{"id":"2c574e87-91ea-4440-bf45-f4306f46a1c4","arxiv_id":"2412.11008","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A convolutional image-restoration network combining star-style element-wise multiplication with dilated strip attention reports state-of-the-art PSNR on dehazing, deblurring, and desnowing benchmarks.","lead":"This paper presents a new convolutional network, CCNet, for cleaning up hazy, blurry, or snowy photos. On three standard benchmarks the authors report top results using far fewer parameters than transformer-based rivals.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unverified H-DSA equation and unspecified strip/kernel sizes leave the claimed extremely large LDIM receptive field unsupported.","rationale":"The reader's weakest_assumption is the same central gap: LDIM's receptive-field claim is load-bearing yet underspecified. I agree with that analysis. However, I would not move the verdict further because the reported CNN results, if reproduced, would still be a solid contribution; the issue is verification, not demonstrated internal inconsistency. The most decisive check is code release with the exact LDIM configuration and a computed receptive-field comparison against the baselines (SANet, OKNet). If the receptive field turns out to be small, the paper's explanation of why LDIM helps would be wrong, but the empirical result could still stand; hence CONDITIONAL remains appropriate.","tokens_in":15310,"tokens_out":1113,"duration_ms":9888,"concrete_test":"Request the authors' released code or configuration and compute the exact receptive field of one H-LDSI for the reported K and per-group dilation schedule at the largest used scale (4C channels, H/4 x W/4 spatial size). If the reachable context is below the image diagonal and no larger than that of SANet/OKNet baselines, Eq. (6) and the 'extremely large' square receptive-field claim should be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"CCNet's headline advantage over transformers is attributed to LDIM's extremely large square receptive field, but the paper never specifies the strip size K or how the dilated rates d=1,...,D in Eq. (6) are split across feature groups. Eq. (6) as written applies one dilation d per output channel c, which does not match group-wise multi-scale dilation described in the text and in Fig. 4, so the receptive field cannot be computed from the manuscript alone. The argument also conflates the covered span with the effective receptive field: H-SA followed by H-DSA with dr=(k+1)/2 gives each output pixel a formal dependence on a span of about 2K pixels, not the global or extremely large context claimed; the claim that operated pixels contain the contribution of any input pixel of the H-LDSI is true only within that finite span. Because ablation results (Table 4) show LDIM adds +2.36 dB PSNR over the baseline, the empirical contribution is plausible, but the architectural mechanism credited for the gain is not verifiable as written and no receptive-field size is reported anywhere. The absent RSAM connection is a further unverified claim if the reported network uses it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CCNet, a U-shaped convolutional network for image restoration, with two main components. The efficient residual star module (ERSM) uses element-wise multiplication ('star operation') together with a large depthwise convolution branch, and the large dynamic integration module (LDIM) combines horizontal and vertical strip attention with dilated strip attention to aggregate context. The authors report state-of-the-art PSNR/SSIM on SOTS dehazing, GoPro/HIDE motion deblurring, and CSD desnowing using about 4.3M parameters, and they provide ablations on SOTS-Indoor attributing the gains to ERSM and LDIM.","tokens_in":15530,"tokens_out":4926,"duration_ms":43055,"significance":"If the reported benchmark numbers are reproducible, the efficiency claim is practically significant: a 4.3M-parameter CNN outperforms several transformer-based methods, including Restormer on GoPro and DehazeFormer-L on SOTS-Indoor, while using substantially fewer parameters. The D-RSM versus ERSM comparison in Table 4 is a fair controlled test with matched parameter counts, and the multi-task evaluation across dehazing, deblurring, and desnowing is a strength. However, the paper's central architectural mechanism for LDIM is under-specified: Eq. (6) cannot be used to compute the claimed 'extremely large' receptive field, no values for K or D are given, and the proposed RSAM module in Section 3.4 is never evaluated. These issues make the significance conditional on additional clarification and verification.","major_comments":[{"comment":"Eq. (6) is ambiguous and prevents the reader from verifying the LDIM receptive field. The text describes multi-scale dilation applied across feature groups, but the equation sums only over k and writes d = 1,...,D without showing how d is assigned to channels or groups. As written, each output channel appears to use a single dilation rate, which does not match the group-wise description in the text and Fig. 4. The strip size K and dilation counts D are never specified anywhere in the paper, and no receptive-field size is computed. Please give the group split, report K and D, and provide an explicit formula or numerical value for the receptive field of one H-LDSI/V-LDSI and of the full LDIM.","section":"Section 3.3, Eq. (6)"},{"comment":"The claim that 'its operated pixels contain the contribution of any input pixel of the H-LDSI' and that LDIM possesses an 'extremely large square receptive field' is an overstatement. With strip size K and dilation dr = (K+1)/2, H-SA integrates over K neighboring pixels and H-DSA then samples those K positions, so each output pixel formally depends on a span of about K * dr = K(K+1)/2 pixels, not on any input pixel of the whole feature map. Unless K is large enough to cover the image, this is not global context. The paper should state the effective receptive field explicitly and temper the 'extremely large' claim accordingly, since the +2.36 dB gain of LDIM in Table 4 is attributed to this mechanism.","section":"Section 3.3, paragraphs after Eq. (8)"},{"comment":"The residual star attention module (RSAM) introduced in Section 3.4 is claimed to 'significantly boost model performance,' but it does not appear in the network architecture description of Section 3.1, in Fig. 2, or in any experiment or ablation in Table 4. The module is therefore unsupported by evidence. Either integrate RSAM into the actual architecture and ablate it, or remove Section 3.4 and the associated claim. As it stands, a reader cannot tell whether the reported CCNet includes RSAM at all.","section":"Section 3.4 and Table 4"},{"comment":"The ablation evidence for the star operation's benefit is only conditional. The model with D-RSM alone (star unit without the context branch) drops to 37.86 dB, below the 38.40 dB baseline, while ERSM raises performance to 40.23 dB. This shows that the improvement comes from the added large depthwise convolutional context branch, not from the star operation per se. The abstract and Section 3.2 credit 'context-aware star operation' with high-dimensional, non-linear feature mapping; the paper should either temper this claim or add a control that isolates the star operation's contribution, for example ERSM without the star operation or with the star operation placed differently.","section":"Section 4.3, Table 4"}],"minor_comments":[{"comment":"The HIDE column headers read 'PSNR↓ SSIM↓' but these metrics should be the higher-the-better arrows (↑), as in the GoPro columns.","section":"Table 2"},{"comment":"The text cites the CSD dataset as [47], but reference [47] is the PyTorch software paper; the CSD dataset is defined by reference [46]. Please correct the citation.","section":"Section 4.2, Image Desnowing"},{"comment":"The symbol k is used both for the summation index and for the strip size in dr = (k+1)/2, while the strip size is otherwise denoted K. Please use consistent notation, e.g., K for the strip size and k for the summation index.","section":"Section 3.3, notation"},{"comment":"The loss equation defines s as a normalization denominator but then says 'S denotes the total elements'; the lowercase s and uppercase S are inconsistent. Use one symbol throughout.","section":"Section 3.5, Eq. (11)"},{"comment":"There are several typographical errors, including 'incluing' (Section 3.2), 'genarate' (Section 3.1), and 'ultilize' (Section 4.3). A copyedit pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The reported gains are large enough that independent verification matters; the underspecified LDIM equations and the orphaned RSAM section are the main technical blockers. I would strongly encourage the authors to release code or at least provide detailed hyperparameters (K, D, group splits) and a receptive-field derivation. The paper fits the journal's scope, but the current version does not let a reader confirm the central mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CCNet is a plausible compact CNN for restoration, and the reported numbers are strong enough to justify a careful look—but the core LDIM receptive-field claim is under-specified as written, RSAM is never used or ablated, and one headline comparison (HIDE) is slightly overstated.\n\nWhat is actually new is modest but real. ERSM takes the star operation from StarNet and adds a depth-wise convolution in one branch, turning it into a context-aware unit for low-level tasks. LDIM stacks horizontal and vertical strip attention in two stages, with a dilated second stage, which is a reasonable extension of SANet. The ablation isolates the pieces nicely: D-RSM (star block without the context branch) drops below baseline, ERSM adds about 1.8 dB, LDIM adds about 2.4 dB, and the full model reaches 41.25 dB on SOTS-Indoor. Same-parameter comparison between D-RSM and ERSM is a clean way to show the context branch matters. The efficiency story—4.26M parameters, 43.5G MACs, beating much larger transformers—is worth taking seriously if the numbers reproduce.\n\nThe soft spots are in the mechanism claimed for the main gains. The LDIM section never specifies the strip size K or how the dilation rates d = 1,...,D in Eq. (6) are split across feature groups. As written, Eq. (6) applies one dilation per output channel, which does not match the group-wise multi-scale description in the text and Fig. 4. Without K and D, the \"extremely large square receptive field\" cannot be computed, and the formal span of H-SA followed by H-DSA is finite (about 2K pixels), not global. That does not kill the empirical result—LDIM clearly helps—but it means the paper's central explanation for the improvement is not verifiable from the manuscript. RSAM is described in Section 3.4, appears in no architecture figure, is not included in the main model, and has no ablation; it should be removed or justified. Also, on HIDE the paper is 0.13 dB behind Restormer in PSNR, so the abstract's blanket \"superior performance on image motion deblurring\" overstates the table; the higher SSIM is the defensible claim.\n\nThis paper is for researchers building efficient CNN-based restoration networks and comparing against transformer baselines. It deserves a serious referee, but the review should be conditional on the authors releasing code and specifying K, D, the group split, and all kernel sizes. Without code, the strong benchmark numbers are trust-based, and the receptive-field claim is uncheckable.","headline":"A compact CNN with strong reported numbers on dehazing/deblurring/desnowing, but the load-bearing LDIM receptive-field claim is under-specified and the RSAM is vestigial; deserves review with code.","tokens_in":16088,"tokens_out":2257,"would_cite":false,"duration_ms":22478,"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 4.26M-parameter convolutional network outperforms transformer-based models on three image restoration tasks.","keywords":["image restoration","dehazing","motion deblurring","desnowing","star operation","strip attention","receptive field","convolutional network"],"falsifier":"Inspect the implementation to read off $K$ and the dilation rates; if the effective receptive field of LDIM covers only a small fraction of the $256\\times256$ training patches, or if removing LDIM changes PSNR by far less than the reported 2.36 dB on the SOTS-Indoor ablation, then the large-context explanation is not what carries the result.","tokens_in":15097,"feed_emoji":"🖼️","tokens_out":9534,"duration_ms":71130,"temperature":0.7,"pith_summary":"This paper argues that a purely convolutional image-restoration network can match or beat transformer-based models by combining two ideas: a residual 'star' module that multiplies two learned feature maps element-wise, and a large dynamic integration module that stacks horizontal and vertical strip attention with dilated strip attention to gather context from a very wide area. If the reported results hold, the network achieves 41.25 dB PSNR on SOTS-Indoor dehazing, 40.29 dB on SOTS-Outdoor, 33.31 dB on GoPro motion deblurring, and 38.29 dB on CSD desnowing while using roughly 4.26 million parameters. The paper's point is that long-range context, usually the selling point of transformers, can be obtained dynamically with cheap convolutional operations, making high-performance restoration feasible at low model complexity.","feed_headline":"A 4.26M-parameter CNN beats transformer models on image restoration","feed_subtitle":"Star-operation feature mapping plus a wide strip-attention field drives gains on dehazing, deblurring, and desnowing.","key_machinery":"The paper's two load-bearing mechanisms are the efficient residual star module (ERSM) and the large dynamic integration module (LDIM). The 'star operation' is element-wise multiplication of two feature vectors; ERSM makes it context-aware by feeding one branch through a large depth-wise convolution so that the multiplied features carry local contextual information before being projected into a high-dimensional nonlinear space. LDIM builds a wide square receptive field by composing horizontal strip attention (H-SA) of strip size $K$ with horizontal dilated strip attention (H-DSA) whose dilation rate is $d_r=(k+1)/2$, then repeating the pair vertically; because the first strip attention already mixes pixels along the strip, the subsequent dilated attention lets the module reach pixels far beyond a single kernel without quadratic cost. All attention weights are generated dynamically from the input by convolutions, and the module uses multi-scale receptive fields across feature groups to handle blurs of different sizes.","core_discovery":"On its own terms, the paper claims that the efficient residual star module (ERSM) and the large dynamic integration module (LDIM), placed in a six-scale U-shaped encoder-decoder, yield a context-aware convolutional network that outperforms prior state-of-the-art restoration methods on image dehazing, motion deblurring, and desnowing. ERSM replaces a plain residual block with a context-aware star unit: one branch is a $1\\times1$ convolution, the other is a $1\\times1$ convolution followed by a large depth-wise convolution and GELU, and their element-wise product is refined by a $3\\times3$ convolution with a skip connection. LDIM applies horizontal strip attention, then horizontal dilated strip attention, then the vertical equivalents, with all attention weights learned dynamically from the input features; the paper argues that the strip-then-dilate composition gives every operated pixel access to any input pixel over an extremely large square region. Ablations on SOTS-Indoor attribute 2.36 dB of the 2.85 dB total gain over the baseline to LDIM and show that ERSM adds another 0.49 dB, with the full model reaching 41.25 dB.","pith_inferences":["If the receptive-field claim is verified, the same horizontal-then-vertical dilated strip composition could be applied to other long-range tasks such as single-image deraining or super-resolution, where the degradation is also spatially varying.","A direct measurement of the effective receptive field, for example by gradient propagation or input perturbation, would show whether the benefit comes from true long-range integration or from the dynamic weighting alone; the paper does not provide that measurement.","The ablation suggests LDIM contributes most of the gain, so a natural extension would be to test LDIM alone in a simple ResNet-style backbone, without the star module, to isolate how much of the improvement is due to context integration rather than the star operation."],"forward_implications":["A convolutional restoration network with roughly 4.26 million parameters can outperform transformer-based models on the dehazing and motion-deblurring benchmarks reported here, which would lower the compute barrier for practical restoration.","The strip-then-dilated attention composition offers a parameter-light way to obtain long-range context without quadratic self-attention, so similar modules could be dropped into other low-level vision networks.","Because LDIM's weights are predicted from the input and applied at multiple scales, the network can adapt its context integration to blurred regions of different sizes, matching the variability found in real motion blur.","The dual-domain $L_1$ loss (spatial plus frequency) combined with multi-input/multi-output supervision makes training stable enough that one U-shaped network handles three degradation types with task-specific block counts."],"supporting_citations":[{"why":"Supplies the star-operation idea: element-wise multiplication maps features into high-dimensional nonlinear spaces; ERSM adapts it for restoration by adding a context branch.","marker":"[42]"},{"why":"Introduces the strip attention module that LDIM extends; the paper argues SAM's receptive field is limited and the strip-then-dilated design fixes that.","marker":"[19]"},{"why":"Supplies the transformer baseline for motion deblurring that CCNet reports beating on GoPro.","marker":"[1]"},{"why":"Supplies a recent convolutional baseline on dehazing and deblurring that CCNet reports outperforming.","marker":"[11]"},{"why":"Supplies a recent attention-based convolutional baseline that CCNet reports outperforming on dehazing and desnowing.","marker":"[20]"},{"why":"Supplies the transformer dehazing baseline that CCNet reports beating by 5.11 dB on SOTS-Outdoor.","marker":"[22]"},{"why":"Supplies the transformer desnowing baseline that CCNet reports beating by 4.54 dB on CSD.","marker":"[29]"},{"why":"Provides the RESIDE/SOTS dehazing dataset and evaluation protocol used for training and testing.","marker":"[21]"},{"why":"Provides the GoPro dynamic scene deblurring dataset used for training and testing motion deblurring.","marker":"[44]"},{"why":"Provides the CSD desnowing dataset used for the desnowing experiments.","marker":"[46]"}],"fun_headline_variants":["Tiny CNN with star ops rivals transformers in image restoration","Context-aware CNN beats transformers on dehazing, deblurring, desnowing","Star operation boosts small CNN to top image restoration results","4.26M-param CNN with star ops outshines transformers in restoration","Dynamic strip attention gives small CNN transformer-level restoration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole 'extremely large receptive field' argument depends on the strip size $K$ and the per-group dilation rates being large enough that the strip-then-dilated composition actually reaches across nearly the whole feature map; the paper never states these values or computes the resulting receptive field.","fun_headline_variants_meta":{"raw":{"variants":["Tiny CNN with star ops rivals transformers in image restoration","Context-aware CNN beats transformers on dehazing, deblurring, desnowing","Star operation boosts small CNN to top image restoration results","4.26M-param CNN with star ops outshines transformers in restoration","Dynamic strip attention gives small CNN transformer-level restoration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000775,"raw_usage":{"total_tokens":3487,"prompt_tokens":1062,"completion_tokens":2425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":2336}},"tokens_in":678,"tokens_out":2425,"duration_ms":15865,"temperature":1.0,"reasoning_tokens":2336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:23:50.798679+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the implementation to read off $K$ and the dilation rates; if the effective receptive field of LDIM covers only a small fraction of the $256\\times256$ training patches, or if removing LDIM changes PSNR by far less than the reported 2.36 dB on the SOTS-Indoor ablation, then the large-context explanation is not what carries the result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the star-operation idea: element-wise multiplication maps features into high-dimensional nonlinear spaces; ERSM adapts it for restoration by adding a context branch."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the strip attention module that LDIM extends; the paper argues SAM's receptive field is limited and the strip-then-dilated design fixes that."},{"cited_title":"W., Arora, A., Khan, S., Hayat, M., Khan, F","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer baseline for motion deblurring that CCNet reports beating on GoPro."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies a recent convolutional baseline on dehazing and deblurring that CCNet reports outperforming."},{"cited_title":"(2024, March)","cited_arxiv_id":null,"evidence_quote":"Supplies a recent attention-based convolutional baseline that CCNet reports outperforming on dehazing and desnowing."},{"cited_title":"L., Yan, Q., Anwar, S., Cong, R., Ren, W., & Li, C","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer dehazing baseline that CCNet reports beating by 5.11 dB on SOTS-Outdoor."},{"cited_title":"(2023, June)","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer desnowing baseline that CCNet reports beating by 4.54 dB on CSD."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RESIDE/SOTS dehazing dataset and evaluation protocol used for training and testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GoPro dynamic scene deblurring dataset used for training and testing motion deblurring."},{"cited_title":"T., Fang, H","cited_arxiv_id":null,"evidence_quote":"Provides the CSD desnowing dataset used for the desnowing experiments."}],"review_version":1}