Image segmentation into overlapping tiles

Patent metadata

Field	Value
Publication number	US-10607331-B1
Application number	US-201916456733-A
Country	US
Kind code	B1
Filing date	Jun 28, 2019
Priority date	Jun 28, 2019
Publication date	Mar 31, 2020
Grant date	Mar 31, 2020

Abstract

Systems and techniques for image segmentation into overlapping tiles are described herein. In an example, an anomaly detection system for super-high-resolution images is adapted to divide the image into tiles with a size and overlap between each tile, wherein the size and overlap is determined using a machine-learning model. The anomaly detection system may be further adapted to use a classifier model for each tile to identify anomaly presence in the tile, wherein the classifier model is trained using a data set of tile size images that are labeled according to anomaly presence in each tile size image. The anomaly detection system may be further adapted to determine a classification for the image based on results from the classifier model for the tiles. The anomaly detection system may be further adapted to output the classification.

First claim

What is claimed is:

1. A system for image analysis, comprising:
at least one processor; and
memory including instructions that, when executed by the at least one processor, cause the at least one processor to:
receive an image of a product;
divide the image into tiles with a size and overlap between each tile, wherein the size and overlap is determined using a machine-learning model trained using a range of tile sizes and overlap and is based on accuracy of anomaly detection for each tile size and overlap;
use a classifier model for each tile to identify anomaly presence in the tile, wherein the classifier model is trained using a data set of tile size images that are labeled according to anomaly presence in each tile size image;
determine a classification for the image based on results from the classifier model for the tiles of the image; and
output the classification for the image to a graphical user interface.

2. The system of claim 1 , wherein the image is classified as passing when the anomaly presence is not identified for each of the tiles of the image and wherein the image is classified as failing when the anomaly presence is identified for at least one of the tiles of the image.

3. The system of claim 1 , wherein the image is a super-high-resolution image with a resolution of at least 5000 by 5000 pixels.

4. The system of claim 1 , wherein the range of tile sizes and overlap for the machine-learning model is based on one of random search, grid search, or Bayesian optimization.

5. The system of claim 1 , wherein the data set of tile size images is generated by dividing a sample set of images into tiles and labeling each tile based on anomaly presence within the tile, and wherein size and overlap of the data set of tile size images is based on results of the machine-learning model.

6. The system of claim 5 , wherein the sample set of images includes images with a plurality of anomalies, the plurality of anomalies being of different types and sizes.

7. The system of claim 1 , wherein the machine learning model identifies an optimal tile size and overlap based on a convergence of accuracy for tile size and overlap parameters.

8. A method for image analysis, comprising:
receiving an image of a product;
dividing the image into tiles with a size and overlap between each tile, wherein the size and overlap is determined using a machine-learning model trained using a range of tile sizes and overlap and is based on accuracy of anomaly detection for each tile size and overlap;
using a classifier model for each tile to identify anomaly presence in the tile, wherein the classifier model is trained using a data set of tile size images that are labeled according to anomaly presence in each tile size image;
determining a classification for the image based on results from the classifier model for the tiles of the image; and
outputting the classification for the image to a graphical user interface.

9. The method of claim 8 , wherein the image is classified as passing when the anomaly presence is not identified for each of the tiles of the image and wherein the image is classified as failing when the anomaly presence is identified for at least one of the tiles of the image.

10. The method of claim 8 , wherein the image is a super-high-resolution image with a resolution of at least 5000 by 5000 pixels.

11. The method of claim 8 , wherein the range of tile sizes and overlap for the machine-learning model is based on one of random search, grid search, or Bayesian optimization.

12. The method of claim 8 , wherein the data set of tile size images is generated by dividing a sample set of images into tiles and labeling each tile based on anomaly presence within the tile, and wherein size and overlap of the data set of tile size images is based on results of the machine-learning model.

13. The method of claim 12 , wherein the sample set of images includes images with a plurality of anomalies, the plurality of anomalies being of different types and sizes.

14. The method of claim 8 , wherein the machine learning model identifies an optimal tile size and overlap based on a convergence of accuracy for tile size and overlap parameters.

15. At least one non-transitory computer readable medium including instructions for image analysis that when executed by at least one processor, cause the at least one processor to:
receive an image of a product;
divide the image into tiles with a size and overlap between each tile, wherein the size and overlap is determined using a machine-learning model trained using a range of tile sizes and overlap and is based on accuracy of anomaly detection for each tile size and overlap;
use a classifier model for each tile to identify anomaly presence in the tile, wherein the classifier model is trained using a data set of tile size images that are labeled according to anomaly presence in each tile size image;
determine a classification for the image based on results from the classifier model for the tiles of the image; and
output the classification for the image to a graphical user interface.

16. The at least one non-transitory computer readable medium of claim 15 , wherein the image is classified as passing when the anomaly presence is not identified for each of the tiles of the image and wherein the image is classified as failing when the anomaly presence is identified for at least one of the tiles of the image.

17. The at least one non-transitory computer readable medium of claim 15 , wherein the range of tile sizes and overlap for the machine-learning model is based on one of random search, grid search, or Bayesian optimization.

18. The at least one non-transitory computer readable medium of claim 15 , wherein the data set of tile size images is generated by dividing a sample set of images into tiles and labeling each tile based on anomaly presence within the tile, and wherein size and overlap of the data set of tile size images is based on results of the machine-learning model.

19. The at least one non-transitory computer readable medium of claim 18 , wherein the sample set of images includes images with a plurality of anomalies, the plurality of anomalies being of different types and sizes.

20. The at least one non-transitory computer readable medium of claim 15 , wherein the machine learning model identifies an optimal tile size and overlap based on a convergence of accuracy for tile size and overlap parameters.

Assignees

• Corning Inc

Inventors

• Tandia Adama
• Zhang Yi

Classifications

• G06T7/0004 Primary

  Industrial image inspection · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20021

  Dividing image into blocks, subimages or windows · CPC title

• G06T7/11

  Region-based segmentation · CPC title