Life sciences · Preprint

Spatial Attention Supervision for Defect Localization: Exploiting Ground-Truth Masks as Training Signal in Diffusion-Augmented Defect Detection

arXiv · September 5, 2026

Preliminary Why this matters to you Expert discussion

Early or partial results. Treat as a signal, not a conclusion.

AI Summary

This preprint proposes using ground-truth defect masks as spatial supervision signals during training to improve defect localization in convolutional networks, evaluated on the MVTec-AD bottle benchmark. Attention-guided training improved pixel-level localization (Pixel-AUROC) by 18–18.7% in some models but showed architecture-dependent effects, with no improvement in ConvNeXt-T and no significant change in classification accuracy, while an unsupervised baseline (PatchCore, Pixel-AUROC=0.983) remained superior.

What was studied

Factorial design, computational method comparison across multiple CNN architectures and training configurations.. MVTec-AD bottle class synthetic images; subset with ground-truth defect masks for training supervision, subset without masks (diffusion-generated) for augmentation compatibility. No real defect instances or industrial inspection pipeline tested.. Intervention: Attention-guided training with activation-based attention alignment loss using ground-truth masks as spatial supervision signals, combined with DDPM synthetic augmentation.. Compared with: Standard classification training without attention guidance; mixed-supervision without augmentation; unsupervised PatchCore baseline..

Key findings

Attention-guided training improved Pixel-AUROC by +18.0% for EfficientNetB0 with augmentation (p=0.005, Cohen's d=2.6) and +18.7% for ResNet50 (p=0.008). Significant improvement in 4 of 8 CNN settings (uncorrected for multiple comparisons) with no significant change in classification accuracy. Data × training-mode interaction significant for EfficientNetB0 (p=0.002), suggesting super-additive effect (+13.6% combined vs +1.6% summed individual effects).

Safety

Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.

Who this may help

The source did not state who this applies to in practice.

Plain-language summary

A single-benchmark computational study on synthetic defect data with modest improvements in a surrogate endpoint (localization), unreviewed preprint without peer-reviewed publication, limited to one industrial dataset and showing architecture-dependent effects.

Study design

As stated by the source record.

MethodsEvaluated 85 models in a 2×2 factorial design (four CNN backbones × presence/absence of augmentation × presence/absence of attention training) over five random seeds, plus Swin-V2-T baseline. Attention-guided training applied activation-based attention alignment loss steering convolutional feature maps toward defect regions in mixed-supervision formulation accommodating both annotated and diffusion-generated (DDPM) images. Localization (Pixel-AUROC) measured on held-out defect images excluded from classifier gradient updates; classification accuracy recorded as secondary outcome.

DesignFactorial design, computational method comparison across multiple CNN architectures and training configurations.

PopulationMVTec-AD bottle class synthetic images; subset with ground-truth defect masks for training supervision, subset without masks (diffusion-generated) for augmentation compatibility. No real defect instances or industrial inspection pipeline tested.

InterventionAttention-guided training with activation-based attention alignment loss using ground-truth masks as spatial supervision signals, combined with DDPM synthetic augmentation.

ComparatorStandard classification training without attention guidance; mixed-supervision without augmentation; unsupervised PatchCore baseline.

Reported figures

Quoted from the source exactly as published.

Pixel-AUROC improvement (EfficientNetB0 with augmentation)+18.0%

p-value (EfficientNetB0)0.005

Cohen's d (EfficientNetB0)2.6

Pixel-AUROC improvement (ResNet50)+18.7%

p-value (ResNet50)0.008

Significant settings (CNN models)4 of 8

Data × training-mode interaction p-value (EfficientNetB0)0.002

Super-additive combined effect (EfficientNetB0)+13.6%

Caveats

Single-benchmark evaluation (MVTec-AD bottle only); generalizability to other defect types or datasets unknown.
Surrogate endpoint (pixel-level localization); no assessment of industrial inspection accuracy, false-positive rates, or deployment impact.
Multiple comparisons not corrected; improvements inconsistent across architectures (ConvNeXt-T null effect), limiting reproducibility claims.
Unsupervised PatchCore baseline (AUROC=0.983) substantially outperforms proposed supervised method, unclear practical advantage.

Evidence assessment

Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.

Study design

moderate

Factorial design with multiple backbones, seeds, and augmentation conditions; single benchmark limits scope.

Directness

limited

Surrogate endpoint (Pixel-AUROC for localization); no clinical or real-world deployment outcomes measured.

Consistency

limited

Improvements significant in 4 of 8 CNN settings; ConvNeXt-T shows no effect; inconsistency across architectures.

Risk of bias

limited

Multiple comparisons uncorrected; held-out test set appropriate but single-dataset validation; preprint status.

Sample adequacy

moderate

85 models evaluated over 5 seeds on MVTec-AD; only bottle class tested; synthetic augmentation may not represent real defects.

Source abstract

Ground-truth defect masks in industrial inspection datasets are typically reserved for evaluation. This paper repurposes them as spatial supervision signals during training of classification networks, teaching a model not just what to predict but where to look. The method adds an activation-based attention alignment loss that steers convolutional feature maps toward defect regions, in a mixed-supervision formulation that also accommodates samples without masks, such as diffusion-generated images. Combined with DDPM augmentation, synthetic images contribute quantity while masks contribute spatial precision. We evaluate 85 models (four CNN backbones under a 2x2 data/training factorial over five seeds, plus a Swin-V2-T transformer baseline) on the MVTec-AD bottle benchmark, with localization measured on held-out defect images excluded from classifier gradient updates. Main findings: (1) attention-guided training improves activation-based localization (Pixel-AUROC) by +18.0% for EfficientNetB0 with augmentation (p=0.005, Cohen's d=2.6) and +18.7% for ResNet50 (p=0.008), significant in four of eight CNN settings (uncorrected for multiple comparisons) with no significant change in classification; (2) for EfficientNetB0 a data x training-mode interaction is significant (p=0.002), consistent with a super-additive effect (+13.6% combined vs +1.6% summed individual effects); (3) architectures with weaker spatial representations benefit most, whereas ConvNeXt-T shows no effect, apparently because its depthwise-convolution activations yield spatially uninformative channel-mean maps; (4) unsupervised PatchCore remains the strongest localizer (Pixel-AUROC=0.983), contextualizing the supervised gains. These results show that existing evaluation masks can act as practical training signals that measurably and reproducibly improve where defect classifiers attend.

Follow from this

Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.

Hossein AbediGnana Tulasi MakineniMuskan SarafTom BushSajjad Rezvani Boroujeni