{"id":"046de19d-aee9-4ddf-a113-0d523a00b1e0","arxiv_id":"2508.16034","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Replacing PaDiM's random channel sampling with per-layer wavelet subband selection yields test-set-optimized MVTec AD averages of 99.32% Image-AUC and 92.10% Pixel-AUC, and shows LL bands help detection while detail bands help localization.","lead":"An industrial image inspection method, WE-PaDiM, replaces PaDiM's random channel sampling with a wavelet transform that separates CNN features into frequency bands before anomaly scoring. It reports high MVTec AD accuracy, but the headline numbers come from choosing the best configuration per class on the test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline AUCs are per-class test-set-optimized oracles, not a single system: fixed configs from Tables 1/2 drop to ~98.84/86.09 (Image/Pixel) and ~94.30/90.77, so 99.32/92.10 overstates a deployable result.","rationale":"Reader's assessment is fair. My independent reading confirms the per-class oracle protocol is the central problem: Tables 4/5 rank configurations by test AUC, with per-class candidate counts up to 430, and the two abstract numbers are from separate optimization criteria. Tables 1/2 already provide a rough fixed-config check and show meaningful drops, especially Pixel-AUC. The paper's own §5 limitation explicitly acknowledges the missing controlled PaDiM random-selection baseline; without that, the stated advantage of DWT selection over random selection is not established. However, the method itself is clearly specified, code is available, and internal ablations (Table 3, Fig. 3) are coherent; the defect is in evaluation/reporting and can be fixed by re-running with a pre-registered configuration and an original-PaDiM control. Therefore no verdict movement beyond the existing CONDITIONAL is warranted.","tokens_in":18396,"tokens_out":6880,"duration_ms":58653,"concrete_test":"Run the full 15-class MVTec evaluation with a single configuration fixed across all classes: (a) efficientnet-b6, Haar, J=1, subbands=LL, sigma=2.0, eps=0.1 (Table 1's best average-Image setup) and (b) resnet18, Haar, J=1, subbands=HH_LH_LL, sigma=2.0, eps=0.001 (Table 2's best average-Pixel setup). Compute per-class means for both Image and Pixel AUC under each fixed config. If the means reproduce Tables 1/2 (~0.9884/0.8609 and ~0.9430/0.9077) rather than 0.9932/0.9210, the headline is an artifact of per-class test-set selection. Also report the joint (Image, Pixel) average from the same per-class configurations used in Table 4, and from the same used in Table 5, to confirm the two headline numbers come from different configurations. This single re-analysis settles whether the stated performance is achievable without oracle selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Abstract; §4.3; Conclusion) is that WE-PaDiM reaches 99.32% Image-AUC and 92.10% Pixel-AUC. These are oracle averages: for each MVTec class, Tables 4/5 select the top configuration directly on the test set from a grid that can contain hundreds of candidates (bottle: 430; carpet: 236), and the two headline numbers come from different optimization criteria (Table 4 best-Image configs have average Pixel 0.8383; Table 5 best-Pixel configs have average Image 0.9555). No single WE-PaDiM instantiation achieves both. The paper's own fixed-across-classes backbone configurations are substantially lower (Table 1: efficientnet-b6, Haar, J=1, LL, sigma=2, eps=0.1 -> 0.9884 Image/0.8609 Pixel; Table 2: resnet18, Haar, J=1, HH_LH_LL, sigma=2, eps=0.001 -> 0.9077 Pixel/0.9430 Image), and even these are test-selected at backbone level. Section 5 explicitly states that a direct controlled comparison with original PaDiM's random selection was not the primary focus, so the claim that DWT selection is a competitive alternative to random selection is not directly evidenced. The load-bearing weakness is therefore evaluation protocol, not the algorithm itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Wavelet-Enhanced PaDiM (WE-PaDiM), which replaces the random channel selection in PaDiM with a structured, frequency-domain selection based on a 2D Discrete Wavelet Transform applied independently to multi-layer CNN feature maps. Selected subbands (e.g., LL, LH, HL) are concatenated channel-wise and fed into PaDiM's multivariate Gaussian patch modeling. The method is evaluated on MVTec AD with ResNet-18 and EfficientNet B0-B6 backbones. The authors report average Image AUC of 99.32% and Pixel AUC of 92.10% under per-class optimized configurations, and provide a detailed analysis of the impact of wavelet type, subband choice, decomposition level, and other hyperparameters on detection and localization performance.","tokens_in":18785,"tokens_out":3913,"duration_ms":42408,"significance":"If the headline results corresponded to a fixed, deployable configuration, WE-PaDiM would be a simple, efficient, and interpretable alternative to PaDiM's random channel selection. The paper's strengths include a clear method description, a public code repository, and a systematic ablation of subband choices with significance tests and visualizations. However, the central performance claim is currently based on per-class test-set-optimized configurations, which are oracle maxima over sometimes hundreds of candidate settings. Consequently, the contribution as stated is not yet established; the paper needs a defensible evaluation protocol or a major reframing of the claims.","major_comments":[{"comment":"The headline numbers 99.32% Image-AUC and 92.10% Pixel-AUC are per-class test-set-optimized maxima, not the performance of a fixed WE-PaDiM system. For each MVTec class, the best configuration is selected directly on the test set from grids with as many as 430 candidates (Table 4, bottle) or 236 (carpet). This is selection on the evaluation set. The paper's own fixed-across-classes backbone configurations are substantially lower (Table 1: efficientnet-b6, Haar, J=1, LL, σ=2, ε=0.1 => 0.9884 Image / 0.8609 Pixel; Table 2: resnet18, Haar, J=1, HH_LH_LL, σ=2, ε=0.001 => 0.9077 Pixel / 0.9430 Image), and even these are selected at the backbone level on the test set. Moreover, the two headline figures come from different selection criteria: Table 4's 0.9932 Image average corresponds to only 0.8383 average Pixel AUC, while Table 5's 0.9210 Pixel average corresponds to 0.9555 average Image AUC;","section":"Abstract; §4.3; §4.8; Tables 4/5"},{"comment":"The paper states that 'a direct, controlled comparison with the original PaDiM's random selection ... was not the primary focus of the presented results.' This is load-bearing because the core motivation is to replace random selection with DWT-based selection. Without a controlled comparison using the same backbones, layers, and evaluation protocol (e.g., matching the number of selected dimensions DW to PaDiM's DR), the claim that WE-PaDiM is a 'competitive and interpretable alternative' to random selection is not supported. The component analysis in Fig. 3 shows within-method effects, but not how the structured selection compares with PaDiM's random channel selection on the same backbone. A fixed-configuration PaDiM baseline must be included.","section":"§5, 'Comparison with Original PaDiM'"},{"comment":"The per-class 'best' configurations are obtained under highly uneven search effort: the 'Opt. # Cfgs' column ranges from 1 (e.g., Table 5, screw) to 430 (Table 4, bottle). This means that per-class maxima are not directly comparable across classes, and classes with more searched configurations have a higher chance of a spuriously favorable maximum. Reporting per-class maxima without search-effort or variance information overstates the consistency of the method. The authors should either provide fixed-configuration results across all classes or present the per-class search as a hyperparameter optimization analysis, with the number of trials and the distribution of scores clearly reported.","section":"§4.8, Tables 4/5"},{"comment":"The statistical significance analysis in §4.5 is presented as evidence for the benefit of LL over detail subbands, but the comparisons are performed over all configurations in the search grid, which are not independent (they share backbones, layers, and hyperparameter families) and are selected on the test set. The p-values are therefore optimistic and should be interpreted as descriptive rather than inferential. Table 3's averages over all backbones, classes, and configurations mix very different regimes; the qualitative conclusions about LL for Image AUC and detail subbands for Pixel AUC are plausible, but the confidence intervals and test-set selection need to be acknowledged.","section":"§4.5, Fig. 3; Table 3"}],"minor_comments":[{"comment":"Typographical error: 'regularization term ϵ. )' should read 'regularization term ϵI'.","section":"§3.4, Eq. (7)"},{"comment":"The subsection 'Impact of Subband Selection' appears twice, nearly verbatim (in §4.4 and again in §4.7). One occurrence should be removed or replaced with a cross-reference.","section":"§4.4 and §4.7"},{"comment":"Reference [16] is attributed to Ruff et al. as 'DifferNet's normalizing-flow modeling'; DifferNet is by Rudolph et al. (2021). Please correct the citation and the associated text.","section":"§2.1, reference [16]"},{"comment":"The captions state that parameters are 'unique values across tied top configurations,' but it is unclear how the single reported Image/Pixel AUC values were computed when multiple tied configurations have different secondary-metric values. State the aggregation rule (e.g., average, best secondary metric).","section":"Tables 4/5 captions"},{"comment":"The sentence 'The comparable computational efficiency (as suggested by the abstract, though detailed timings are pending)' is vague. Either provide concrete runtime/memory comparisons or omit the efficiency claim.","section":"§5, 'Interpretability and Practical Implications'"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and clearly presented, but the evaluation protocol is the central issue. The paper needs either a fixed-configuration evaluation with a proper validation split and a PaDiM comparison, or a reframing of the contribution as a per-class hyperparameter search analysis. In its current form, the abstract and conclusion present oracle maxima as the method's performance, which is likely to mislead readers. I would not recommend acceptance until this is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has one genuinely useful idea and one serious evaluation problem. The idea is to apply 2D DWT per feature map layer, keep a chosen subband set (LL, LH/HL, etc.), align, concatenate, then run PaDiM's Gaussian modeling. It is a simple, training-free plug-in, and the authors' component analysis (Table 3, Fig. 3) is the most valuable part: LL consistently helps image-level detection, detail bands help pixel-level localization, and HH can hurt. That is a citable empirical observation. Code is linked, so the work is reproducible.\n\nThe problem is how the headline results are reported. The 99.32% Image AUC and 92.10% Pixel AUC in the abstract are per-class maxima selected on the test set from grids that can contain hundreds of configurations (e.g., 430 for bottle, 236 for carpet). Worse, the two numbers come from different optimization criteria: the best-Image configurations average 83.83% Pixel AUC, and the best-Pixel configurations average 95.55% Image AUC. No single WE-PaDiM instantiation achieves both. The paper's own fixed-across-classes configurations are lower (e.g., efficientnet-b6 with Haar/LL gives 98.84% image / 86.09% pixel; resnet18 with Haar/HH_LH_LL gives 94.30% / 90.77%). And even those are test-selected at the backbone level.\n\nThe paper does not present a controlled comparison with original PaDiM on the same backbones and same training setup; Section 5 says that was \"not the primary focus.\" So the claim that DWT selection is a competitive alternative to random selection is not directly evidenced. The efficiency claim is also unmeasured.\n\nTo be fair, these are evaluation and reporting flaws, not fundamental defects in the algorithm. The method is coherent, the math is standard, and the subband analysis is a real contribution. The paper would be much stronger if the authors fixed configurations across classes for a fair comparison, added a proper PaDiM baseline, disclosed the search space size, and reported a single operational configuration with both metrics.\n\nWho should read this: people who want a cheap structured feature-selection trick for PaDiM or want to see how frequency decomposition interacts with CNN features. It deserves a serious referee, but it needs major revision before it can be accepted as-is. I would send it to review with a request for the fixed-config experiments and a direct PaDiM baseline.","headline":"A plausible, clearly described wavelet-channel-selection plug-in for PaDiM, but the headline AUCs are per-class test-set-optimized oracles, not the performance of a single system.","tokens_in":19271,"tokens_out":2654,"would_cite":false,"duration_ms":27844,"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":"Replacing PaDiM's random channel selection with wavelet frequency subbands keeps detection accuracy and adds interpretability.","keywords":["anomaly detection","anomaly localization","wavelet transform","PaDiM","feature selection","MVTec AD","industrial inspection"],"falsifier":"Run WE-PaDiM with a fixed configuration (e.g., Haar, level 1, subbands LL) on all 15 MVTec classes and compare the average Image and Pixel AUC to the per-class optimized numbers. If the fixed-configuration average is substantially lower, the headline results are mostly test-set fitting rather than robust gains.","tokens_in":1311,"feed_emoji":"🔍","tokens_out":2182,"duration_ms":65279,"temperature":0.7,"pith_summary":"WE-PaDiM claims that a structured, frequency-based feature selection can replace the random channel sampling at the heart of PaDiM without losing accuracy. The authors apply a two-dimensional discrete wavelet transform to each layer's CNN feature map, keep chosen subbands (low-frequency LL, horizontal LH, vertical HL), align them, concatenate them, and then fit PaDiM's per-patch multivariate Gaussians. On the MVTec AD benchmark, with per-class tuned configurations, the method reports an average image-level AUC of 99.32% and pixel-level AUC of 92.10%, competitive with established embedding-based detectors. The point of the work is not just accuracy: choosing subbands by frequency content gives practitioners an interpretable handle on what the model attends to (coarse structure vs. fine edges), which random selection cannot provide.","feed_headline":"Wavelet subbands replace random channels in PaDiM","feed_subtitle":"Structured DWT selection gives 99.32 Image AUC and 92.10 Pixel AUC on MVTec AD","key_machinery":"The central object is the DWT-before-concatenation pipeline: per-layer 2D DWT, subband selection S, spatial alignment, channel-wise concatenation, then PaDiM's per-patch multivariate Gaussian modeling. The load-bearing idea is that frequency content is a meaningful organizing principle: LL carries coarse/global information, LH/HL/HH carry directional detail, so selecting subbands is a structured alternative to random channel sampling. The Mahalanobis distance to the fitted normal distribution produces both the pixel-level anomaly map and the image-level score.","core_discovery":"On its own terms, the paper establishes that DWT-before-concatenation is a viable and principled substitute for random dimensionality reduction in PaDiM. Instead of drawing a fixed-size random subset of concatenated CNN channels, WE-PaDiM decomposes each layer's feature maps into LL, LH, HL, and HH subbands, selects a subset, aligns the maps to a common resolution, and concatenates them before fitting a multivariate Gaussian per spatial location and scoring with Mahalanobis distance. Across ResNet-18 and EfficientNet B0-B6 backbones on MVTec AD, the method reaches an average of 99.32% image AUC when the best configuration is selected per class, and 92.10% pixel AUC under a separate per-class","pith_inferences":["Editorial: the 99.32% Image AUC and 92.10% Pixel AUC come from two different per-class selection passes (Table 4 vs Table 5), so they do not describe a single model; a fixed configuration would likely score lower on at least one of the metrics.","Editorial: because configurations were selected per class on the test set from tens to hundreds of candidates per class, the averages may overstate deployable performance; a fair comparison would hold the configuration fixed across classes or use a held-out validation split.","Editorial: the same DWT-before-concatenation idea could be dropped into other patch-embedding anomaly detectors, not just PaDiM, with testable consequences for localization on fine defects.","Editorial: an explicit head-to-head against random channel selection at matched feature dimension would isolate whether gains come from the frequency structure itself or simply from using more or different channels."],"forward_implications":["Low-frequency (LL) subbands alone can serve as a reduced, interpretable feature set for image-level detection, since they consistently lead in image AUC.","For localization, adding directional detail subbands (LH, HL, and usually HH) improves pixel-level accuracy, so a practitioner can tune subband selection to the expected defect type.","The method stays within PaDiM's training-free, Gaussian-modeling regime, so it inherits PaDiM's practical speed and memory profile while adding only DWT overhead.","Different backbones and categories favor different wavelets and subbands, meaning configuration choice is a meaningful axis of the method rather than a nuisance parameter."],"supporting_citations":[{"why":"Supplies the MVTec AD benchmark dataset and the one-class evaluation protocol used for all reported Image AUC and Pixel AUC numbers.","marker":"[1]"},{"why":"Defines PaDiM, the base framework whose random channel selection is replaced by wavelet-based subband selection.","marker":"[2]"},{"why":"Provides the wavelet/multiresolution decomposition theory that motivates the LL/LH/HL/HH subband structure.","marker":"[3]"},{"why":"Supplies ResNet-18, one of the two backbone families used to extract multi-layer CNN features.","marker":"[4]"},{"why":"Supplies the EfficientNet B0-B6 backbone family used to demonstrate generality across architectures.","marker":"[5]"},{"why":"Serves as a strong embedding-based anomaly detection baseline that the method's results are positioned against.","marker":"[11]"}],"fun_headline_variants":["PaDiM gets wavelet boost: 99.32 Image AUC on MVTec","DWT-before-concat beats random channels in PaDiM","Wavelet subband selection lifts PaDiM to 99.32 AUC","Structured wavelets replace random feature picks in PaDiM","PaDiM with DWT: 99.32 Image AUC on industrial anomaly"],"cache_read_input_tokens":20864,"weakest_assumption_plain":"The headline numbers assume that picking the best configuration per class on the MVTec test set is a fair way to judge the method; if a single fixed configuration must be chosen across all classes, the reported averages likely drop.","fun_headline_variants_meta":{"raw":{"variants":["PaDiM gets wavelet boost: 99.32 Image AUC on MVTec","DWT-before-concat beats random channels in PaDiM","Wavelet subband selection lifts PaDiM to 99.32 AUC","Structured wavelets replace random feature picks in PaDiM","PaDiM with DWT: 99.32 Image AUC on industrial anomaly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000413,"raw_usage":{"total_tokens":2032,"prompt_tokens":862,"completion_tokens":1170,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":1085}},"tokens_in":606,"tokens_out":1170,"duration_ms":9041,"temperature":1.0,"reasoning_tokens":1085,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:34:17.974101+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run WE-PaDiM with a fixed configuration (e.g., Haar, level 1, subbands LL) on all 15 MVTec classes and compare the average Image and Pixel AUC to the per-class optimized numbers. If the fixed-configuration average is substantially lower, the headline results are mostly test-set fitting rather than robust gains.","supporting_citations":[{"cited_title":"A theory for multiresolution signal decomposition: The wavelet representation","cited_arxiv_id":null,"evidence_quote":"Provides the wavelet/multiresolution decomposition theory that motivates the LL/LH/HL/HH subband structure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EfficientNet B0-B6 backbone family used to demonstrate generality across architectures."}],"review_version":1}