Granted Patent B2

US 12,505,526 · App. 18/054,524 · Granted Dec 23, 2025

Inventors: Kihyuk Sohn (Mountain View, CA); Jinsung Yoon (San Jose, CA); Chun-Liang Li (Mountain View, CA); Tomas Jon Pfister (Foster City, CA); Chen-Yu Lee (Mountain View, CA)

Assignee: GOOGLE LLC

G06T7/0004G06V10/7625G06V10/7635G06T2207/20081G06T2207/20084G06V10/764G06V10/82

Abstract

Claims (40)

1 . A computer-implemented method when executed by data processing hardware causes the data processing hardware to perform operations comprising:

receiving an anomaly clustering request requesting the data processing hardware assign each image of a plurality of images into one of a plurality of groups;

obtaining the plurality of images; and

for each respective image of the plurality of images:

extracting, using a trained model, a respective set of patch embeddings from the respective image;

determining a distance between the respective set of patch embeddings and each other set of patch embeddings;

assigning, using the distances between the respective set of patch embeddings and each other set of patch embeddings, the respective image into one of the plurality of groups; and

determining a weight vector for the respective set of patch embeddings, the weight vector comprising a weight for each patch embedding in the respective set of patch embeddings, each weight indicating a likelihood the patch embedding includes an anomaly.

2 . The method of claim 1 , wherein assigning the respective image into one of the plurality of groups comprises applying a distance-based clustering technique.

3 . The method of claim 2 , wherein the distance-based clustering technique comprises one of hierarchical clustering or spectral clustering.

4 . The method of claim 1 , wherein each image of the plurality of images is unlabeled or labeled.

5 . The method of claim 1 , wherein determining the distance between the respective set of patch embeddings and each other set of patch embeddings comprises determining a weighted average distance between each set of patch embeddings using the weight vectors corresponding to the respective set of patch embeddings and each other set of patch embeddings.

6 . The method of claim 1 , wherein the distance between the respective set of patch embeddings and each other set of patch embeddings comprises a Euclidean distance.

7 . The method of claim 1 , wherein determining the distance between the respective set of patch embeddings and each other set of patch embeddings comprises using an unsupervised model.

8 . The method of claim 1 , wherein determining the distance between the respective set of patch embeddings and each other set of patch embeddings comprises using a semi-supervised model.

9 . The method of claim 1 , wherein the plurality of groups comprises at least one of:

a normalcy group representing images without any anomalies;

a first anomaly group representing images that comprise a first manufacturing defect; and

a second anomaly group representing images that comprise a second manufacturing defect.

10 . A system comprising:

data processing hardware; and

memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:

receiving an anomaly clustering request requesting the data processing hardware assign each image of a plurality of images into one of a plurality of groups;

obtaining the plurality of images; and

for each respective image of the plurality of images:

extracting, using a trained model, a respective set of patch embeddings from the respective image;

determining a distance between the respective set of patch embeddings and each other set of patch embeddings;

assigning, using the distances between the respective set of patch embeddings and each other set of patch embeddings, the respective image into one of the plurality of groups; and

determining a weight vector for the respective set of patch embeddings, the weight vector comprising a weight for each patch embedding in the respective set of patch embeddings, each weight indicating a likelihood the patch embedding includes an anomaly.

11 . The system of claim 10 , wherein assigning the respective image into one of the plurality of groups comprises applying a distance-based clustering technique.

12 . The system of claim 11 , wherein the distance-based clustering technique comprises one of hierarchical clustering or spectral clustering.

13 . The system of claim 10 , wherein each image of the plurality of images is unlabeled or labeled.

14 . The system of claim 10 , wherein determining the distance between the respective set of patch embeddings and each other set of patch embeddings comprises determining a weighted average distance between each set of patch embeddings using the weight vectors corresponding to the respective set of patch embeddings and each other set of patch embeddings.

15 . The system of claim 10 , wherein the distance between the respective set of patch embeddings and each other set of patch embeddings comprises a Euclidean distance.

16 . The system of claim 10 , wherein determining the distance between the respective set of patch embeddings and each other set of patch embeddings comprises using an unsupervised model.

17 . The system of claim 10 , wherein determining the distance between the respective set of patch embeddings and each other set of patch embeddings comprises using a semi-supervised model.

18 . The system of claim 10 , wherein the plurality of groups comprises at least one of:

a normalcy group representing images without any anomalies;

a first anomaly group representing images that comprise a first manufacturing defect; and

a second anomaly group representing images that comprise a second manufacturing defect.

Assignments (1)

Continuity (2)

References Cited (20)