{"id":"de33a2f5-fb73-4304-8ca6-14b49ab1d0e3","arxiv_id":"2412.20870","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SoftPatch+ removes outlier patches before building a PatchCore memory bank and softens the anomaly score, making unsupervised defect detection robust when training data contains 10% to 40% noise.","lead":"This paper reports a method for industrial defect detection when the training images themselves contain defective samples. It filters out anomalous image patches before building a memory bank and re-weights the remaining features, reporting stable performance with up to 40% training noise.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The high-noise robustness claim is established only in the Overlap protocol, where augmented test anomalies are placed in training; no-overlap results show SoftPatch+ does not beat PatchCore for classification, so the headline claim may not generalize to unseen defects.","rationale":"The paper's central claim is robustness to 10–40% noise in fully unsupervised AD. After reading, the most load-bearing condition is not the statistical-outlier assumption per se (the paper gives evidence that LOF/Gaussian fusion is stable across noise ratios) but the validity of the evaluation used to demonstrate it. The no-overlap results, which are the only ones that measure generalization to unseen anomalies, show that SoftPatch+ does not beat PatchCore for image-level classification; the impressive high-noise numbers come from the Overlap protocol, where augmented versions of test anomalies are inserted into the training set. This makes the headline claim underdetermined: if the protocol is invalid or unrepresentative, the claimed robustness at 40% noise is not established. A no-overlap high-noise test is feasible on VisA/BTAD or with synthetic anomalies and would settle it. I do not allege manipulation; the paper discloses the protocol and even flags dataset-distribution dependence. The reader's conditional verdict is appropriate; my concern is a sharper statement of the same methodological risk, so no verdict change is needed. Credit is due for released code and fixed hyperparameters, but those do not resolve the evaluation-protocol issue.","tokens_in":21734,"tokens_out":7116,"duration_ms":76102,"concrete_test":"Construct a strictly no-overlap high-noise benchmark by splitting anomalous images by instance/defect type: train with one disjoint subset of anomalies as noise and test on the held-out subset, at 20% and 40% noise, on VisA or BTAD-02 (or MVTecAD with synthetic defect injection). Use the paper's fixed hyperparameters (LOF-K=6, tau_cls=0.15, tau_seg=0.50, coreset ratio 10%) and report I-AUROC and P-AUROC over at least 3 random splits. If SoftPatch+'s margin over PatchCore disappears or reverses, the claimed high-noise robustness is an artifact of the Overlap protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4 defines No-Overlap and Overlap settings. No-Overlap removes from the test set the anomalous samples injected into training; Overlap keeps them in the test set after augmentation (rotation ±3°, translation ±3%). Because MVTecAD and VisA have limited anomalous test samples, the 10–40% noise experiments behind the headline robustness claim are run only in Overlap (Sec. 5.2, Fig. 4), and this is exactly where SoftPatch+'s advantage is largest. In the No-Overlap classification tables, SoftPatch+ actually trails PatchCore: 0.982 vs 0.984 on MVTecAD (Table 1) and 0.937 vs 0.944 on VisA (Table 3). The Sec. 4 Discussion asserts that \"fully unsupervised anomaly detection does not face overfitting issues with overlap settings,\" but no evidence is provided, and the protocol makes test anomalies near-duplicates of training noise after augmentation. The high-noise comparison therefore tests recognition of previously seen defect instances (or filtering of known noisy images) more than generalization to unseen anomalies. The paper's own Limitations section concedes that \"the current evaluation method may be related to the specific dataset distribution.\" This is the load-bearing weakness: the central claim of superior robustness at 10–40% noise is not validated for unseen anomalies.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses unsupervised anomaly detection when the training set itself contains anomalous (noisy) samples, a setting it calls fully unsupervised industrial AD. The proposed methods, SoftPatch and SoftPatch+, build on PatchCore. Features from a pretrained backbone are grouped by spatial position; per-position outlier scores are computed with nearest-neighbor distance, multivariate Gaussian/Mahalanobis distance, or LOF; the top tau percent of patches are removed before coreset construction; and the remaining patches are stored with soft outlier weights that re-weight the PatchCore nearest-neighbor anomaly score. SoftPatch+ fuses rank-normalized LOF and Gaussian scores. Experiments on MVTecAD, VisA, and BTAD are reported under two protocols, no-overlap and overlap, with noise ratios up to 10-40%. The paper claims that both methods outperform state-of-the-art AD methods under noisy training data and remain competitive in the clean setting.","tokens_in":22039,"tokens_out":10467,"duration_ms":94699,"significance":"If the claims are validated, the patch-level denoising and soft-weighting mechanism would be a practically useful contribution, since production-line training data cannot always be guaranteed clean. The methods are clearly specified, the code is released, and the experiments include repeated runs with standard deviations, which aids reproducibility. The ablation study supports the value of patch-level rather than image-level denoising. However, the evaluation protocol leaves the headline claim conditional: the high-noise results are obtained only in the overlap protocol, where test anomalies have near-duplicates in the training noise, and in no-overlap classification the method does not surpass PatchCore. The significance of the work therefore depends on whether the authors can validate robustness to unseen anomaly instances or anomaly types.","major_comments":[{"comment":"The abstract claims that both SoftPatch and SoftPatch+ outperform state-of-the-art AD methods on MVTecAD and VisA, but the no-overlap classification results do not support this for SoftPatch+. On MVTecAD at 10% noise, SoftPatch+ obtains 0.982 image-level AUROC versus 0.984 for PatchCore; on VisA at 8% noise, SoftPatch+ obtains 0.933 and SoftPatch-LOF obtains 0.937, both below PatchCore's 0.944. The no-overlap setting is the only protocol that evaluates anomaly instances not present in the training noise, so the claimed superiority in robustness is not established for unseen defects. Please either restrict the claim to the overlap protocol, or add no-overlap evaluations at higher noise ratios, for example by holding out entire anomaly types rather than individual instances.","section":"Sec. 5.2, Tables 1 and 3"},{"comment":"The high-noise curves in Fig. 4 and the headline 10-40% results are generated only under the overlap protocol, where the training noise consists of augmented copies of the test anomaly images. The Discussion in Sec. 4 asserts that fully unsupervised anomaly detection does not face overfitting issues with overlap settings, but no analysis is provided to support this assertion. Because each test anomaly has a near-duplicate (rotation/translation) in the training set, the overlap protocol largely measures the method's ability to filter or recognize the specific injected defect instances, not its ability to generalize to unseen anomaly appearances. The Limitations section itself concedes that the current evaluation method may be related to the specific dataset distribution. Please provide evidence from a protocol with strictly disjoint anomaly instances or anomaly types before claiming robust performance under 10-40% noise.","section":"Sec. 4, Overlap protocol and Fig. 4"},{"comment":"The removal thresholds tau_cls=0.15 and tau_seg=0.50 are described as fixed constants, but Fig. 5 shows that these values were selected by sweeping the threshold and measuring AUROC on the evaluation benchmark. In a fully unsupervised setting, no labels are available for such selection, so benchmark-tuned thresholds can inflate the reported gains. Please demonstrate that the conclusions are robust across a plausible range of thresholds, or propose an unsupervised criterion for setting tau_cls and tau_seg.","section":"Sec. 5.1 and Sec. 5.3.2"}],"minor_comments":[{"comment":"Tables 2 and 10 report different pixel-level AUROC for SoftPatch+ at 10% noise on MVTecAD (0.978 in Table 2 Overlap versus 0.951 in Table 10); please reconcile these numbers.","section":"Tables 2 and 10"},{"comment":"The indicator variables s_m are defined only as elements of {0,1}; the text should state the rule for setting them, for example which discriminators are included in SoftPatch+.","section":"Sec. 3.3, Eq. (8)"},{"comment":"Figure 4 has no legend or line labels in the caption, making it difficult to identify which curve corresponds to which method.","section":"Fig. 4"},{"comment":"There are several typographical errors, including 'SoftPach' in Sec. 5.2 and 'articularly' in the Abstract; a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The overlap protocol is both the main novelty and the main weakness of this journal version; I would push the authors to provide a disjoint-instance or disjoint-type evaluation before accepting the high-noise robustness claim. The incremental contribution over the NeurIPS 2022 SoftPatch paper is moderate, but the broader evaluation and the multi-discriminator fusion make it a reasonable journal candidate if the evaluation gap is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper tackles a real and under-studied problem—defect detection on production lines where the training data contains defective samples—and the patch-level denoising idea is a genuine step. But the central claim that SoftPatch+ is robust to 10–40% noise is established only in the \"overlap\" protocol, where the test anomaly images themselves, lightly augmented, are added to training. In the \"no-overlap\" protocol, which tests generalization to unseen defects, SoftPatch+ slightly trails PatchCore on MVTecAD classification (0.982 vs 0.984) and on VisA (0.937 vs 0.944). The headline overreaches.\n\nWhat is actually new: removing outlier patches before coreset construction, storing the outlier scores as soft weights, and using a rank-based fusion of LOF and Gaussian discriminators. That fusion is simple and does help in high-noise overlap scenarios. Code is released, and the paper is transparent about limitations, admitting in the Limitations section that the evaluation may be dataset-specific.\n\nSoft spots, in proportion: the overlap protocol is the main evidence for high-noise robustness, but it has train/test exposure—the same anomalies appear in both training and testing. The Discussion's claim that fully unsupervised AD \"does not face overfitting issues with overlap settings\" is unsupported, and the no-overlap results undercut the robustness advantage. Hyperparameters tau_cls and tau_seg are selected on the same benchmarks (Sec. 5.3.2), so the fixed values 0.15/0.50 are effectively tuned to the test set. The main tables (Tables 1–4) lack error bars, despite the ablation tables reporting them, so the small no-overlap differences are likely within run-to-run noise. The \"first\" claim for fully unsupervised industrial AD is also a bit strong given IGD already handles contaminated data, though the patch-level treatment and industrial framing are new.\n\nThe math is not circular: the anomaly score in Eq. 10 is a re-weighted nearest-neighbor distance, not a fitted quantity. The circularity burden is evaluative, not algorithmic.\n\nWho this is for: anyone working on anomaly detection under label noise or on memory-bank methods. It deserves a serious referee—the problem is important, the method is clearly specified, and the code is public. But it needs major revision: reframe the claims to separate overlap from no-overlap, add evidence for generalization to unseen defects (e.g., synthetic anomalies or held-out defect types), report error bars in all tables, and validate the thresholds on a separate set.\n\nRecommendation: send to peer review, but expect heavy revision.","headline":"Patch-level denoising for contaminated training data is a good idea, but the paper's headline robustness claim only holds in the overlap protocol, which leaks test anomalies into training; in the cleaner no-overlap setting, SoftPatch+ actually trails PatchCore.","tokens_in":22539,"tokens_out":4455,"would_cite":false,"duration_ms":45329,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SoftPatch+ claims that fully unsupervised industrial anomaly detection with up to 40% noisy training data is practical: patch-level outlier scoring removes defective patches before coreset construction, and the stored outlier scores…","keywords":["fully unsupervised anomaly detection","noisy training data","patch-level denoising","coreset memory bank","local outlier factor","Mahalanobis distance","industrial defect detection","anomaly segmentation"],"falsifier":"Build a training set where defects are large, dense, and visually close to normal texture (e.g., roughness variations covering most of a machined part), inject 40% such noisy images, and measure the denoiser's patch-level outlier AUROC against ground-truth defect maps; if the denoiser ranks defective patches near chance, the memory bank becomes corrupted and SoftPatch+ should fall to the PatchCore baseline. Extending the overlap setting to 45-50% noise on MVTecAD would also test the declared ceiling.","tokens_in":21564,"feed_emoji":"🏭","tokens_out":4151,"duration_ms":40702,"temperature":0.7,"pith_summary":"This paper argues that unsupervised anomaly detection in real production lines must tolerate contaminated training data, since defective samples inevitably slip into the nominal set. Existing memory-bank methods like PatchCore collapse under such noise because they build their coreset from every training patch without question. SoftPatch+ removes defective patches before coreset construction by scoring each patch's outlierness within its spatial position across the training set, then uses those scores as soft weights to down-weight any remaining noise during inference. The paper claims this patch-level denoising, combined with a fusion of LOF and Gaussian discriminators, keeps both image-level classification and pixel-level segmentation accurate even when 10% to 40% of training data is defective. A sympathetic reader would care because it addresses a practical bottleneck: deploying inspection systems on new production lines without manual data cleaning.","feed_headline":"Detects defects even when 40% of training data is defective","feed_subtitle":"Patch-level denoising removes defective patches before memory-bank construction and beats prior methods on three industrial benchmarks.","key_machinery":"The central object is the position-grouped patch feature stack φ_i(h,w), where each patch's outlier score is computed per spatial location across all training images. Three discriminators supply these scores: nearest-neighbor distance W_nn, Mahalanobis distance W_mvg under a per-position multivariate Gaussian, and Local Outlier Factor W_LOF, which measures a patch's local density relative to its k-nearest neighbors. SoftPatch+ rank-normalizes each score and averages them into a fused vote W_vote, then removes the top τ% of patches by fused score before greedy coreset sampling. The remaining scores are stored as soft weights W_{m*} in the memory bank, and each test patch's anomaly score is computed as s_{h,w} = W_{m*} ||p - m*||_2, which re-weights the nearest-neighbor distance by the noise level of the matched memory patch.","core_discovery":"The paper's central claim is that noisy training data in unsupervised anomaly detection is best handled at the patch level, not the image level, because defects typically occupy only a tiny image region. The proposed SoftPatch+ method groups patch features by position, computes outlier scores for every training patch using nearest-neighbor distance, per-position multivariate Gaussian Mahalanobis distance, and Local Outlier Factor, removes the top τ% of patches before coreset sampling, and stores the remaining outlier scores as soft weights in the memory bank. At inference, the nearest-neighbor anomaly score of each test patch is multiplied by the stored soft weight of its matched memory patch, softening the decision boundary. The paper reports that SoftPatch and SoftPatch+ outperform previous state-of-the-art methods on MVTecAD, VisA, and BTAD under noise ratios from 10% to 40%, and that SoftPatch+ remains close to the clean-training upper bound in both overlap and no-overlap settings.","pith_inferences":["The per-position outlier assumption implies the method would struggle on whole-surface defects that cover most of the image, where few normal patches remain to anchor the feature distribution; this failure mode is not covered by the paper's benchmarks.","The patch-level denoising and soft-weighting scheme could be ported to other nearest-neighbor memory methods, such as graph-based or transformer-based detectors, as a preprocessing step independent of the coreset algorithm.","A natural testable extension is to use the stored soft weights as confidence values for active learning, requesting human labels only for the lowest-weight patches in the memory bank.","Because the paper keeps τ and LOF-K constant across all noise ratios, an adaptive threshold estimator that infers the noise level from the outlier-score distribution could push robustness beyond the reported 40% ceiling."],"forward_implications":["A memory-bank detector can be trained on unfiltered production-line images containing up to roughly 40% defective samples and still localize defects at test time.","Patch-level denoising outperforms image-level denoising because normal regions of defective images are retained and contribute to the coreset.","Fusing LOF with Gaussian (Mahalanobis) discriminators is more robust at high noise ratios than either alone, since LOF degrades with heavy contamination while Gaussian stays stable.","With no added noise, SoftPatch and SoftPatch+ match conventional clean-data performance, so adopting them does not sacrifice standard-setting results.","Using separate removal thresholds for classification (τ = 0.15) and segmentation (τ = 0.50) improves both image-level and pixel-level AUROC."],"supporting_citations":[{"why":"Supplies the PatchCore memory-bank baseline whose coreset construction and nearest-neighbor scoring SoftPatch+ modifies.","marker":"[10]"},{"why":"Supplies the Local Outlier Factor density-based outlier discriminator used in SoftPatch and fused in SoftPatch+.","marker":"[40]"},{"why":"Supplies the PaDiM multivariate Gaussian patch-modeling approach used for the Mahalanobis discriminator.","marker":"[29]"},{"why":"Provides the MVTecAD benchmark, the primary dataset for evaluating classification and segmentation under noise.","marker":"[13]"},{"why":"Provides the VisA large-scale industrial dataset used to evaluate generalization across object and texture domains.","marker":"[14]"},{"why":"Provides IGD, the prior method for unsupervised AD with contaminated data that SoftPatch+ is compared against.","marker":"[38]"},{"why":"The preliminary conference version of SoftPatch that this journal work extends with multi-discriminator fusion and broader evaluation.","marker":"[16]"}],"fun_headline_variants":["Patch-level denoising tames noisy training data for anomaly detection","First fully unsupervised AD that handles noisy data at patch level","Denoise patches before memory bank to beat noisy-training AD","Patch-level denoising for unsupervised AD under 40% noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Defective patches are statistical outliers in the per-position feature distribution of the noisy training set, so LOF, Mahalanobis, or nearest-neighbor scores can rank them before coreset construction, and if defective patches are densely clustered or highly similar to normal patches, the denoiser will retain noise and remove normal patches, breaking the memory bank.","fun_headline_variants_meta":{"raw":{"variants":["Patch-level denoising tames noisy training data for anomaly detection","First fully unsupervised AD that handles noisy data at patch level","Denoise patches before memory bank to beat noisy-training AD","Patch-level denoising for unsupervised AD under 40% noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000795,"raw_usage":{"total_tokens":3533,"prompt_tokens":1011,"completion_tokens":2522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":2451}},"tokens_in":627,"tokens_out":2522,"duration_ms":16938,"temperature":1.0,"reasoning_tokens":2451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:07:46.003671+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a training set where defects are large, dense, and visually close to normal texture (e.g., roughness variations covering most of a machined part), inject 40% such noisy images, and measure the denoiser's patch-level outlier AUROC against ground-truth defect maps; if the denoiser ranks defective patches near chance, the memory bank becomes corrupted and SoftPatch+ should fall to the PatchCore baseline. Extending the overlap setting to 45-50% noise on MVTecAD would also test the declared ceiling.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PatchCore memory-bank baseline whose coreset construction and nearest-neighbor scoring SoftPatch+ modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Local Outlier Factor density-based outlier discriminator used in SoftPatch and fused in SoftPatch+."},{"cited_title":"Defard, A","cited_arxiv_id":null,"evidence_quote":"Supplies the PaDiM multivariate Gaussian patch-modeling approach used for the Mahalanobis discriminator."},{"cited_title":"Bergmann, M","cited_arxiv_id":null,"evidence_quote":"Provides the MVTecAD benchmark, the primary dataset for evaluating classification and segmentation under noise."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the VisA large-scale industrial dataset used to evaluate generalization across object and texture domains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides IGD, the prior method for unsupervised AD with contaminated data that SoftPatch+ is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The preliminary conference version of SoftPatch that this journal work extends with multi-discriminator fusion and broader evaluation."}],"review_version":1}