Focus and Supplement: Dual-Enhanced Vision Transformer for Multi-Class Anomaly Classification

1 Introduction

• Industrial anomaly classification distinguishes fine-grained anomaly classes, supporting applications such as root-cause analysis and process optimization.The task is especially challenging when anomaly types are unknown in real-world production.Source: §1 · p. 1
• MACO addresses noisy and incomplete anomaly representations and unknown class counts with enhanced features and a Correlation-based Number Estimation strategy.The strategy computes average correlations among labeled classes and transfers their separability cue to the unlabeled set.Source: §1 · p. 1
• 6.5% and 16.3% ARI improvements are reported with a known class number, respectively, and the unknown-number setting yields an 11.2% NMI gain on MTD.For class-number estimation, MACO improves UPS by 24.1% and 40% on the two datasets.Source: §1 · p. 2
• DE-ViT combines soft-focused attention with Auxiliary Classification Tokens to extract discriminative, comprehensive anomaly representations.The tokens collectively attend to diverse abnormal sub-regions, complementing the [CLS] token.Source: §1 · p. 2

2.1 Multi-class Anomaly Classification

• Clustering approaches use anomaly-weighted feature aggregation or select high-scoring patch features, while NCD methods have shifted toward one-stage use of labeled and unlabeled data.Industrial approaches include mask-guided attention, which requires binarizing anomaly maps and can propagate missed detections into misclassification.Source: Clustering-based methods · p. 2, NCD-based methods · p. 2

2.2 [CLS] token of vision transformer

• Auxiliary Classification Tokens supplement the [CLS] token with anomaly-specific cues to represent diverse anomaly patterns more comprehensively.The cited discussion says [CLS] can overfit dominant features and overlook subtle inter-anomaly distinctions.Source: §2.2 · p. 2

2.3 Estimating the class number

• Existing class-number estimators use clustering metrics, split-and-merge procedures, candidate refinement, protoscores, or labeled-class accuracy, often at substantial computational cost.MACO instead uses labeled-class prototypes to estimate inter-prototype gaps among unlabeled classes.Source: §2.3 · p. 2

3 Preliminary

• The task categorizes unlabeled images into novel classes, under settings where the novel-class count is known a priori or unknown.The setup uses labeled anomaly images and an unlabeled set containing novel anomalies.Source: Problem Definition · p. 2
• The described NCD framework combines supervised and self-supervised contrastive learning with ground-truth and pseudo-label classification losses, plus mean-entropy regularization.A teacher generates pseudo-labels for unlabeled images, while the student classifies all images.Source: Novel Class Discovery · pp. 2–3
• ViT self-attention computes dependencies among patch tokens through an attention map.The framework pipeline then extracts anomaly features with DE-ViT, classifies them using a linear classifier with Orthogonal Weight Loss, and estimates an unavailable class count with Correlation-based Number Estimation.Source: Self-attention · p. 3, Figure 2 · p. 3

4 Method

• The framework combines DE-ViT anomaly features, orthogonal weight loss, and correlation-based class-number estimation for multi-class anomaly classification.The Under-estimation Penalized Score evaluates estimated class counts with asymmetric penalties.Source: §4 · p. 3

4.1 Soft-focused Attention

• Soft-focused attention uses anomaly maps to guide the [CLS] token toward likely anomaly regions, reducing background interference without discarding lower-probability regions.Unlike binary mask-guided attention, continuous maps can retain regions with potential anomalies.Source: §4.1 · pp. 3–4

4.2 Auxiliary Classification Tokens

• Auxiliary [A-CLS] tokens complement [CLS] by attending to different anomaly regions, including missed regions, producing more complete representations.[ac1] tokens use anomaly-map guidance, while unguided [ac2] tokens can discover regions missed by the map.Source: §4.2 · pp. 4–5

4.3 Orthogonal Weight Loss

• Orthogonal weight loss constrains classifier prototypes to remain separated in feature space, encouraging inter-class separation from a prototype perspective.The loss operates on classifier parameters; W represents prototypes for labeled and unlabeled classes.Source: §4.3 · p. 5

4.4 Correlation-based Number Estimation

• Correlation-based Number Estimation transfers the average labeled-prototype correlation threshold to unlabeled prototypes to dynamically refine an initially overestimated class count.It selects prototypes by minimum maximum similarity to those already selected, stopping when that similarity exceeds the labeled-class threshold.Source: §4.4 · p. 5

4.5 Under-estimating Penalized Score

• The Under-estimation Penalized Score assigns asymmetric penalties because underestimating classes can merge distinct anomaly types, while overestimation mainly increases manual verification.The desired score is 1 for an exact count, rises monotonically as an underestimate approaches the true count, and decreases monotonically for overestimates.Source: §4.5 · p. 6

5.1 Experimental Setting

• Experiments use MVTec AD and MTD, report NMI/ARI and F1, assess unknown-count estimates with MAPE/UPS, and use ViT-B/8 pretrained with DINO at 224 × 224 resolution.Source: Datasets · p. 6, Evaluation protocol · p. 6, Implementation details · p. 6

5.2 Comparison with the State-of-the-Arts

• 24.1% UPS improvement over the best prior method on MVTec AD; on MTD, MACO achieves MAPE=0.000 and UPS=1.000, with best MAPE and UPS on both datasets.Source: Table 2 · p. 6, Unlabeled classes number Cu is unknown · p. 7
• 6.5% ARI gain on MVTec AD and 16.3% on MTD accompany MACO’s higher scores than both industrial-specific approaches across all metrics.The MTD dataset has noisier anomaly maps.Source: Unlabeled classes number Cu is known · p. 6

5.3 Ablation Study

• Against mask-guided attention, soft-focus reduces binarization-error propagation; missed-detection rates reach 16.02% on MVTec AD and 40.46% on MTD, while ARI increases 2.6% and 18.3%, respectively.Source: Discussion of the soft-focus attention · p. 7
• Using both [ac1] and [ac2] yields the best overall performance across most metrics, though [ac2] alone slightly improves some metrics when anomaly maps misguide [ac1].The tokens are also reported as applicable to other visual and non-visual tasks.Source: Table 4 · p. 7, Effectiveness of [A-CLS] tokens · p. 7
• Soft-focused attention shifts [CLS] toward true-positive regions; guided [ca1] covers other true positives, while unguided [ca2] attends more to false negatives.The authors link this complementary coverage to visible and missed anomalies.Source: Table 5 · p. 7, Discussion of anomaly attention… · p. 7
• The [A-CLS] tokens are also reported as applicable to other visual and non-visual tasks.Source: Effectiveness of [A-CLS] tokens · p. 7
• On MTD, Lac improves F1 by 0.5% and Low raises ARI by 1.1%, while Low decreases NMI by 0.2%.The authors attribute the NMI decrease to reduced clustering flexibility when visually similar anomaly classes overlap.Source: Table 6 · p. 7, Effect of Lac and Low · p. 7

• MACO addresses noisy and incomplete anomaly representations and unknown class counts using soft-focus attention, Auxiliary Classification Tokens, and Correlation-based Number Estimation.Source: §6 · p. 7