acceptodds
Desk rejected at ICLR 2027

TRRC-HETMM: Risk-Controlled Texture-Disentangled Mutual Matching for Industrial Anomaly Detection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Industrial chip inspection must jointly limit missed defects, over-inspection of normal products, and computational cost. This balance is difficult when normal texture appears primarily on the carrier strip, which also belongs to the valid inspection region and therefore cannot simply be masked out. We propose Texture Reweighting and Risk Control for HETMM (TRRC-HETMM), combining texture-disentangled mutual matching, miss-risk calibration, and safe dynamic computation. The method estimates position-conditioned variation subspaces from normal templates and reweights residuals within and orthogonal to those subspaces. Independently calibrated release thresholds and guard rules govern fast decisions and fallback to full matching. Controlled high-frequency suppression and an equal-energy residual-shuffling control further support the contribution of normal strip texture to over-inspection. On a sealed industrial test set with 1,000 defective and 19,000 normal images, the frozen system misses 4 defective samples and over-inspects 2,765 normal samples, corresponding to rates of 0.40% and 14.55%, respectively. At the same four-miss operating point, HETMM over-inspects 6,718 normal samples; TRRC-HETMM reduces this number by 58.84%. Experiments on public benchmarks further assess generalization across object categories.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
industrial anomaly detection, template matching, texture disentanglement, risk-controlled inference, adaptive computation, visual inspection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ghBXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents TRRC-HETMM, an extension of HETMM for industrial inspection settings in which normal texture on a carrier strip causes false alarms but cannot be removed from the inspection region. The method has three parts. It first estimates a position-conditioned subspace of normal feature variation and reweights query-template residuals inside and outside that subspace. It then calibrates fast and full decision thresholds under separate miss-risk budgets. Finally, a dynamic policy makes fast decisions for apparently unambiguous samples and sends gray-zone or guard-triggered samples to full PTS-60 matching.

The main evaluation uses an entity-disjoint private chip dataset with 1,000 defective and 19,000 normal test images. The final system reports four missed defects and 2,765 false positives, compared with 48 misses and 3,895 false positives for HETMM at its frozen operating point. The paper also reports public-benchmark results on MVTec AD, VisA, and Real-IAD, a targeted texture-suppression intervention with an equal-energy control, module ablations, and end-to-end timing.

Critical strengths
  1. The paper takes the deployment setting seriously. Templates, validation, calibration, and test data are separated by physical chip identity, and the authors distinguish continuous-score ranking from errors at a frozen operating point. The paper is also careful about the scope of its statistical claim: the conformal result assumes exchangeability and does not guarantee safety under process drift.

  2. The mechanism study is stronger than a routine ablation. The authors combine residual-direction statistics with a targeted strip-frequency intervention, an RMS-matched shuffled control, and region-specific detection counts. This provides concrete evidence that normal strip texture contributes to false positives and that the improvement is not explained by indiscriminate response suppression alone.

Critical weaknesses
  1. The supplementary artifact does not adequately support the central numerical results and raises concerns about the paper's reproducibility claims. The paper states that the archive provides the implementation used on the experimental server. However, the archive reportedly states that it is not a verified export of that environment, that the original server artifacts and end-to-end numerical regression are unavailable, and that the exact fast-template order and score scaling have not been regression-tested. Calibrated thresholds must also be copied manually into the inference configuration. Although the 20 unit tests pass, they verify only core operations rather than the reported industrial counts, AUROC, or timing. Since the private data and several intervention settings are also unavailable, the main experimental results cannot currently be independently verified.

  2. The methodological novelty is limited, and the advantage of the proposed texture-disentangled metric over simpler alternatives is not sufficiently established. The framework combines a position-conditioned low-rank residual metric, standard conformal risk calibration, and heuristic dynamic routing on top of HETMM. While this integration addresses a practical industrial problem, the individual components are closely related to existing techniques. In particular, the texture-disentangled distance is essentially a position-conditioned quadratic metric based on covariance eigendecomposition and residual reweighting. Its relationship to established approaches such as Mahalanobis distance, PCA residual distance, and feature whitening is not sufficiently investigated. Without comparisons against these simpler statistical alternatives under the same HETMM framework, it remains unclear whether the proposed metric provides a distinct methodological contribution or whether similar improvements could be obtained through standard covariance-aware reweighting.

  3. The public evidence is not strong enough to establish broad generality. Table 2 reports AUROC pooled over all test samples, without per-category results, standard category-macro averages, confidence intervals, or analysis of variability across different data splits. Pooled AUROC can be affected by differences in anomaly-score scales across categories and may not adequately reflect performance within individual object categories. Although the private industrial results are impressive, the public experiments do not yet provide a sufficiently comprehensive and independently auditable substitute, particularly given the reproducibility limitations of the private dataset.

Questions for authors
  1. The paper states that the supplementary archive contains the implementation used on the experimental server, whereas the archive reportedly indicates that it is not a verified server export and that the original artifacts and numerical regression are unavailable. Which statement is accurate? Can Tables 1–5 still be regenerated from a preserved internal environment, and can the authors provide hashes or derived per-sample scores linking that environment to the reported results?

  2. How were the first 30 fast-route templates selected and ordered in the actual experiments, and what exact score scaling produced the thresholds 0.58, 0.68, and 0.72? The archive reportedly states that these details have not been regression-tested. Would changing the template order affect the routing counts or miss rate?

  3. How does the proposed texture-disentangled distance compare with simpler position-wise covariance-aware metrics, such as Mahalanobis distance, PCA residual distance, or feature whitening, under the same HETMM framework? Such comparisons would help establish whether the proposed formulation offers a distinct methodological advantage beyond standard statistical reweighting.

  4. Please report category-wise image and pixel AUROC, standard category-macro averages, and uncertainty estimates for MVTec AD, VisA, and Real-IAD. Why is pooled AUROC used as the main public metric when anomaly-score scales can differ across categories?

  5. With 600 defective calibration samples, the minimum corrected risk is 1/6011/601, so the two branch budgets are feasible only with zero calibration misses. How stable are the selected thresholds and test miss rate across repeated entity-disjoint calibration splits?

1 comment

Response to Reviewer ghBX

Thank you for the detailed questions. We clarify the implementation, metric comparison and calibration below.

  1. Implementation and verification We have prepared an updated implementation with explicit scoring conventions and calibration artifacts linked to the source, weights, templates and scoring policy. Inference loads calibrated thresholds automatically. Verification includes implementation tests, GPU/real-image checks and numerical tests for the added comparison, documented by fixed configurations, reference selections, per-image scores and hashes. Exact regeneration of Tables 1-5 and verification of their original configuration remain incomplete.

  2. Fast templates and score conventions The prepared implementation uses L3 prototype indices 0-29 from the PTS-60 bank, records their source indices, and applies a sigmoid to the raw fast score. PTS planes need not correspond to whole reference photographs.

Changing which prototypes enter the first 30 can affect scores and routing. Permuting an unchanged selected set preserves the mathematical min/max reductions. Subset sensitivity has not been measured. The values 0.58 and 0.68 denote fast-normal and full-normal release; 0.72 is the validation-fixed fast-defect cutoff. Release thresholds require calibration for each frozen scoring policy.

  1. Comparison with simpler statistical metrics TD is a position-conditioned quadratic metric. PCA residual retains only the orthogonal component; full covariance whitening followed by squared Euclidean distance equals Mahalanobis distance. We therefore compare cosine, PCA residual, shrinkage Mahalanobis and TD, without counting whitening as a separate method.

We completed a comparison on all 15 MVTec AD categories and 1,725 official test images. Methods share normalized features, normal-training covariance estimates, 16 fixed normal references per category, an ALL bank, matching neighborhoods and aggregation. Input and pixel evaluation are 256x256. TD retains the industrial weights 0.25/1.0 without test-result retuning. This is a common-bank metric control, distinct from the industrial 640/PTS-60 protocol and an official HETMM reproduction.

Distance               Image macro AUROC   Pixel macro AUROC
Cosine                       86.30%              93.13%
PCA residual                 94.70%              96.46%
Shrinkage Mahalanobis        85.64%              95.55%
TD                           75.49%              84.50%

PCA residual performs best in this setting, so a general advantage for the fixed TD metric is not established. TD's parallel coefficient is 0.25/max(lambda,1e-6), allowing greater weight on low-variance directions. Its intended distinction is position-conditioned directional costs within mutual matching, coupled to calibrated release. We agree that the constituent PCA and risk-calibration tools are established techniques.

  1. Category aggregation and uncertainty Prepared result records include each category's image/pixel AUROC and texture/object averages. The macro averages give categories equal weight. TD minus cosine is -10.815 image-AUROC percentage points, with paired 95% interval [-13.657, -8.185] from 1,000 category-and-label-stratified image bootstrap replicates. This interval conditions on one reference bank and does not measure repeated-reference or pixel-level uncertainty. Pooled AUROC can depend on cross-category score scales; category-macro results address this concern. The new experiment covers MVTec AD; VisA/Real-IAD category-wise reruns remain outstanding.

Category-wise image/pixel AUROC (%) under the control above:

Category          Cosine           TD
bottle       99.60/97.55  91.03/96.95
cable        70.00/92.38  76.26/74.50
capsule      62.90/92.39  77.62/93.43
carpet       99.92/99.04  96.27/98.37
grid         88.39/92.89  56.73/87.26
hazelnut     97.07/96.40  41.36/72.35
leather     100.00/99.44  99.49/99.06
metal_nut    79.08/75.26  48.68/74.93
pill         72.64/86.40  54.45/60.53
screw        56.63/94.45  62.45/77.37
tile         97.69/89.91  92.06/81.23
toothbrush   99.44/98.57  77.22/79.78
transistor   94.21/94.76  73.04/86.46
wood         99.21/93.05  96.93/88.97
zipper       77.73/94.53  88.71/96.25
  1. Calibration with 600 defective samples The calculation is correct: zero misses gives 1/601=0.1664%, below both branch budgets of 0.18% and 0.22%; one miss gives 2/601=0.3328%, above both. Each feasible threshold therefore requires zero calibration misses. The implementation disables an infeasible normal-release branch. Stability across repeated entity-disjoint calibration splits has not been measured. The CRC guarantee concerns expected risk under exchangeability, which process changes can invalidate.
Reviewer_XbTLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRRC-HETMM, a framework for industrial anomaly detection that aims to reduce false positives while controlling missed defects and computational costs. The method integrates position-conditioned texture-disentangled matching, conformal risk calibration, and dynamic inference with fallback to full template matching. Experiments on a proprietary industrial chip dataset demonstrate improvements over HETMM in detection performance and efficiency. Additional evaluations on MVTec AD, VisA, and Real-IAD suggest applicability beyond the specific industrial scenario.

Critical strengths
  1. The paper addresses a practically important problem in industrial anomaly detection, where reducing false positives must be balanced against stringent missed-defect constraints. The integration of texture-aware matching, risk calibration, and dynamic inference provides a coherent framework for this setting.

  2. The reported improvements over HETMM are substantial on the industrial dataset. The authors also provide module ablations, mechanism-oriented experiments, and evaluations on multiple public benchmarks, offering useful empirical evidence of the method's effectiveness.

Critical weaknesses
  1. The proposed texture-disentangled metric operates on features extracted by the backbone rather than directly improving fine-grained feature representations. Its ability to preserve subtle anomaly evidence, particularly when defects overlap with normal texture variations, remains insufficiently validated. Furthermore, the position-conditioned modeling assumes reasonably consistent spatial correspondence, raising my concerns about robustness to misalignment and viewpoint variations.

  2. The experimental analysis on public benchmarks is insufficiently detailed. Although the paper reports improvements on MVTec AD, VisA, and Real-IAD, only pooled AUROC results are provided. Since the primary industrial dataset is proprietary and cannot be independently evaluated, the lack of category-level results makes it difficult to determine whether the improvements generalize consistently across different objects and anomaly characteristics.

Questions for authors
  1. The proposed texture-disentangled metric operates on extracted features, but subtle anomalies may already lose discriminative information during feature extraction, as highlighted by prior work such as AnomalyCLIP (ICLR 2024). Could the authors clarify how the proposed method preserves fine-grained anomaly evidence, particularly for small or low-contrast defects, and whether anomaly responses overlapping with normal-variation subspaces might be suppressed?

  2. Since the main industrial dataset is proprietary, more detailed evaluations on public benchmarks would help establish the generality of the method. Could the authors provide category-wise image-level and pixel-level results on MVTec AD and other public datasets, particularly comparing texture and object categories and their improvements relative to HETMM?

  3. The position-conditioned subspace modeling relies on consistent physical meaning at corresponding spatial locations. Could the authors clarify how robust the method is to spatial misalignment, object displacement, or viewpoint variations, particularly in less controlled inspection settings?

1 comment

Response to Reviewer XbTL

Thank you for the questions on subtle defects, category-level evidence and alignment. We clarify the mechanism and report the newly completed comparison below.

  1. Fine-grained evidence and normal-subspace overlap We agree that a metric applied to frozen features cannot recover information already discarded by the encoder. TD changes how available residuals are compared; it is not a representation-learning solution to the limitation raised through AnomalyCLIP.

TD includes both the orthogonal residual and the variance-normalized retained-subspace component. This does not guarantee preservation of small or low-contrast defects: a defect can align with normal variation, and a truncated normal subspace need not contain all normal variability. The coefficient of a retained direction is w_parallel/max(lambda,1e-6), which can exceed the orthogonal weight. Consequently, the orthogonal complement cannot simply be equated with anomalies, and not every parallel component is suppressed. Dedicated small-defect, contrast and subspace-overlap subgroup experiments have not been completed.

  1. Category-wise public evidence We completed the four-distance comparison on all 15 MVTec AD categories and all 1,725 official test images. Features, 16 fixed normal references per category, an ALL reference bank, bidirectional matching and aggregation are shared. Inputs and pixel evaluation are 256x256; TD uses the fixed industrial weights 0.25/1.0 without test-result retuning. This is a controlled metric comparison, distinct from the original 640/PTS-60 setting and not an official HETMM reproduction.

Image/pixel macro AUROC (%) is 86.30/93.13 for cosine, 94.70/96.46 for PCA residual, 85.64/95.55 for shrinkage Mahalanobis, and 75.49/84.50 for TD. PCA residual performs best in this setting; consistent cross-category superiority of the fixed TD metric is not established.

For the five texture categories, TD versus cosine image/pixel macro AUROC is 88.30/90.98% versus 97.04/94.86%. For the ten object categories it is 69.08/81.26% versus 80.93/92.27%. Complete per-category tables and per-image score records have been prepared.

The TD-minus-cosine image-macro difference is -10.815 percentage points, with paired 95% interval [-13.657, -8.185] from 1,000 category-and-label-stratified image bootstrap replicates. This describes test-sample uncertainty conditional on one fixed reference bank. All ranking scores come from the full branch on every test image. We report equal-category macro averages because pooling can confound within-category discrimination with cross-category score scales. The new experiment covers MVTec AD; it does not add complete VisA/Real-IAD runs or evaluate CRC.

Category-wise image/pixel AUROC (%) under the control above:

Category          Cosine          PCA  Mahalanobis           TD
bottle       99.60/97.55 100.00/98.43  99.84/97.45  91.03/96.95
cable        70.00/92.38  97.11/97.14  91.29/96.62  76.26/74.50
capsule      62.90/92.39  86.20/95.33  84.84/95.03  77.62/93.43
carpet       99.92/99.04  99.80/99.18  99.56/98.90  96.27/98.37
grid         88.39/92.89  96.57/98.03  97.41/97.36  56.73/87.26
hazelnut     97.07/96.40  99.82/97.83  47.79/96.64  41.36/72.35
leather     100.00/99.44 100.00/99.44 100.00/99.10  99.49/99.06
metal_nut    79.08/75.26  98.97/93.16  75.95/94.66  48.68/74.93
pill         72.64/86.40  93.48/93.51  63.58/88.77  54.45/60.53
screw        56.63/94.45  80.36/97.99  67.31/92.69  62.45/77.37
tile         97.69/89.91  97.26/92.94  96.57/93.52  92.06/81.23
toothbrush   99.44/98.57  92.22/99.13  81.39/95.30  77.22/79.78
transistor   94.21/94.76  99.96/96.73  88.38/96.55  73.04/86.46
wood         99.21/93.05  98.95/93.13  98.51/92.79  96.93/88.97
zipper       77.73/94.53  79.78/94.98  92.12/97.92  88.71/96.25
  1. Spatial correspondence and viewpoint Position-conditioned statistics assume sufficiently consistent physical correspondence. Local matching neighborhoods offer limited tolerance; they do not make the position-wise subspace invariant to translation, rotation or viewpoint. The new unperturbed MVTec experiment does not isolate misalignment robustness, and a dedicated perturbation/multi-view study has not been completed.
Reviewer_Em7gOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper discusses a core problem in industrial anomaly detection: how a detector can effectively suppress interference from normal variations while fully preserving the structural residuals that are orthogonal to them. Motivated by this question, the paper introduces a position-constrained texture disentanglement metric, which separates normal texture directions from orthogonal structural residuals, and achieves competitive results on an entity-disjoint industrial chip dataset.

Critical strengths

The paper discusses a very valuable problem in industrial anomaly detection, namely how to suppress interference from normal variations in the detector. The authors' experimental effort is commendable.

Critical weaknesses
  1. In the Introduction, the paper only discusses baseline methods but omits any discussion of existing template-matching approaches (e.g., MuSc), which weakens the motivation of the paper.
  2. It seems that none of the Figures in the paper are cited and discussed in the main text, and only Table 1 is referenced. These figures and tables that are not cited or discussed in the text increase the difficulty of reading and reduce the compactness of the paper's structure.
Questions for authors
  1. The main figures and sub-figures (Fig. 1 and Fig. 2) are both rendered as flowcharts. Although this form is easy to follow, they amount to little more than a concatenation of engineering steps, which fails to convey the novelty of the paper. They are also not sufficiently intuitive or visually pleasing.
  2. As noted in Critical Weaknesses, the paper lacks any discussion of Fig. 3. In Case 1 and Case 3, is the TRRC-HETMM map not consistently better than the HETMM map? From the visualizations, TRRC-HETMM can filter out some interference, but it also reduces the detection of anomalous pixels.
  3. The PRELIMINARIES section is so brief that it could be merged into the METHOD section.
  4. Why are the scores 0.58 and 0.72? Are these two values sensitive to the dataset? Do they always remain valid?
1 comment

Response to Reviewer Em7g

Thank you for the suggestions on motivation, presentation and interpretation. We address them directly below.

  1. Related matching methods We agree that the Introduction should discuss MuSc (https://arxiv.org/abs/2401.16753) and connect existing reference/mutual-matching approaches more clearly to our setting. Our motivation is a fixed inspection station where normal strip texture and genuine strip defects share the valid inspection region, so excluding that region is inappropriate. TD addresses this setting through position-conditioned quadratic reweighting within mutual matching.

  2. Figures, references and organization Some interpretation already appears in the appendix, including the three cases in Figure 3. However, the main-text references and explanations are insufficient. We agree that each figure/table should be introduced and connected to the specific claim it supports.

Figures 1-2 currently emphasize execution order more than the metric's mechanism. The intended mechanism is to separate a query-template residual into retained normal-variation and orthogonal components, apply different directional costs, and use the resulting distance at the query anchor in forward matching and the template anchor in backward matching. A geometric illustration would convey this distinction more clearly. We also agree that merging the short Preliminaries into the Method would improve continuity.

  1. Figure 3, Cases 1 and 3 Case 1 is a normal sample with a strip-texture false positive under HETMM; it illustrates removal of nuisance response, not improved coverage of anomalous pixels. Case 2 illustrates a detected true anomaly. Case 3 was originally presented as a shared failure. Its displayed annotation does not explicitly identify a defect, so that panel cannot establish pixel-level anomaly retention or superiority.

Normal and anomalous residuals can share directions, so reducing nuisance responses can also weaken true-defect responses. The three examples alone cannot establish general pixel-level preservation. Absolute heatmap scores require a shared color scale for comparison.

  1. Why 0.58 and 0.72, and their validity Under the stated protocol, 0.58 is the fast-normal calibration threshold, 0.68 is the full-normal calibration threshold, and 0.72 is a validation-frozen fast-defect routing threshold. Guards precede both fast exits. The fast-defect exit does not directly create missed defects, but its threshold affects false positives and computation. Equality at the fast-normal threshold causes fallback; equality at the full-normal threshold means normal release.

These are dataset- and score-policy-dependent values, not universal constants. Changing features, templates, aggregation or routing requires freezing the new policy and independently recalibrating release thresholds.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.