Objectionable content detector

• US 9,530,082 B2
• Filed: 04/24/2015
• Issued: 12/27/2016
• Est. Priority Date: 04/24/2015
• Status: Active Grant

• Current Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Original Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Inventors: Abdulkader, Ahmad Abdulmageed Mohammed, Rajaram, Giridhar
• Primary Examiner(s): Carter, Aaron W
• Application Number:
  US14/695,691
• Publication Number:
  US 20160314380A1
• Time in Patent Office: 613 Days
• US Class Current: 1/1
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06V 10/26 Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
• CPC Class Codes: G06V 10/44 Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
• CPC Class Codes: G06V 10/56 relating to colour
• CPC Class Codes: G06V 10/809 of classification results, e.g. where the classifiers operate on the same input data
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

A arbitrated image classifier can be trained to identify whether an image contains specified features, such as sexual, violent, or other potentially objectionable content. An arbitrated image classifier can include a configuration of classifiers and an arbitrator that determines a final image classification based on classification results from the classifiers. An arbitrated image classifier can be trained to identify image features by dividing images labeled as including or not including a specified feature into portions, which are provided to the classifiers of the arbitrated image classifier. The arbitrator of the arbitrated image classifier can determine a result for whether or not the image includes the specified feature. If the final result does not match the image label, parameter values for various of the classifiers or the arbitrator combining procedure can be adjusted. A trained arbitrated image classifier can then be used to determine whether new images include the particular feature.

Claims

1. A method for training an arbitrated image classifier to identify objectionable content, comprising:

obtaining the arbitrated image classifier comprising a configuration of multiple classifiers and an arbitrator, wherein each of the multiple classifiers in the configuration is associated with an image zone;

receiving an image associated with a label;

splitting the received image into a set of patches;

matching each patch, from the set of patches, to one of the multiple classifiers, wherein matching each patch, from the set of patches, to one of the multiple classifiers comprises;

determining an area of the received image from which that patch originated; and

matching the area to one of the image zones;

generating, by the matched classifiers, classification results, wherein individual ones of the classification results correspond to a patch from the set of patches;

providing the classification results to an arbitrator;

computing, by the arbitrator, an image classification, wherein the image classification is based on one or more of the classification results;

determining that the image classification does not match the label; and

in response to determining that the image classification does not match the label, updating parameters of the arbitrated image classifier.

2. The method of claim 1, wherein each of the multiple classifiers in the configuration is a deep neural network.

3. The method of claim 1, wherein each of the multiple classifiers in the configuration is a support vector machine.

4. The method of claim 1, wherein dividing the image into patches is performed by applying an established patch template.

5. The method of claim 1, wherein dividing the image into patches is performed using a randomization algorithm.

6. The method of claim 1,

wherein each classification result is associated with a weight value based on the image zone associated with the classifier that generated that classification result;

and

wherein computing, by the arbitrator, the image classification uses an algorithm that accounts for the weight values associated with the classification results.

7. The method of claim 1,

wherein each classification result is associated with a weight value;

and

wherein computing, by the arbitrator, the image classification uses an algorithm that accounts for the weight values associated with the classification results.

8. The method of claim 1, wherein the label indicates that the image contains sexual content.

9. The method of claim 1,

wherein the label indicates that the image contains sexual content, and

wherein none of the classification results indicate that the corresponding patch contains sexual content.

10. The method of claim 1, wherein updating parameters of the arbitrated image classifier comprises one or more of:

adjusting one or more weight values associated with individual ones of the multiple classifiers;

adjusting internal parameters of one or more of the multiple classifiers;

modifying an image zone associated with one or more of the multiple classifiers;

modifying an established patch template;

or

any combination thereof.

11. The method of claim 1, wherein the received image is a frame from a video file.

12. A method for applying an arbitrated image classifier to identify an image feature, the method comprising:

obtaining the arbitrated image classifier comprising a configuration of multiple classifiers and an arbitrator, the configuration trained using training data labeled for the image feature, wherein each of the multiple classifiers in the configuration is associated with an image zone;

receiving an image;

splitting the received image into a set of patches;

matching each patch, from the set of patches, to one of the multiple classifiers, wherein matching each patch, from the set of patches, to one of the multiple classifiers comprises;

determining an area of the received image from which that patch originated; and

matching the area to one of the image zones;

generating, by the matched classifiers, classification results for the image feature, wherein individual ones of the classification results correspond to a patch from the set of patches;

providing the classification results to an arbitrator;

computing, by the arbitrator, an image classification identifying whether the received image contains the image feature, wherein the image classification is based on one or more of the classification results; and

providing the image classification.

13. The method of claim 12,

wherein each classification result is associated with a weight value;

and

wherein computing, by the arbitrator, the image classification uses an algorithm that accounts for the weight values associated with the classification results.

14. The method of claim 12, wherein each of the multiple classifiers in the configuration is a deep neural network.

15. The method of claim 12, wherein dividing the image into patches is performed by applying an established patch template.

16. The method of claim 12, wherein the image feature is whether or not sexual content is present.

17. A system for training an arbitrated image classifier to identify objectionable content comprising:

a memory;

a network interface configured to;

obtain the arbitrated image classifier, the arbitrated image classifier comprising a configuration of multiple classifiers and an arbitrator, wherein each of the multiple classifiers in the configuration is associated with an image zone; and

receive an image associated with a label; and

one or more processors configured to;

split the received image into a set of patches;

match each patch, from the set of patches, to one of the multiple classifiers of the arbitrated image classifier, wherein matching each patch, from the set of patches, to one of the multiple classifiers comprises;

determining an area of the received image from which that patch originated; and

matching the area to one of the image zones;

use the matched classifiers of the arbitrated image classifier to compute classification results, wherein individual ones of the classification results correspond to a patch from the set of patches;

use the arbitrator to compute an image classification, wherein the image classification is based on one or more of the classification results;

determine that the image classification does not match the label; and

in response to determining that the image classification does not match the label, update;

weights associated with classification results from at least one of the classifiers of the arbitrated image classifier; and

internal features of one or more of the classifiers used to create classification results by those one or more classifiers.

18. The system of claim 17,

wherein the one or more processors use the arbitrator to compute an image classification by applying an algorithm that accounts for the weights associated with a classification results.

19. The system of claim 17, wherein the received image is a frame from a video file.

20. The system of claim 17,

wherein each classification result is associated with a weight value based on the image zone associated with the classifier that generated that classification result;

and

wherein computing, by the arbitrator, the image classification uses an algorithm that accounts for the weight values associated with the classification results.