Concept Distribution Modeling for Few-Shot Anomaly Detection

02 — Approach

Method

The workflow below is a visual summary of the documented implementation; it is not a claim of additional system components.

1. 01

   Engineered a novel concept-based distribution framework that semantically divides industrial images into unique visual prototypes using k-means clustering and a saliency-assisted voting mechanism

   →
2. 02

   Developed a robust reference bank of normal prototypes by modeling multivariate Gaussian distributions of deep features, significantly reducing the dependency on large volumes of training data

   →
3. 03

   Achieved State-of-the-Art (SOTA) performance on the MVTec AD and ViSA benchmarks, securing 96.3% and 97.1% pAUROC respectively in 1-shot scenarios while outperforming existing methods in F1-max and PRO scores

03 — Contribution

Contribution

• 01

  Developed a robust reference bank of normal prototypes by modeling multivariate Gaussian distributions of deep features, significantly reducing the dependency on large volumes of training data

• 02

  Achieved State-of-the-Art (SOTA) performance on the MVTec AD and ViSA benchmarks, securing 96.3% and 97.1% pAUROC respectively in 1-shot scenarios while outperforming existing methods in F1-max and PRO scores

• 03

  Optimized anomaly localization precision by implementing Mahalanobis distance scoring for test patches, enabling the model to effectively discern defects in small and complex objects without global feature reliance

04 — Evidence

Results