannot be statistically assessed from the data presented.","section":"Section III.D"}],"minor_comments":[{"comment":"The hyperparameter table lists 'NSN Features' twice; the second occurrence should presumably read 'SEN Features'. The duplicate row makes the architecture specification ambiguous.","section":"Table I"},{"comment":"There are several typographical and language issues: 'RT-Net' should be 'RTF-Net', 'y mage' should be 'y image', and the Indonesian 'dan' appears in the Conclusions ('NLSF-CNN [13] dan NLSF-CNN [13]').","section":"Section II.B / IV"},{"comment":"The discussion of deblurring, JPEG artifact removal, and super-resolution uses different images (Airplane, Parrot, Girl) and is not part of the salt-and-pepper denoising evaluation. The connection to the central claim is unexplained, and the captions use verbs like 'deblurred' and 'artifact removed' that do not match the denoising task. This material should be removed or explicitly framed as unrelated pilot experiments.","section":"Section III.E / Figures 12-14"},{"comment":"The figure captions are confusing: subfigures (c)-(e) show NSN outputs at 30%, 50%, and 70% noise, then (f)-(g) show noisy inputs at 50% and 70%, out of sequence. Reordering the panels or clarifying the intermediate outputs would help the reader.","section":"Section III.D / Figures 15-17"},{"comment":"No code, trained model, or data-generation script is provided. Given that the comparison depends on exact noise realizations, releasing the evaluation code with fixed seeds is important for reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early draft: the hyperparameter table has a duplicated row, the conclusions contain an Indonesian word, and Figures 12-14 show experiments on deblurring/JPEG/SISR that are unrelated to the denoising claim. The BSD300 entries in Table II are the main worry; I would ask the editor to have the authors supply the raw per-image results and the aggregation script before any further consideration, since a transcription or protocol mismatch would invalidate the headline claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper proposes RTF-Net, a two-stage salt-and-pepper denoiser. A residual-block network (NSN) estimates a noise map, then a CvT-based network (SEN) adds back structure. The combination is new, and the components are clearly described. There's no new principle, but the design is sensible.\n\nWhat's good: the architecture is simple enough to reproduce from the text, training uses DIV2K, and the paper reports PSNR on Lena, Bridge, Pepper, and BSD300 at three noise levels. The visual examples look plausible.\n\nWhere it falls down: the central superiority claim is not supported. No error bars, no repeated trials, no statistical tests. No ablations and no plain CvT or plain ResNet baseline, so you cannot tell whether the two-stage fusion is what helps. Most damaging, the BSD300 rows in Table II are internally implausible: RTF-Net gets 44.56 dB on BSD300 at 30% noise while getting 38.87 dB on Lena at the same noise level. A 300-image average should not be 5 dB higher than a single clean image. The PARIGI entries (12.04, 6.01, 5.42 dB) are at or below the PSNR of the corrupted images themselves, suggesting the comparison protocol or transcription is off. The paper gives no aggregation rule for BSD300, no noise seeds, and no code.\n\nSo the claim that the method 'excels in all cases except one' collapses on the BSD300 rows. The other comparisons are single-run PSNRs with no variance, suggestive but not conclusive. The authors also omit a comparison against their own earlier CvT denoiser [7], which a referee would rightly ask for.\n\nBottom line: the architecture is a reasonable incremental idea and the writing is clear. But the evaluation as written is not credible, and the BSD300 numbers need to be corrected or removed. With code, seeds, per-image results, and ablations, this could become a decent workshop or short conference paper. As is, I would not send it to peer review.\n\nRecommendation: desk reject, with an invitation to resubmit after fixing the evaluation.","headline":"A plausible two-stage ResNet/CvT denoiser whose claimed superiority rests on an unverifiable and internally suspicious BSD300 comparison table.","tokens_in":8360,"tokens_out":3392,"would_cite":false,"duration_ms":32897,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Residual Transformer Fusion Network (RTF-Net) combines a residual-block noise suppressor with a convolutional vision transformer enhancer and reports higher PSNR than six existing salt-and-pepper denoisers in 11 of 12 tested cases.","keywords":["image denoising","salt-and-pepper noise","Residual Transformer Fusion Network","Convolutional Vision Transformer","residual network","PSNR comparison","Noise Suppression Network","Structure Enhancement Network"],"falsifier":"Regenerate salt-and-pepper noise on the same test images with the paper's own noise model and rerun the six baseline methods and RTF-Net under identical random seeds; if the baseline PSNR values differ from Table II by more than rounding, or if any baseline matches or beats RTF-Net outside the reported Pepper-30% case on the regenerated noise, the central claim fails.","tokens_in":7382,"feed_emoji":"🖼️","tokens_out":5018,"duration_ms":44539,"temperature":0.7,"pith_summary":"This paper proposes RTF-Net, a two-stage neural architecture for removing salt-and-pepper noise from images without needing prior knowledge of the noise model or level. The first stage, a Noise Suppression Network built from residual blocks, estimates a noise map that is subtracted from the noisy input. The second stage, a Structure Enhancement Network built from a Convolutional Vision Transformer, adds back the fine details lost in the first stage. The authors report that RTF-Net achieves higher PSNR than six existing salt-and-pepper denoisers on Lena, Bridge, Pepper, and BSD300 test images at 30, 50, and 70 percent noise, with one exception: on Pepper at 30 percent noise, NLSF-CNN scores 32.99 dB versus 31.70 dB for the proposed method.","feed_headline":"Hybrid transformer net tops PSNR table for salt-and-pepper noise","feed_subtitle":"RTF-Net wins 11 of 12 PSNR comparisons on Lena, Bridge, Pepper, and BSD300 at 30–70% noise.","key_machinery":"The central identity is the two-stage fusion $$I^* = y + \\mathrm{SEN}(y; \\$\\theta$), \\quad y = \\tilde{I} - \\mathrm{NSN}(\\tilde{I}; \\$\\theta$),$$ in which the residual-block branch learns a noise map to subtract and the Convolutional Vision Transformer (CvT) branch learns structural detail to add back. The attention inside the CvT uses pixel unshuffle to cut spatial size, then pixel shuffle and projection afterward, so the transformer can attend globally at reduced computational cost.","core_discovery":"The central claim is that combining a residual-block noise suppression network with a convolutional vision transformer enhancement network yields a salt-and-pepper denoiser that outperforms classical decision-based, non-local, and learned baselines across nearly all tested cases. The paper asserts that the residual branch learns the noise map directly from the noisy image, and the transformer branch restores structural detail after subtraction, so no noise prior is required. On the reported PSNR table, the proposed method exceeds the best baseline by roughly 3 to 14 dB depending on image and noise level, with the single Pepper-30% exception.","pith_inferences":["If the reported BSD300 gains, such as 44.56 dB at 30 percent noise, are reproducible, they suggest the network is capturing dataset-level structure rather than only impulse statistics; that should be tested by cross-dataset generalization, which the paper does not report.","The same residual-plus-transformer fusion could be applied to random-valued impulse noise or mixed noise with minimal modification, since neither branch is tied to the specific salt-and-pepper model.","Because training used DIV2K and testing used classic images, a held-out modern dataset would clarify whether the advantage survives distribution shift.","The pixel-unshuffle attention could be compared directly against full-resolution attention to isolate how much of the gain comes from the transformer versus the residual branch."],"forward_implications":["RTF-Net needs no noise-level prior at test time, so it can be dropped into pipelines where the noise density is unknown.","The two-stage design separates noise removal from detail restoration, a pattern that could be reused for other degradation types with the same residual-plus-transformer structure.","At 70 percent noise the reported margin over the best baseline is largest, for example 32.85 dB versus 30.18 dB on Lena, suggesting the architecture is especially strong under very heavy corruption.","The one reported failure, Pepper at 30 percent noise, is a single-image, single-level case and does not overturn the overall pattern claimed by the paper."],"supporting_citations":[{"why":"Supplies the Convolutional Vision Transformer architecture used in the Structure Enhancement Network.","marker":"[6]"},{"why":"Motivates the residual identity mapping used inside the Noise Suppression Network blocks.","marker":"[8]"},{"why":"Provides the multi-head attention mechanism applied within each transformer block.","marker":"[4]"},{"why":"Baseline decision-based denoiser whose PSNR is compared against RTF-Net.","marker":"[10]"},{"why":"Baseline adaptive switching non-local filter whose PSNR is compared against RTF-Net.","marker":"[11]"},{"why":"Baseline patch-based impulse-Gaussian denoiser whose PSNR is compared against RTF-Net.","marker":"[12]"},{"why":"Baseline CNN denoisers NLSF and NLSF-CNN whose PSNR values are the main comparisons.","marker":"[13]"},{"why":"Baseline plain neural network denoiser NLSF-MLP whose PSNR is compared against RTF-Net.","marker":"[14]"}],"fun_headline_variants":["Transformer-ResNet fusion beats six denoisers on salt-and-pepper","RTF-Net wins nearly all PSNR tests for salt-and-pepper noise","Residual transformer hybrid dominates salt-and-pepper denoising","One exception: NLSF-CNN beats RTF-Net on Pepper at 30% noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported PSNR table is only meaningful if all compared methods were evaluated on the same noisy inputs under the same protocol; the paper does not state the noise generation or averaging details, so a mismatch in that protocol would invalidate the claimed superiority.","fun_headline_variants_meta":{"raw":{"variants":["Transformer-ResNet fusion beats six denoisers on salt-and-pepper","RTF-Net wins nearly all PSNR tests for salt-and-pepper noise","Residual transformer hybrid dominates salt-and-pepper denoising","One exception: NLSF-CNN beats RTF-Net on Pepper at 30% noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000402,"raw_usage":{"total_tokens":2131,"prompt_tokens":1012,"completion_tokens":1119,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1033}},"tokens_in":628,"tokens_out":1119,"duration_ms":9255,"temperature":1.0,"reasoning_tokens":1033,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:57:20.548378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Regenerate salt-and-pepper noise on the same test images with the paper's own noise model and rerun the six baseline methods and RTF-Net under identical random seeds; if the baseline PSNR values differ from Table II by more than rounding, or if any baseline matches or beats RTF-Net outside the reported Pepper-30% case on the regenerated noise, the central claim fails.","supporting_citations":[{"cited_title":"A new fast and efficient decision-based algorithm for removal of high-density impulse noises,","cited_arxiv_id":null,"evidence_quote":"Baseline decision-based denoiser whose PSNR is compared against RTF-Net."},{"cited_title":"Adaptive switching non-local filter for the restoration of salt and pepper impulse-corrupted digital images,","cited_arxiv_id":null,"evidence_quote":"Baseline adaptive switching non-local filter whose PSNR is compared against RTF-Net."},{"cited_title":"PARIGI: a Patch-based Approach to Remove Impulse-Gaussian Noise from Images,","cited_arxiv_id":null,"evidence_quote":"Baseline patch-based impulse-Gaussian denoiser whose PSNR is compared against RTF-Net."},{"cited_title":"A Convolutional Neural Networks Denoising Approach for Salt and Pepper Noise","cited_arxiv_id":"1807.08176","evidence_quote":"Baseline CNN denoisers NLSF and NLSF-CNN whose PSNR values are the main comparisons."},{"cited_title":"Image denoising: Can plain neural networks compete with bm3d?","cited_arxiv_id":null,"evidence_quote":"Baseline plain neural network denoiser NLSF-MLP whose PSNR is compared against RTF-Net."}],"review_version":1}