PatchCore
PatchCore offers competitive inference time while achieving state-of-the-art performance for both anomaly detection and localization.
PatchCore is presented as an effective method for..
- Maximizing nominal information available at test time
- Reducing biased towards ImageNet classes by using mid-level network patch features.
- Retaining high inference speed by applying “Coreset” selection to reduce the size of the “memory banks” where all the features are stored.
“Coreset” Selection: Aims to find a subset S ⊂ A such that problem solutions over A can be most closely and especially more quickly approximated by those computed over S
The overall architecture is the same as PaDiM.
What Previous Work Couldn’t Do
SPADE:
Similarly used a “memory bank”, but did not have neighborhood-aware patch-level features critical to achieving higher performance, as more nominal context is retained and a better fitting inductive bias is incorporated.
PaDiM:
Makes use of an efficient patch-feature “memory bank” but it is limited due to using the Mahalanobis distance measures specific to each patch. This makes the model more reliant on alignment of the patches in the image.
Reference:Towards Total Recall in Industrial Anomaly Detection
