METHOD AND DEVICE FOR CLASSIFYING AN OBJECT OF AN IMAGE AND CORRESPONDING COMPUTER PROGRAM PRODUCT AND COMPUTER-READABLE MEDIUM

• US 20170061252A1
• Filed: 08/28/2016
• Published: 03/02/2017
• Est. Priority Date: 08/28/2015
• Status: Active Grant

• Current Assignee: Thomson Licensing (Vantiva SA)
• Original Assignee: Thomson Licensing (Vantiva SA)
• Inventors: BURGOS, Xavier, Kabbaj, Nezha, LE CLERC, Francois
• Granted Patent: US 10,169,683 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06F 18/24147 Distances to closest patterns, e.g. nearest neighbour classification
• CPC Class Codes: G06F 18/24323 Tree-organised classifiers
• CPC Class Codes: G06T 11/60 Editing figures and text; Combining figures or text
• CPC Class Codes: G06V 10/50 by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 30/19147 Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 30/19173 Classification techniques
• CPC Class Codes: G06V 40/171 Local features and components; Facial parts eye characteristics G06V40/18; Occluding parts, e.g. glasses; Geometrical relationships

Abstract

The disclosure relates to a method for classifying an object of a current image, a plurality of first landmarks representative of the shape of the object being associated with the current image, a first unique identifier being associated with each first landmark. According to the disclosure, the method includes, for at least a first landmark, a step of obtaining at least a first descriptor describing an area of the current image having the at least first selected landmark. Then, the first landmark is selected according to its first identifier, and when its first identifier corresponds to a second identifier of a second landmark, a second descriptor is used in a classifier for classifying the object. Finally, the method determines information representative of confidence of the shape according to the first descriptor and according to weighting information associated with the second descriptor.

Claims

1. A method for classifying an object of a current image, a plurality of first landmarks representative of the shape of said object being associated with the current image, a first unique identifier being associated with each first landmark, the method comprising:

for at least a first landmark selected among the plurality of first landmarks, obtaining at least a first descriptor describing an area of the current image comprising said at least first selected landmark,

said at least first landmark being selected according to its first identifier,

said at least first landmark being selected when its first identifier corresponds to a second identifier of a second landmark, a second descriptor of which being used in a classifier for classifying said object,

determining information representative of confidence of said shape according to said at least first descriptor and according to weighting information associated with said second descriptor.

2. The method for classifying an object of a current image according to claim 1, wherein said determining of information representative of confidence comprises comparing said at least first descriptor to a threshold associated with said at least second descriptor.

3. The method for classifying an object of a current image according to claim 1, wherein said information representative of confidence is normalised to provide a probability value of confidence.

4. The method for classifying an object of a current image according to claim 1, wherein:

for said current image, first descriptors associated with each landmark of said plurality of first landmarks are concatenated, following an order depending on their identifier, to form a first vector, and

for said classifier, second descriptors associated with each landmark of a plurality of second landmarks are concatenated, following said order depending on their identifier, to form a second vector.

5. The method for classifying an object of a current image according to claim 1, wherein said classifier is an iterative classifier implementing a plurality of iterations, wherein, from the second iteration, each iteration uses the result of the previous iteration.

6. The method for classifying an object of a current image according to claim 1, wherein said classifier belongs to the group comprising:

a classifier implementing an Adaboost algorithm,

a classifier implementing binary decision trees,

a support vector machine classifier,

a nearest neighbour classifier.

7. The method for classifying an object of a current image according to claim 1, wherein said first and second descriptors are of the same type, said type belonging to the group comprising:

a histogram of oriented gradients,

information representing a contour,

a luminance value,

an intensity value,

a texture value.

8. The method for classifying an object of a current image according to claim 1, wherein said method comprises a previous training of said classifier.

9. The method for classifying an object of a current image according to claim 8, wherein said training of said classifier is performed by using a dataset of training images comprising:

a first set of training images, which provide a positive classifying result when being processed by said classifier, and

a second set of training images, which provides a negative classifying result when being processed by said classifier.

10. The method for classifying an object of a current image according to claim 9, wherein said second set of training images contains images where the object is not present and/or images where object landmark annotation is erroneous.

11. The method for classifying an object of a current image according to claim 1, wherein said method comprises a previous landmark annotating of said object of said current image

12. The method for classifying an object of a current image according to claim 9, wherein said landmark annotating of said object of said current image is performed by using one of the method belonging to the group comprising:

a cascaded pose regression,

a robust cascaded pose regression.

25. A non-transitory computer-readable medium comprising a computer program product recorded thereon and capable of being run by a processor, including program code instructions for implementing the steps of the method for detecting at least one fitting error according to claim 1 when it is executed by a processor.

13. A device for classifying an object of a current image, a plurality of first landmarks representative of the shape of said object being associated with the current image, a first unique identifier being associated with each first landmark, said device comprising a processor configured to:

For at least a first landmark selected among the plurality of first landmarks, obtain at least a first descriptor describing an area of the current image comprising said at least first selected landmark,

said at least first landmark being selected according to its first identifier,

said at least first landmark being selected when its first identifier corresponds to a second identifier of a second landmark, a second descriptor of which being used in a classifier for classifying said object,

determine information representative of confidence of said shape according to said at least first descriptor and according to weighting information associated with said second descriptor.

14. The device for classifying an object of a current image according to claim 13, wherein said determining of information representative of confidence comprises comparing said at least first descriptor to a threshold associated with said at least second descriptor.

15. The device for classifying an object of a current image according to claim 13, wherein said information representative of confidence is normalised to provide a probability value of confidence.

16. The device for classifying an object of a current image according to claim 13, wherein:

for said current image, first descriptors associated with each landmark of said plurality of first landmarks are concatenated, following an order depending on their identifier, to form a first vector, and

for said classifier, second descriptors associated with each landmark of a plurality of second landmarks are concatenated, following said order depending on their identifier, to form a second vector.

17. The device for classifying an object of a current image according to claim 13, wherein said classifier is an iterative classifier implementing a plurality of iterations, wherein, from the second iteration, each iteration uses the result of the previous iteration.

18. The device for classifying an object of a current image according to claim 13, wherein said classifier belongs to the group comprising:

a classifier implementing an Adaboost algorithm,

a classifier implementing binary decision trees,

a support vector machine classifier,

a nearest neighbour classifier.

19. The device for classifying an object of a current image according to claim 13, wherein said first and second descriptors are of the same type, said type belonging to the group comprising:

a histogram of oriented gradients,

information representing a contour,

a luminance value,

an intensity value,

a texture value.

20. The device for classifying an object of a current image according to claim 13, wherein said processor is further configured to process a previous training of said classifier.

21. The device for classifying an object of a current image according to claim 20, wherein said training of said classifier is performed by using a dataset of training images comprising:

a first set of training images, which provide a positive classifying result when being processed by said classifier, and

a second set of training images, which provides a negative classifying result when being processed by said classifier.

22. The device for classifying an object of a current image according to claim 21, wherein said second set of training images contains images where the object is not present and/or images where object landmark annotation is erroneous.

23. The device for classifying an object of a current image according to claim 13, wherein said method comprises a previous landmark annotating of said object of said current image

24. The device for classifying an object of a current image according to claim 21, wherein said landmark annotating of said object of said current image is performed by using one of the method belonging to the group comprising:

a cascaded pose regression,

a robust cascaded pose regression.