{"id":"c1490e94-212e-4fa4-89a0-dd0a1fec800a","arxiv_id":"2506.18364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DCT-based low-frequency fusion into the input of a few-shot classifier gives roughly 1 to 3 point gains, but the paper's headline comparison uses a stronger backbone.","lead":"Few-shot image classifiers must learn from one or five labeled examples per category. This paper adds a preprocessing step that keeps the low-frequency part of an image and adds it back to the original before training, reporting small accuracy gains on three datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) and the text define two different low-frequency masks (r vs sqrt(r)); the reported gains cannot be attributed to the method until this ambiguity is resolved.","rationale":"The reader's CONDITIONAL verdict is appropriate, and the reader already noted the Eq. (3)/text mask inconsistency in their rationale. My pass sharpens this into the single most load-bearing concern: the central ablation in Table 3 is attached to an underspecified preprocessing operator, and the two possible readings of the mask produce substantially different inputs. This is not a manufactured objection; it is an internal inconsistency in the manuscript. The reader's weakest_assumption focused on sensitivity to r and the fusion operation, which is related but not identical; hence partial agreement. No additional evidence in the paper resolves the ambiguity, and no code or data are provided. The paper's other weaknesses—the backbone confound in Table 2 and the unspecified ResNet12+ViT combination—reinforce the need for a conditional verdict, but they do not independently change the recommendation. If the authors resolve the mask definition and show the ablation is robust to the two interpretations, the claim could be accepted; as written, the empirical result is not fully interpretable.","tokens_in":11319,"tokens_out":5239,"duration_ms":57282,"concrete_test":"Implement the preprocessing twice, exactly following Eq. (3) with r=0.15 (side length floor(h*sqrt(r))×floor(w*sqrt(r))) and following the prose (side length floor(hr)×floor(wr)), using the same ResNet12+ViT backbone and training schedule as in Table 3. Re-run the 5-way 1-shot and 5-way 5-shot ablations on CUB-200-2011 and Stanford Dogs. If both masks reproduce the reported gains, the ambiguity is not load-bearing; if one mask changes the gaps by more than about 1 point or flips their sign, the paper's central claim is tied to an unspecified implementation detail and needs correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that fusing low-frequency DCT information with the original image improves few-shot classification—rests entirely on the Table 3 ablation. However, the exact preprocessing is not defined unambiguously. Eq. (3) sets L_{i,j}=D_{i,j} only when i<floor(h*sqrt(r)) and j<floor(w*sqrt(r)), retaining a square block whose side is sqrt(r)≈0.387 of the image size and whose area is r=15% of all coefficients. The sentence immediately below the equation says the top-left floor(hr)×floor(wr) block is kept, which for r=0.15 is only 2.25% of the coefficients—a factor of 6.7 smaller in area and 2.58 smaller per side. These are very different images: one is mild low-pass filtering, the other is aggressive smoothing that removes almost all high-frequency content. The reported gains (1.1–3.3 points in Table 3) are attached to an unspecified r and mask shape. If the authors used the square-root definition, the claim is about a modest low-frequency emphasis; if they used the direct-r definition, the improvements may be due to strong regularization rather than to the claimed frequency–spatial fusion. No code or data are released, so the ambiguity cannot be resolved from the document. Because every downstream conclusion (including the comparison in Table 2) depends on this preprocessing, the central result is not reproducible or interpretable as stated. This is a correctness and reproducibility risk, not a disagreement with the general idea that low-frequency information can help; the specific implementation must be pinned down.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-preprocessing method for few-shot image classification: apply a 2D DCT to each input image, zero out all coefficients above a low-frequency cutoff, reconstruct via IDCT, and add the reconstruction to the original image. The authors call the resulting network SFIFNet, using ResNet12 and ViT as a joint backbone. They evaluate on CUB-200-2011, Stanford Dogs, and a custom 30-class animal dataset under 5-way 1-shot and 5-way 5-shot tasks, report improved accuracy over FRN, BSNet, DeepEMD, and DeepBDC (Table 2), and provide a same-backbone ablation (Table 3) showing gains of roughly 1.1 to 3.3 points.","tokens_in":11645,"tokens_out":4231,"duration_ms":42904,"significance":"If the same-backbone gains in Table 3 are robust, the method offers a simple, computationally cheap input-level augmentation that could be dropped into many existing few-shot classifiers, and the idea of fusing frequency-domain structure with spatial detail is plausible for fine-grained recognition. The paper does not ship code or data, and because of the unresolved definitional ambiguity in Eq. (3) and the missing statistical analysis, the claimed gains cannot currently be attributed to the described preprocessing. The contribution is therefore potentially useful but not yet verifiable.","major_comments":[{"comment":"Eq. (3) defines the low-frequency mask as L_{i,j} = D_{i,j} when i < floor(h*sqrt(r)) and j < floor(w*sqrt(r)), which retains a square block of side approximately sqrt(r) of the image dimension, i.e. area r = 15% of all coefficients. The sentence immediately below Eq. (3) says the retained block is the top-left floor(hr) by floor(wr) region, which with r = 0.15 keeps only r^2 approximately 2.25% of coefficients. These two masks produce very different images, and every result in Tables 2 and 3 depends on which one is used. The paper must unambiguously state the mask shape, provide a sensitivity analysis over r, and ideally release the preprocessing code.","section":"Mathematical components, Eq. (3)"},{"comment":"Table 3 reports a single set of runs and uses the standard deviation over 1000 randomly generated tasks, not the variability across training runs. The reported gains of 1.1 to 3.3 percentage points are not accompanied by any significance test, multiple training seeds, or paired comparison, so they may be within run-to-run variability. The central claim needs at least five independent training runs with mean plus/minus standard deviation and a paired test, for example across the same sampled tasks.","section":"Ablation Studies, Table 3"},{"comment":"Table 2 compares SFIFNet (ResNet12 plus ViT) against FRN, DeepEMD, and DeepBDC with ResNet12 and against BSNet with CNN, while the frameworks also differ (PyTorch 1.7.1 for SFIFNet and 2.2.0 for the others). These backbone and framework differences confound the comparison, so the large accuracy gaps cannot be attributed to the proposed frequency preprocessing. The only controlled evidence is Table 3; Table 2 should be explicitly framed as an overall system comparison, or all methods should be evaluated with the same backbone and training pipeline.","section":"Performance comparison, Table 2"},{"comment":"The manuscript states that fused images are fed into 'a backbone composed of ResNet12 and ViT' and Figure 1 shows the pipeline, but it does not describe how ResNet12 and ViT are combined, which layers are used, how the two networks interact, or whether the same preprocessing is applied to support and query images. The fusion operation is also described only as 'directly add' with no specification of the fusion weight. Without these details the method is not reproducible and the experiments cannot be independently verified.","section":"Materials and Methods / Results"}],"minor_comments":[{"comment":"The manuscript contains numerous grammatical errors, e.g. 'which would effect the classification performance' should be 'affect', and the opening sentence is a fragment. A thorough language edit is needed.","section":"Abstract and Introduction"},{"comment":"The text says 'N and N are the width and height of the images respectively'; this should be 'N and M are the width and height'. The same inconsistency appears in the surrounding DCT/IDCT discussion.","section":"Eq. (1) and Eq. (4)"},{"comment":"Table 1 has column headers 'Ntrain', 'Nval', and 'Ntest' without definitions; these should be clarified, and the captions of Tables 1, 2 and 3 should be formatted consistently.","section":"Table 1"},{"comment":"The data availability statement says 'All data generated or analysed during this study are included in this published article', but the custom animal dataset and the preprocessing code are not actually available. This statement should be revised to describe what is and is not released.","section":"Data availability statement"},{"comment":"Figure 1 is referenced but is not described in sufficient detail; please enlarge it and annotate each component so the pipeline is understandable without the text.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The scientific idea is simple and potentially useful, but the paper is not currently publishable because the core preprocessing is defined ambiguously in Eq. (3), the statistical evidence is limited to single-run task-sampling standard deviations, and the architecture description is incomplete. I recommend major revision with a demand for code/data release, a corrected and unambiguous preprocessing definition, and seed-level statistical analysis. The writing quality is also far below the standard of the venue, despite the interesting direction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The core idea is as simple as it sounds: take the DCT, keep only low frequencies, IDCT, add the result to the original image, and feed that into a few-shot backbone. The only controlled evidence for that idea is one ablation table, and the paper defines the low-frequency mask two different ways.\n\nWhat's new and what works: the fusion sequence as a preprocessing module dropped in front of an existing classifier is not something I've seen spelled out. The ablation in Table 3 is the right control—same backbone, unmodified vs. modified—and the gains are modest but consistent across three datasets (roughly 1 to 3 points). The idea is cheap to implement and could plausibly transfer to other pipelines. That earns credit.\n\nNow the soft spots. The Eq. (3) vs. text mismatch is not a minor typo—it changes the method materially. Eq. (3) retains a square block of size floor(h*sqrt(r)) x floor(w*sqrt(r)), which for r=0.15 means 15% of coefficients. The text says top-left floor(hr) x floor(wr), which is 2.25% of coefficients. Those are very different images: one is mild low-pass filtering, the other is aggressive smoothing. Since the entire empirical claim rests on this preprocessing, the reported gains are attached to an unspecified procedure. No code or data are released, so a reader cannot resolve which mask was actually used.\n\nThe other issues are proportionate but real. No sensitivity analysis for r or the fusion weight, no multiple seeds or paired tests, and the headline comparison in Table 2 is confounded because SFIFNet uses ResNet12+ViT while the baselines use ResNet12 or a plain CNN. The gap over FRN could easily come from the stronger backbone alone. There is also no comparison with the frequency-domain few-shot methods the paper itself cites (refs 41 and 61).\n\nAll that said, the central idea is plausible and the ablation provides some genuine signal. The flaw is load-bearing but fixable: pin down the mask, add sensitivity experiments, and compare on equal backbone footing. As written, the paper is not reproducible.\n\nThis is a paper for readers interested in cheap input augmentation for few-shot learning; it could give them a starting point, but they would need to re-derive the exact method. It deserves a serious referee—the question is well-scoped and the authors can address the ambiguity—but I'd send it back for major revision and treat the headline numbers skeptically until the mask is clarified.","headline":"Simple DCT low-pass plus fusion as a few-shot preprocessing trick shows modest same-backbone gains, but the paper defines the central mask two incompatible ways and never pins down the method.","tokens_in":12162,"tokens_out":2118,"would_cite":false,"duration_ms":22411,"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":"Adding low-frequency DCT structure to images improves few-shot classification.","keywords":["Few-shot learning","Frequency domain information","Spatial domain information","Data preprocessing","Data augmentation","Classification performance","Discrete Cosine Transform","Fine-grained image classification"],"falsifier":"Run the same ResNet12+ViT backbone and training schedule on the fine-grained bird dataset while sweeping the retention ratio $r$ from 0.05 to 0.5 and varying the fusion weight from 0.25 to 2.0. If no retention ratio or weight beats the unmodified input by a margin outside the reported confidence intervals, or only a single narrow parameter value does, then the claimed benefit is not robust evidence for frequency fusion.","tokens_in":1855,"feed_emoji":"🖼️","tokens_out":2602,"duration_ms":89735,"temperature":0.7,"pith_summary":"The paper proposes SFIFNet, a network that feeds a few-shot classifier a modified input formed by adding an image's low-frequency Discrete Cosine Transform (DCT) reconstruction back onto the original image. The central claim is that this frequency-domain fusion gives the backbone, here a ResNet12 plus Vision Transformer, richer and more stable feature representations than using spatial pixels alone, and that this improves classification when labeled examples are scarce. The claim is supported mainly by an ablation in which the same backbone with the preprocessing outperforms the unmodified version by about 1.1 to 3.3 percentage points across a fine-grained bird dataset, a dog-breed dataset, and a custom animal dataset, in both 1-shot and 5-shot settings. If true, the method matters because it is a cheap input-level augmentation that can be dropped into existing few-shot classifiers without changing their training objectives.","feed_headline":"Low-frequency fusion lifts few-shot accuracy","feed_subtitle":"Adding an image's DCT low-frequency reconstruction to the original pixels improves fine-grained classification with few labels.","key_machinery":"The load-bearing mechanism is a three-step preprocessing pipeline. First, the image is transformed with the 2D Discrete Cosine Transform, converting pixel values into frequency coefficients. Second, a low-pass mask keeps only the top-left $\\lfloor h\\sqrt{r}\\rfloor \\times \\lfloor w\\sqrt{r}\\rfloor$ coefficients with retention ratio $r=0.15$ and zeroes the rest. Third, the inverse DCT reconstructs a low-frequency image that is added directly to the original pixels, and the sum becomes the network input. The paper also describes and rejects a gradient-threshold low-pass filter as too slow. This machinery is what carries the causal claim: any accuracy gain is attributed to the frequency-domain fusion rather than to a new classifier.","core_discovery":"SFIFNet's central discovery is that the low-frequency component of an image, recovered by taking a two-dimensional DCT, zeroing all coefficients outside the top-left block defined by $\\lfloor h\\sqrt{r}\\rfloor \\times \\lfloor w\\sqrt{r}\\rfloor$ with $r=0.15$, and applying the inverse DCT, can be added pixel-wise to the original image to produce an input that yields better few-shot classification than the original image alone. The paper interprets this as integrating frequency-domain and spatial-domain information: the low-frequency reconstruction highlights main structure and suppresses high-frequency detail and noise, while the preserved original image keeps the details, so the fused input carries both. Across the three datasets, the fused input outperforms the unmodified input on the same ResNet12+ViT backbone in all six reported 5-way 1-shot and 5-way 5-shot settings, and SFIFNet also reports higher accuracy than the FRN, DeepEMD, DeepBDC, and BSNet baselines.","pith_inferences":["The paper never varies the retention ratio $r=0.15$ or the fusion weight, so a natural next experiment is a sweep over both; if the gain holds over a wide range, the method is robust, and if it peaks only at a single narrow setting, the method is a tuned preprocessing trick rather than a general principle.","The headline comparison against FRN, DeepEMD, DeepBDC, and BSNet uses different backbones and different software versions, so the Table 2 margins are not a clean test; the same-backbone ablation in Table 3 is the comparison that actually isolates the contribution of frequency fusion.","A testable extension is to apply the same preprocessing to medical or camera-trap images where lesions or rare species are distinguished by local texture; if high-frequency cues are diagnostic, the current low-pass fusion may need to be replaced by a band-pass fusion that keeps some mid-frequency detail.","The paper's rejection of gradient-based low-frequency selection on speed grounds leaves open whether a better low-frequency mask could yield larger gains; the claim is about the fusion concept, not about the optimality of the square-corner mask."],"forward_implications":["Fusing the low-frequency DCT reconstruction into input images improves average classification accuracy on all three evaluated datasets in both 5-way 1-shot and 5-way 5-shot settings, by about 1.1 to 3.3 percentage points over the same ResNet12+ViT backbone.","Because the fusion happens before the backbone, the preprocessing can be inserted into existing few-shot pipelines as a drop-in input transform, as long as the DCT, IDCT, and addition preserve the image size.","Retaining only low-frequency structure suppresses high-frequency noise and detail variation, which the paper argues reduces the model's tendency to over-fit sample-specific noise when training data are scarce.","The method's gains on fine-grained benchmarks suggest that, for few-shot fine-grained classification, low-frequency structural and contour cues are a reliable source of discriminative signal."],"supporting_citations":[{"why":"Supplies the premise that frequency-domain representations carry useful feature information beyond spatial pixels.","marker":"[41]"},{"why":"Gives the 2D DCT definition on which Eq. (1) and the low-frequency extraction are built.","marker":"[43]"},{"why":"Provides the original discrete cosine transform formulation that Eq. (1) applies.","marker":"[44]"},{"why":"Defines the 2D DCT/IDCT pair used in Eq. (4) to reconstruct the low-frequency image.","marker":"[45]"},{"why":"Introduces the low-frequency generalization versus high-frequency discrimination distinction used to justify keeping only low-frequency coefficients.","marker":"[61]"},{"why":"Supplies the Vision Transformer half of the ResNet12+ViT backbone used in the ablation and comparisons.","marker":"[51]"},{"why":"Provides the fine-grained bird dataset used for the main performance comparison and ablation.","marker":"[47]"},{"why":"Provides the fine-grained dog-breed dataset used for the main performance comparison and ablation.","marker":"[48]"},{"why":"Defines the episodic 5-way 1-shot and 5-way 5-shot evaluation protocol used in all experiments.","marker":"[50]"}],"fun_headline_variants":["Fusing DCT low-freq boosts few-shot learning","Low-freq DCT fusion sharpens few-shot models","Add low-frequency DCT to images for better few-shot","Spatial-frequency fusion improves few-shot accuracy","Low-frequency image fusion aids few-shot classification"],"cache_read_input_tokens":14208,"weakest_assumption_plain":"The argument depends on the assumption that the low-frequency block of the DCT, fixed at 15% retention and added back additively, captures class-discriminative structure without destroying the high-frequency details needed to tell similar classes apart, a premise the paper does not test by varying the retention ratio or the fusion weight.","fun_headline_variants_meta":{"raw":{"variants":["Fusing DCT low-freq boosts few-shot learning","Low-freq DCT fusion sharpens few-shot models","Add low-frequency DCT to images for better few-shot","Spatial-frequency fusion improves few-shot accuracy","Low-frequency image fusion aids few-shot classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1223,"prompt_tokens":892,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":508,"tokens_out":331,"duration_ms":3565,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:50:29.863627+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ResNet12+ViT backbone and training schedule on the fine-grained bird dataset while sweeping the retention ratio $r$ from 0.05 to 0.5 and varying the fusion weight from 0.25 to 2.0. If no retention ratio or weight beats the unmodified input by a margin outside the reported confidence intervals, or only a single narrow parameter value does, then the claimed benefit is not robust evidence for frequency fusion.","supporting_citations":[{"cited_title":"InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1740–1749 (2020)","cited_arxiv_id":null,"evidence_quote":"Supplies the premise that frequency-domain representations carry useful feature information beyond spatial pixels."},{"cited_title":"& Jalloh, I","cited_arxiv_id":null,"evidence_quote":"Defines the 2D DCT/IDCT pair used in Eq. (4) to reconstruct the low-frequency image."},{"cited_title":"& Belongie, S","cited_arxiv_id":null,"evidence_quote":"Provides the fine-grained bird dataset used for the main performance comparison and ablation."},{"cited_title":"& Li, F.-F","cited_arxiv_id":null,"evidence_quote":"Provides the fine-grained dog-breed dataset used for the main performance comparison and ablation."},{"cited_title":"neural information processing systems29(2016)","cited_arxiv_id":null,"evidence_quote":"Defines the episodic 5-way 1-shot and 5-way 5-shot evaluation protocol used in all experiments."}],"review_version":1}