Method and apparatus for recognizing image and method and apparatus for training recognition model based on data augmentation

Patent metadata

Field	Value
Publication number	US-10957073-B2
Application number	US-201916264752-A
Country	US
Kind code	B2
Filing date	Feb 1, 2019
Priority date	Aug 23, 2018
Publication date	Mar 23, 2021
Grant date	Mar 23, 2021

Abstract

An image recognition method includes: selecting an augmentation process from augmentation processes based on a probability table, in response to an acquisition of an input image; acquiring an augmented image by augmenting the input image based on the selected augmentation process; and recognizing an object from the augmented image based on a recognition model.

First claim

What is claimed is:

1. An image recognition method, comprising:
analyzing a characteristic of an input image, in response to an acquisition of the input image;
selecting a probability table from probability tables, in response to the probability table corresponding to the analyzed characteristic;
selecting an augmentation process from augmentation processes based on the selected probability table;
acquiring an augmented image by augmenting the input image based on the selected augmentation process; and
recognizing an object from the augmented image based on a recognition model.

2. The image recognition method of claim 1 , wherein the selecting of the augmentation process comprises determining a parameter of the selected augmentation process based on another probability table.

3. The image recognition method of claim 1 , further comprising:
receiving a user input designating the probability table; and
selecting the probability table from a plurality of probability tables, in response to a reception of the user input.

4. The image recognition method of claim 1 , further comprising:
collecting environmental information about a surrounding environment of an image recognition apparatus; and
selecting the probability table from a plurality of probability tables, in response to the probability table corresponding to the environmental information.

5. The image recognition method of claim 1 , further comprising:
updating the probability table based on an output calculated from the augmented image, using the recognition model.

6. The image recognition method of claim 5 , wherein the updating of the probability table comprises:
calculating a contribution score of each of the augmentation processes based on the output calculated from the augmented image, using the recognition model; and
adjusting a probability that designates each of the augmentation processes in the probability table based on the contribution score.

7. The image recognition method of claim 6 , wherein the updating of the probability table comprises calculating contribution scores for each parameter of an individual augmentation process among the augmentation processes.

8. The image recognition method of claim 6 , wherein
the calculating of the contribution score comprises
calculating, from the augmented image, a contribution score of an augmentation process, among the augmentation processes, selected every time an object is recognized, and
generating a contribution histogram by classifying and accumulating contribution scores of the selected augmentation processes based on the selected augmentation processes, and
the adjusting of the probability comprises adjusting the probability based on the contribution histogram.

9. A recognition model training method, comprising:
analyzing a characteristic of an input data, in response to an acquisition of the input data;
selecting a probability table from probability tables, in response to the probability table corresponding to the analyzed characteristic;
selecting an augmentation process, among augmentation processes, based on the selected probability table;
acquiring augmented data by augmenting the input data based on the selected augmentation process, in response to the selecting of the augmentation process; and
training a recognition model based on the augmented data.

10. The recognition model training method of claim 9 , wherein the selecting of the augmentation process comprises
determining a parameter of the selected augmentation process based on another probability table.

11. The recognition model training method of claim 9 , further comprising:
randomly selecting a new augmentation process based on the probability table, in response to an augmentation count of the input data being less than a threshold count;
acquiring new augmented data by augmenting the input data based on the selected new augmentation process; and
retraining the recognition model based on the new augmented data.

12. The recognition model training method of claim 9 , wherein the selecting of the augmentation process comprises
acquiring n pieces of input data, and
selecting the augmentation process based on the probability table for each of the n pieces of input data, and
wherein n is an integer greater than or equal to 1.

13. The recognition model training method of claim 9 , further comprising:
updating the probability table based on an output calculated from the augmented data, using the recognition model.

14. The recognition model training method of claim 13 , wherein the updating of the probability table comprises
calculating a contribution score of each augmentation process based on the output calculated from the augmented data, using the recognition model, and
adjusting a probability that designates each augmentation process in the probability table based on the contribution score.

15. The recognition model training method of claim 14 , wherein the updating of the probability table comprises
calculating contribution scores for each augmentation process, and
calculating contribution scores for each parameter of an individual augmentation process among the augmentation processes.

16. The recognition model training method of claim 14 , wherein the calculating of the contribution score comprises
calculating, from the augmented data, a contribution score of an augmentation process, among the augmentation processes, selected for each training, and
generating a contribution histogram by classifying and accumulating contribution scores of the selected augmentation processes based on the selected augmentation processes, and
wherein the adjusting of the probability comprises adjusting the probability based on the contribution histogram.

17. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1 .

18. A recognition model training apparatus, comprising:
a data acquirer, including an image sensor, configured to acquire input data; and
a processor configured to analyze a characteristic of an input image in response to an acquisition of the input image, to select a probability table from probability tables in response to the probability table corresponding to the analyzed characteristic, to select an augmentation process based on the selected probability table, to acquire augmented data by augmenting the input data based on the selected augmentation process, in response to the selection of the augmentation process, and to train a recognition model based on the augmented data.

Assignees

• Samsung Electronics Co Ltd

Inventors

• Sung Jaemo
• Kim Chang Hyun

Classifications

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06V10/7796

  based on specific statistical tests · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/30

  Noise filtering · CPC title

• G06V10/82

  using neural networks · CPC title