Method of training image deep learning model and device thereof
Disclosed herein is an image deep learning model training method. The method includes sampling a twin negative comprising a first negative sample and a second negative sample by selecting the first negative sample with a highest similarity out of an anchor sample and a positive sample constituting a matching pair in each class and by selecting the second negative sample with a highest similarity to the first negative sample, and training the samples to minimize a loss of a loss function in each class by utilizing the anchor sample, the positive sample, the first and second negative samples for each class. The first negative sample is selected in a different class from a class comprising the matching pair, and the second negative sample is selected in a different class from classes comprising the matching pair and the first negative sample.
1. Method of training an image deep learning model comprising:
sampling a twin negative comprising a first negative sample and a second negative sample, the sampling includes:
selecting the first negative sample, the first negative sample having a highest, among a plurality of predefined classes, similarity to either of an anchor sample and a positive sample in which the anchor sample and the positive sample are a matching pair in a class of the plurality of predefined classes and different from a class of the first negative sample; and
selecting the second negative sample in which the second negative sample has a highest, among classes of the plurality of predefined classes that are different from the class of the matching pair and different from the class of the first negative sample, similarity to the first negative sample; and
training the image deep learning model, based on the anchor sample, the positive sample, and the first and second negative samples, to minimize a loss of a loss function in each class.
2. The method of claim 1 ,
wherein a multiplicity of first non-matching pairs comprising the positive sample of the matching pair and anchor samples belonging to classes different from the class respectively as pair elements and a multiplicity of second non-matching pairs comprising the anchor sample of the matching pair and positive samples belonging to classes different from the class respectively as pair elements are adopted as first negative candidates, and the anchor or positive sample of the different class belonging to a non-matching pair with the highest similarity among the first and second non-matching pairs is selected as the first negative sample.
3. The method of claim 1 ,
wherein a multiplicity of third non-matching pair comprising any one sample of the anchor sample and the positive sample constituting the first negative sample as a first element of a pair and respective samples belonging to different classes from the matching pair and the first negative sample and to a different kind from that of the first negative sample as a second element of the pair is adopted as a second negative candidate, and a sample of the different kind belonging to a third non-matching pair with a highest similarity among the third non-matching pairs is selected as the second negative sample.
4. The method of claim 1 ,
wherein the class makes clusters of the anchor sample and the positive sample for images based on a same object, and the matching pair is a pair of the anchor sample and the positive sample, which are extracted at a same location in an image of the same object, and the anchor sample, the positive sample, and the first and second negative samples include patches that are obtained in pixels of predetermined unit around a feature extracted from an image.
5. The method of claim 1 ,
wherein the similarity is calculated based on a distance between the anchor sample and the positive sample, and the highest similarity judges a sample with a shortest distance among distances of the anchor sample and the positive sample as a negative sample.
6. The method of claim 5 ,
wherein the distance is calculated using a descriptor describing a patch of the anchor sample and the positive sample, and the descriptor has an n-dimensional vector.
7. The method of claim 1 ,
wherein the loss function comprises an inter class distance function defining a distance between different classes and an intra class distance function defining a distance between the anchor sample and the positive sample within a same class.
8. The method of claim 7 ,
wherein the inter class distance function and the intra class distance function are not computed by adding up all classes of the plurality of predefined classes but are computed for each class, and then the image deep learning model are learned to minimize the loss of the loss function in each class.
9. The method of claim 7 ,
wherein the inter class distance function comprises a first margin and the intra class distance function comprises a second margin, and the first margin is set to a larger value than the second margin.
10. A device for training an image deep learning model comprising:
a transceiver receiving samples for a training target;
a memory storing instructions to train an image deep learning model for the samples; and
a processor configured to execute the instructions, configuring the processor to be configured to:
sample a twin negative comprising a first negative sample and a second negative sample, the sampling includes:
selecting, among a plurality of predefined classes, with a highest similarity to either of an anchor sample and a positive sample in which the anchor sample and the positive sample are a matching pair in a class of the plurality of predefined classes, as the first negative sample, the first negative sample is selected from a class different from the class of the matching pair; and
selecting the second negative sample in which the second negative sample has a highest, among classes of the plurality of the predefined classes that are different from the class of the matching pair and different from a class of the first negative sample, similarity to the first negative sample, and
training the image deep learning model, based on the anchor sample, the positive sample, and the first and second negative samples, to minimize a loss of a loss function in each class.
11. The device of claim 10 ,
wherein a multiplicity of first non-matching pairs comprising the positive sample of the matching pair and anchor samples belonging to classes different from the class respectively as pair elements and a multiplicity of second non-matching pairs comprising the anchor sample of the matching pair and positive samples belonging to classes different from the class respectively as pair elements are adopted as first negative candidates, and the anchor or positive sample of the different class belonging to a non-matching pair with the highest similarity among the first and second non-matching pairs is selected as the first negative sample.
12. The device of claim 10 ,
wherein a multiplicity of third non-matching pair comprising any one sample of the anchor sample and the positive sample constituting the first negative sample as a first element of a pair and respective samples belonging to different classes from the matching pair and the first negative sample and to a different kind from that of the first negative sample as a second element of the pair is adopted as a second negative candidate, and a sample of the different kind belonging to a third non-matching pair with a highest similarity among the third non-matching pairs is selected as the second negative sample.
13. The device of claim 10 ,
wherein the class makes clusters of the anchor sample and the positive sample for images based on a same object, and the matching pair is a pair of the anchor sample and the positive sample, which are extracted at a same location in an image of the same object, and the anchor sample, the positive sample, and the first and second negative samples include patches that are obtained in pixels of predetermined unit around a feature extracted from an image.
14. The device of claim 10 ,
wherein the similarity is calculated based on a distance between the anchor sample and the positive sample, and the highest similarity judges a sample with a shortest distance among distances of the anchor sample and the positive sample as a negative sample.
15. The device of claim 14 ,
wherein the distance is calculated using a descriptor describing a patch of the anchor sample and the positive sample, and the descriptor has an n-dimensional vector.
16. The device of claim 10 ,
wherein the loss function comprises an inter class distance function defining a distance between different classes and an intra class distance function defining a distance between the anchor sample and the positive sample within a same class.
17. The device of claim 16 ,
wherein the inter class distance function and the intra class distance function are not computed by adding up all classes of the plurality of predefined classes but are computed for each class, and then the image deep learning model are learned to minimize the loss of the loss function in each class.
18. The device of claim 16 ,
wherein the inter class distance function comprises a first margin and the intra class distance function comprises a second margin, and the first margin is set to a larger value than the second margin.