Dataset Augmentation Based on Occlusion and Inpainting

• US 20150379422A1
• Filed: 06/30/2014
• Published: 12/31/2015
• Est. Priority Date: 06/30/2014
• Status: Active Grant

• Current Assignee: Hewlett Packard Enterprise Development LP (Hewlett-Packard Enterprise Company)
• Original Assignee: Hewlett-Packard Development Company, L.P. (HP Inc.)
• Inventors: Chandler, Benjamin Orth
• Granted Patent: US 11,017,311 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Augmenting a dataset in a machine learning classifier is disclosed. One example is a system including a training dataset with at least one training data, and a label preserving transformation including an occluder, and an inpainter. The occluder occludes a selected portion of the at least one training data. The inpainter inpaints the occluded portion of the at least one training data, where the inpainting is based on data from a portion different from the occluded portion. In one example, the augmented dataset is deployed to train a machine learning classifier.

Claims

1. A system comprising:

a training dataset for machine learning, the dataset including at least one training data; and

a label preserving transformation to augment the dataset, the label preserving transformation comprising;

an occluder to occlude, via the processing system, a selected portion of the at least one training data, and

an inpainter to inpaint, via the processing system, the occluded portion of the at least one training data, wherein the inpainting is based on data from a portion different from the occluded portion.

2. The system of claim 1, further comprising a deployer to deploy, via the processing system, the augmented dataset to train a machine learning classifier.

3. The system of claim 2, further comprising inferring, based on the trained machine learning classifier, a label for unlabeled data.

4. The system of claim 2, wherein the machine learning classifier is a deep learning classifier.

5. The system of claim 1, wherein the dataset is a dataset of images, and the at least one training data is an image, and the selected portion of the image is a plurality of pixels in the image.

6. The system of claim 1, wherein the portion of the at least one training data is selected randomly.

7. The system of claim 1, wherein the label preserving transformation is applied N times to the at least one training data to provide N copies of the at least one training data.

8. A method to augment a dataset, the method comprising:

receiving, via a processor, a training dataset for machine learning, the dataset including at least one training data

applying, via the processor, a label preserving transformation to augment the dataset, the label preserving transformation comprising;

occluding, via the processor, a selected portion of the at least one training data, and

inpainting, via the processor, the occluded portion of the at least one training data, wherein the inpainting is based on data from a portion different from the occluded portion; and

deploying, via the processor, the augmented dataset to train a machine learning classifier.

9. The method of claim 8, wherein the machine learning classifier is a deep learning classifier.

10. The method of claim 8, wherein the dataset is a dataset of images, and the at least one training data is an image, and the selected portion of the image is a plurality of pixels in the image.

11. The method of claim 10, further comprising:

receiving, via the processor, another image for image processing;

identifying, based on the trained machine learning classifier, at least ne occlusion in the another image; and

inpainting, based on the trained machine learning classifier, the at least one occlusion.

12. The method of claim 8, wherein the portion of the at least one training data is selected randomly.

13. The method of claim 8, wherein the label preserving transformation is applied N times to the at least one training data to provide N copies of the at least one training data.

14. A non-transitory computer readable medium comprising executable instructions to:

receive, via a processor, a training dataset for machine learning, the dataset including at least one training data; and

apply, via the processor, a label preserving transformation to augment the dataset, the instructions to apply the label preserving transformation comprising further instructions to;

occlude a selected portion of the at least one training data, and

inpaint the occluded portion of the at least one training data, wherein the inpainting is based on data from a portion different from the occluded portion.

15. The non-transitory computer readable medium of claim 14, further comprising executable instructions to deploy, via the processor, the augmented dataset to train a machine learning classifier.