Abstract

1. https://github.com/wogur110/PNI_Anomaly_Detection，based on PatchCore & better
2. Ranked #4 on Anomaly Detection on BTAD (using extra training data)
3. PNI Ensemble is ranked #7 on MVTec AD ****(using extra training data)

1. cables are normal only when the color order is correct,
2. if the color order is incorrect, the product is defective, even though all the local features are normal.
3. PatchCore: image-level memory bank of neighbourhood-aware patch-level features 如何理解左下anomaly的粗的地方score高？

Introduction

baseline: PatchCore.

Proposed three components:

1. Neighbor
   1. train a MLP to model the probability distribution of normal features conditioned by neighboring features
      1. the input is the concatenated neighboring features
      2. MLP observes a large support region, while the features remain local, allowing the proposed method to produce a finely detailed localization map.
      3. 和大patch有啥区别？
2. Position
   1. Construct a histogram of all the training features to model a conditional probability distribution given the positional information.
3. These two distributions are combined to estimate the likelihood and anomaly score
4. Anomaly map refinement
   1. Instead of simply resizing the anomaly map ⇒ an additional refine network trained on synthetic anomaly images to better interpolate and account for the shape and edge of the input image.

1. Neighbor: 98.92% to 99.44%
2. Position: provides an additional gain in object subcategories, but there is little improvement observed in texture subcategories.
3. Refine: more effective in texture subcategories and has complementary properties with position information