{"id":"75bb7edf-8f01-42b2-9128-b02ef745a966","arxiv_id":"1908.02648","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Linear depthwise convolution plus determinant-based filter attention (ALDNet) yields a lightweight SR network with modest accuracy gains, though the claimed advantages are not fully established.","lead":"ALDNet is a super-resolution network that removes the nonlinearity between depthwise and pointwise convolutions and adds a channel attention branch driven by the determinant of each depthwise filter. The paper reports small PSNR/SSIM gains over older baselines and over depthwise variants, but the evidence is weakened by unfair comparisons and missing code.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I compares ALDSR trained on DIV2K against baselines trained on smaller datasets, so the headline state-of-the-art claim is not established.","rationale":"I agree with the reader's weakest assumption. The cross-dataset protocol mismatch is the most load-bearing concern because it directly targets the paper's headline claim, not a peripheral design detail. The paper's own internal ablations (linear depthwise vs. depthwise) are performed under a consistent protocol and provide some support for the proposed building block, which is why the paper merits a conditional, not outright rejection. However, the determinant descriptor issue in Table IV is a secondary concern: the paper's own results show average/max pooling sometimes give better PSNR, so the specific claim about determinant is not consistently supported. Still, that would only weaken one of the contributions, not the empirical SOTA claim. The static, weight-derived nature of the attention branch is conceptually interesting but is not needed to decide the verdict; the protocol mismatch is sufficient to withhold acceptance until the authors provide fair comparisons. Therefore, the reader's CONDITIONAL verdict remains appropriate, pending retrained baselines and statistical evaluation.","tokens_in":11008,"tokens_out":10074,"duration_ms":111867,"concrete_test":"Retrain VDSR, LapSRN, MemNet, and IDN on the identical DIV2K training set (800 images) with the same patch size (48x48), augmentation (random flips/rotations), optimizer (Adam lr=1e-4 halved at epoch 200), and 300-epoch budget used for ALDSR, then evaluate on Set5/Set14/B100/Urban100. If ALDSR no longer achieves higher PSNR/SSIM than these retrained baselines on the majority of datasets, the state-of-the-art claim in Section V is not supported. Alternatively, train ALDSR on the same 291-image set used by VDSR; if the performance gap shrinks to <0.1 dB or reverses, the Table I advantage is explained by training-data size.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is \"state-of-the-art SR performance with low computational cost.\" The only direct SOTA comparison is Table I, where ALDSR (trained on the 800-image DIV2K training set) is compared against published PSNR/SSIM values for VDSR, LapSRN, MemNet, and IDN. These baselines were trained on different and generally much smaller datasets (e.g., 291 images for VDSR and LapSRN, with different augmentation and training schedules). SR performance is highly sensitive to training data quantity and protocol; a 0.1-0.3 dB advantage could be entirely attributable to the larger training set. The paper never retrains any baseline under its own protocol, nor does it report error bars or multiple seeds. Consequently, the \"state-of-the-art\" claim is unsupported. The internal ablations (Tables II and V) compare models under the same DIV2K protocol and are more trustworthy, but they only demonstrate small relative gains (e.g., 0.08 dB for LDW over DW on Urban100), and without statistical evaluation, even these could be noise. This is load-bearing because the conclusion's headline assertion depends on the fairness of Table I; if the baselines were retrained on DIV2K, ALDSR might not beat them, and the paper's main selling point collapses even if the depthwise mechanism has some merit.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ALDNet, a super-resolution network built on attention-aware linear depthwise convolution. The linear depthwise convolution removes the ReLU between the depthwise and pointwise convolutions, and an attention branch computes the determinant of each 3x3 depthwise filter and uses a gating mechanism to recalibrate feature maps. The authors construct a lightweight baseline, ALDSR, and also adapt the Residual Dense Network into ALD-RDN by replacing standard convolutions with the proposed ALD convolutions. Experiments are reported on Set5, Set14, BSD100, Urban100, and the DIV2K validation set, with internal ablations comparing depthwise, linear depthwise, and attention-aware variants under a common training protocol.","tokens_in":11240,"tokens_out":4241,"duration_ms":42450,"significance":"If the results hold, the work offers a lightweight SR architecture with a novel filter-descriptor-based attention mechanism. The internal ablation design is a strength: DW-RDN, LDW-RDN, and ALD-RDN are trained under the same protocol, and Table III provides useful parameter counts for the building blocks. The central claim, however, is not established because the headline comparison in Table I is not protocol-matched, and the determinant descriptor is not consistently the best on PSNR according to the paper's own Table IV. The paper would be significantly strengthened by retraining baselines under the same protocol and by reporting repeated-run statistics.","major_comments":[{"comment":"The state-of-the-art comparison in Table I is not protocol-matched: ALDSR is trained on the 800-image DIV2K training set, while VDSR, LapSRN, and MemNet are commonly trained on the much smaller 291-image dataset with different augmentation and training schedules. Since SR performance is highly sensitive to training data quantity, the reported PSNR gaps do not demonstrate that ALDSR surpasses these methods. No baseline is retrained under the authors' protocol, and no error bars or multiple-seed results are provided. To support the conclusion that ALDNet achieves state-of-the-art performance, the authors must either retrain the baselines under the same protocol or visibly restrict the state-of-the-art claim to the internally re-implemented comparisons.","section":"IV-B, Table I"},{"comment":"The conclusion states that the determinant descriptor shows superior performance to average and max descriptors, but this is contradicted by Table IV. On PSNR, average pooling or max pooling beats or ties the determinant on most configurations: for ALD-RDN, max pooling gives 32.23 dB on Set5 versus 32.18 dB for determinant, average pooling gives 28.66 dB on Set14 versus 28.62 dB, and the two descriptors tie on B100 and Urban100. For ALDSR, determinant is slightly worse on Urban100 and DIV2K validation than average pooling. The discussion in Section IV-B acknowledges this, so the concluding claim should be tempered and should distinguish PSNR from SSIM.","section":"V Conclusion and IV-B, Table IV"},{"comment":"The evidence for the proposed mechanisms is based on small PSNR differences without statistical evaluation. For example, LDW-RDN outperforms DW-RDN by 0.08 dB on Urban100, and ALD-RDN outperforms DW-RDN by 0.17 dB on Urban100, but no repeated runs or variance estimates are reported. These differences may be within run-to-run variability, especially for networks trained with Adam and random patch sampling. Because these internal ablations are the main support for the claim that linearity and attention-aware depthwise convolution help SR, the authors should report the mean and standard deviation over at least two or three random seeds, or otherwise demonstrate that the improvements are reproducible.","section":"IV-A, IV-B, Tables II and V"},{"comment":"The reduction ratio r in the attention branch is set to 16 with no sensitivity analysis. Since the authors emphasize the low computational cost of the proposed block, the effect of varying r on both performance and parameter count should be reported, or at least a brief justification should be given for choosing the SENet-inspired value.","section":"II-B, Eq. (3)"}],"minor_comments":[{"comment":"There are several typos in the introduction and elsewhere, including 'depthwsie', 'mentined ealier', 'dose not', and 'convoltuion'. These should be corrected.","section":"I"},{"comment":"The manuscript uses 'ALDSR' and 'ALDNet' interchangeably. The relationship between the general ALDNet and the specific ALDSR baseline should be stated explicitly once in Section III.","section":"III-A"},{"comment":"The text says the pointwise convolution layer is omitted from Figure 6 for simplicity, but this makes the block diagram misleading because the ALD convolution is defined as depthwise plus pointwise. The figure should include the pointwise layer or clearly annotate where it is inserted.","section":"Fig. 6"},{"comment":"The discussion of why max pooling helps PSNR and determinant helps SSIM is speculative; a short quantitative connection, such as correlation of descriptor values with reconstruction error, would make the interpretation more credible.","section":"IV-B, Discussion"},{"comment":"Reference [4] (EDSR) is cited for the L1 loss and ESPCN but the discussion of batch normalization removal also relies on it; the citation is appropriate but the sentence structure could be clearer.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the internal ablations are a positive feature, but the headline state-of-the-art claim is not supported by the protocol-unmatched comparison in Table I, and the determinant descriptor's advantage over simpler pooling schemes is not clear on PSNR. I would be willing to consider a revised version that retrains baselines under the same protocol, adds repeated-run statistics, and tempers the claims about the determinant descriptor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea here—dropping the ReLU between depthwise and pointwise convolutions and using a filter-interdependency attention branch—is sensible and not something I've seen before. But the state-of-the-art claim in the conclusion rests on Table I, which compares ALDSR trained on DIV2K (800 images) against published numbers for VDSR, LapSRN, MemNet, and IDN that were typically trained on 291 images. That comparison can't carry the weight. The differences are around 0.1–0.3 dB, exactly the size of typical training-data effects.\n\nWhat the paper does well: the internal ablations are fairly designed. Replacing standard convs in RDN with depthwise separable (DW-RDN), then linear depthwise (LDW-RDN), then adding the attention branch (ALD-RDN) under the same protocol shows small but consistent gains, with the biggest on Urban100 (26.06 to 26.14 to 26.23). That suggests the mechanism is doing something real. The parameter counts are also reported clearly, and Table III shows ALDB is much lighter than RDB.\n\nSoft spots: the determinant descriptor doesn't actually win on PSNR in the paper's own Table IV—average and max pooling are often better. The authors acknowledge this and fall back on SSIM, which is fine, but it undercuts the 'determinant captures filter shape' story. Also, no error bars, no multiple seeds, so even the internal gains could be noise; 0.08 dB on Urban100 is small. The paper doesn't release code, and the exact architecture of ALDSR (number of blocks, channels) is only partially specified.\n\nOverall, this is an empirical architecture paper with a plausible idea and insufficient evidence for its headline. The stress-test concern lands: the SOTA claim is not established. The right fix is retraining baselines under the same protocol, reporting variance, and releasing code. The paper deserves a serious referee—not a desk reject—because the linear depthwise idea is simple and potentially useful for mobile SR. But I wouldn't cite the state-of-the-art claim as it stands. Readers working on efficient SR or lightweight networks should look at the internal ablations, not Table I.","headline":"Plausible lightweight SR block, but the state-of-the-art claim rests on an unfair comparison and the paper's own ablations don't single out the determinant descriptor.","tokens_in":11782,"tokens_out":1992,"would_cite":false,"duration_ms":22821,"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":"Removing the ReLU between depthwise and pointwise stages and scoring filters by determinant lets a lightweight network match heavier super-resolution models.","keywords":["single image super-resolution","depthwise separable convolution","linear depthwise convolution","channel attention","determinant","filter interdependency","lightweight network"],"falsifier":"Retrain VDSR, LapSRN, MemNet, and IDN on the same DIV2K 800 training images with the same 4× bicubic protocol, loss, and training scheme used for ALDSR, and compare PSNR/SSIM on the four benchmark sets; if ALDSR no longer matches or exceeds them, the paper's headline superiority claim is not established. The same experiment could also test linearity by inserting a ReLU between the depthwise and pointwise stages of ALD-RDN and checking whether Urban100 performance drops.","tokens_in":10790,"feed_emoji":"🖼️","tokens_out":7733,"duration_ms":79317,"temperature":0.7,"pith_summary":"Single-image super-resolution aims to restore a high-resolution image from a low-resolution input, and deep CNN models do this well but at growing computational cost; this paper attacks both problems at once. It proposes ALDNet, in which each depthwise convolution is made \"linear\" by removing the ReLU between the depthwise and pointwise stages, and an attention branch scores each depthwise filter by its determinant and rescales the channel features. The authors build this block into a lightweight baseline, ALDSR, and report that it matches or exceeds published results of VDSR, LapSRN, MemNet, and IDN on mainstream benchmarks at 4× upscaling with about 731k parameters. They also drop the block into RDN to form ALD-RDN and show it outperforms the corresponding depthwise-separable RDN on every benchmark. If the claims hold, the practical payoff is that a compact SR network can preserve fine details, such as repetitive lattice patterns, without paying the parameter cost of heavy models.","feed_headline":"A 731k-parameter SR network rivals much heavier models","feed_subtitle":"Dropping one ReLU and adding determinant-based filter attention preserves edges and repetitive patterns.","key_machinery":"The load-bearing unit is the attention-aware linear depthwise convolution (ALD convolution). It combines a depthwise $3\\times3$ convolution directly followed by a $1\\times1$ pointwise convolution with no activation in between, plus a branch that computes $\\det W_3^c$ for each filter, passes the $C$-dimensional descriptor vector through FC-ReLU-FC-Sigmoid with reduction ratio $r=16$, and rescales the feature map as $f_c'=s_c f_c$, followed by residual scaling $D=(1+s_c)f_c$. The determinant is the paper's chosen descriptor of filter shape: Sarrus' rule gives the $3\\times3$ determinant from products of diagonal entries, so the attention is driven by interdependency among a filter's weights rather than by spatial pooling of feature maps. A local residual connection every two ALD convolutions and a bottleneck $1\\times1$ fusion realize the block's memory mechanism at low parameter count.","core_discovery":"The paper's central claim is that depthwise separable convolution—originally designed for image classification—can be made effective for single-image super-resolution through two modifications. The first is removing the nonlinearity between the depthwise spatial filter and the 1×1 pointwise filter, because at that intermediate stage the feature map reflects only spatial information and a ReLU there discards reconstruction details. The second is an attention branch that exploits the fact that each depthwise filter is a small square matrix: the determinant of each $3\\times3$ filter is computed by Sarrus' rule, reduced by a gating network, and used to recalibrate the channel features, with a residual connection so repeated attention does not degrade performance. The paper reports that the resulting ALDSR matches or beats the published numbers of several heavier SR networks, and that when inserted into RDN the proposed block (ALD-RDN) beats a depthwise-separable RDN on all tested datasets, with the largest gain on Urban100, which contains repetitive structures.","pith_inferences":["Because the block is modular, the same linear-depthwise-plus-determinant design could plausibly transfer to other image restoration tasks such as denoising, deblurring, or JPEG artifact removal, where preserving fine details is also critical; the paper only evaluates SR.","The determinant descriptor is one member of a family of matrix descriptors; a natural next experiment is comparing trace, spectral norm, or singular values in the attention branch, since the paper's ablation compares only determinant against average and max pooling.","A practical consequence the authors leave implicit is that replacing standard convolutions with ALD convolutions should cut FLOPs roughly by the depthwise factor, making real-time or on-device 4× SR feasible; the parameter counts in Table III support this direction."],"forward_implications":["Linear depthwise convolution alone improves super-resolution: LDW-RDN beats DW-RDN on Urban100 by 0.08 dB with the same parameter count.","Adding the attention branch improves it further: ALD-RDN beats DW-RDN on all five test sets, with a 0.17 dB gain on Urban100.","The gains are cheap: one ALDB has 41,472 parameters versus 1,363,968 for an RDB, so the proposed block is suited to compact SR networks.","Determinant-based attention yields the best SSIM among the tested descriptors on every benchmark except Set5 for ALD-RDN, indicating better structural reconstruction.","The clearest effects appear on repeated patterns and textures, as shown on Urban100 images where ALDSR and ALD-RDN reconstruct lattices more accurately than the compared methods."],"supporting_citations":[{"why":"Defines depthwise separable convolution, the operation the paper modifies into linear depthwise convolution.","marker":"[6]"},{"why":"Supplies the squeeze-and-excitation gating design that the attention branch adapts to filter descriptors.","marker":"[8]"},{"why":"Motivates the residual scaling $D=(1+s_c)f_c$ that avoids degrading performance from stacked attention masks.","marker":"[12]"},{"why":"Provides the precedent for removing batch normalization and using L1 loss in SR networks, which the paper adopts.","marker":"[4]"},{"why":"Supplies the Residual Dense Network architecture and contiguous memory mechanism that ALD-RDN replaces with ALD layers.","marker":"[5]"},{"why":"Provides global residual learning for the ALDSR structure and is one of the published baselines in the headline comparison.","marker":"[3]"}],"fun_headline_variants":["Drop ReLU, add determinant attention: sharper SR","Linear depthwise conv preserves SR details","Attention-aware depthwise net rivals heavy SR models","Determinant-based attention boosts depthwise SR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison in Table I assumes the published PSNR/SSIM values of VDSR, LapSRN, MemNet, and IDN—trained on different, generally smaller datasets—can be directly compared with ALDSR's scores, which were produced on the 800-image DIV2K training set with a different protocol.","fun_headline_variants_meta":{"raw":{"variants":["Drop ReLU, add determinant attention: sharper SR","Linear depthwise conv preserves SR details","Attention-aware depthwise net rivals heavy SR models","Determinant-based attention boosts depthwise SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1870,"prompt_tokens":889,"completion_tokens":981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":923}},"tokens_in":505,"tokens_out":981,"duration_ms":10219,"temperature":1.0,"reasoning_tokens":923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:38:23.883903+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain VDSR, LapSRN, MemNet, and IDN on the same DIV2K 800 training images with the same 4× bicubic protocol, loss, and training scheme used for ALDSR, and compare PSNR/SSIM on the four benchmark sets; if ALDSR no longer matches or exceeds them, the paper's headline superiority claim is not established. The same experiment could also test linearity by inserting a ReLU between the depthwise and pointwise stages of ALD-RDN and checking whether Urban100 performance drops.","supporting_citations":[{"cited_title":"Accurate image super-resolution using very deep convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Provides global residual learning for the ALDSR structure and is one of the published baselines in the headline comparison."}],"review_version":1}