Tagging similar images using neural network

Patent metadata

Field	Value
Publication number	US-9740966-B1
Application number	US-201615016607-A
Country	US
Kind code	B1
Filing date	Feb 5, 2016
Priority date	Feb 5, 2016
Publication date	Aug 22, 2017
Grant date	Aug 22, 2017

Abstract

An approach is provided in which a knowledge manager selects an extraction layer from a convolutional neural network that was trained on an initial set of images. The knowledge manager processes subsequent images obtained from crawling a computer network that includes extracting image feature sets of the subsequent images from the selected extraction layer and generating tags from metadata associated with the subsequent images. In turn, the knowledge manager receives a new image, extracts a new image feature set from the selected extraction layer, and assigns one or more of the tags to the new image based upon evaluating the new image feature set to the image features sets of the subsequent images.

First claim

The invention claimed is:

1. A method implemented by an information handling system that includes a memory and a processor, the method comprising:
training a classifier on a first set of images, wherein the classifier comprises a convolutional neural network that includes a plurality of layers;
processing a second set of images using the trained classifier, wherein the processing comprises:
extracting a plurality of feature sets from a selected layer included in the plurality of layers; and
generating a plurality of tags based upon metadata corresponding to the second set of images;

selecting a subset of the plurality of tags based on performing a nearest neighbors search on a new image feature set corresponding to a new image, wherein the new image feature set is extracted from the selected layer during processing of the new image; and
assigning at least one of the subset of the plurality of tags to the new image based upon voting on the subset of the plurality of tags.

2. The method of claim 1 wherein the selected layer is not a last layer in the plurality of layers.

3. The method of claim 2 wherein the plurality of layers includes a set of convolutional layers and a subsequent set of fully connected non-convolutional layers, and wherein the selected layer is a first layer in the subsequent set of fully-connected non-convolutional layers.

4. The method of claim 1 wherein the first set of images includes a first set of object types that are different from a second set of object types included in the second set of images.

5. The method of claim 1 further comprising searching the Internet to obtain the second set of images.

6. An information handling system comprising:
one or more processors;
a memory coupled to at least one of the processors; and
a set of computer program instructions stored in the memory and executed by at least one of the processors in order to perform actions of:
training a classifier on a first set of images, wherein the classifier comprises a convolutional neural network that includes a plurality of layers;
processing a second set of images using the trained classifier, wherein the processing comprises:
extracting a plurality of feature sets from a selected layer included in the plurality of layers; and
generating a plurality of tags based upon metadata corresponding to the second set of images;

selecting a subset of the plurality of tags based on performing a nearest neighbors search on a new image feature set corresponding to a new image, wherein the new image feature set is extracted from the selected layer during processing of the new image; and
assigning at least one of the subset of the plurality of tags to the new image based upon voting on the subset of the plurality of tags.

7. The information handling system of claim 6 wherein the selected layer is not a last layer in the plurality of layers.

8. The information handling system of claim 7 wherein the plurality of layers includes a set of convolutional layers and a subsequent set of fully connected non-convolutional layers, and wherein the selected layer is a first layer in the subsequent set of fully-connected non-convolutional layers.

9. The information handling system of claim 6 wherein the first set of images includes a first set of object types that are different from a second set of object types included in the second set of images.

10. The information handling system of claim 6 wherein at least one of the one or more processors perform additional actions comprising:
searching the Internet to obtain the second set of images.

11. A computer program product stored in a computer readable storage medium, comprising computer program code that, when executed by an information handling system, causes the information handling system to perform actions comprising:
training a classifier on a first set of images, wherein the classifier comprises a convolutional neural network that includes a plurality of layers;
processing a second set of images using the trained classifier, wherein the processing comprises:
extracting a plurality of feature sets from a selected layer included in the plurality of layers; and
generating a plurality of tags based upon metadata corresponding to the second set of images;

selecting a subset of the plurality of tags based on performing a nearest neighbors search on a new image feature set corresponding to a new image, wherein the new image feature set is extracted from the selected layer during processing of the new image; and
assigning at least one of the subset of the plurality of tags to the new image based upon voting on the subset of the plurality of tags.

12. The computer program product of claim 11 wherein the selected layer is not a last layer in the plurality of layers.

13. The computer program product of claim 12 wherein the plurality of layers includes a set of convolutional layers and a subsequent set of fully connected non-convolutional layers, and wherein the selected layer is a first layer in the subsequent set of fully-connected non-convolutional layers.

14. The computer program product of claim 11 wherein the first set of images includes a first set of object types that are different from a second set of object types included in the second set of images.

Assignees

• IBM
• IBM

Inventors

• Chavez Aaron J
• Harper Devin R
• Lineback Nicholas A
• Turner Elliot B

Classifications

• G06V30/19173

  Classification techniques · CPC title

• G06V10/82 Primary

  using neural networks · CPC title

• G06F18/24147

  Distances to closest patterns, e.g. nearest neighbour classification · CPC title

• G06F18/2321

  using statistics or function optimisation, e.g. modelling of probability density functions · CPC title

• G06N3/045

  Combinations of networks · CPC title