Long-tail large scale face recognition by non-linear feature level domain adaptation

1. A computer-implemented method for facial recognition, the method comprising:

receiving, by a processor device, a plurality of images;

extracting, by the processor device with a feature extractor utilizing a convolutional neural network (CNN) with an enlarged intra-class variance of long-tail classes, feature vectors for each of the plurality of images;

generating, by the processor device with a feature generator, discriminative feature vectors for each of the feature vectors;

classifying, by the processor device utilizing a fully connected classifier, an identity from the discriminative feature vector; and

controlling an operation of a processor-based machine to react in accordance with the identity.

2. The computer-implemented method as recited in claim 1 , further comprising employing an alternative bi-stage strategy to train the feature extractor, the feature generator, and the fully connected classifier.

3. The computer-implemented method as recited in claim 1 , wherein generating includes employing a center-based non-linear feature transfer.

4. The computer-implemented method as recited in claim 1 , wherein generating includes utilizing a pair-wise non-center-based non-linear feature transfer.

5. The computer-implemented method as recited in claim 1 , wherein extracting includes utilizing a pixel-wise reconstruction loss.

6. The computer-implemented method as recited in claim 1 , wherein extracting includes sharing covariance matrices across all classes to transfer intra-class variance from regular classes to the long-tail classes.

7. The computer-implemented method as recited in claim 1 , wherein generating includes optimizing a softmax loss by joint regularization of weights and features through a magnitude of an inner product of the weights and features.

8. The computer-implemented method as recited in claim 1 , wherein extracting includes averaging the feature vector with a flipped feature vector, the flipped feature vector being generated from a horizontally flipped frame from one of the plurality of images.

9. The computer-implemented method as recited in claim 1 , wherein each of the plurality of images is selected from the group consisting of an image, a video, and a frame from the video.

10. The computer-implemented method as recited in claim 2 , wherein one stage of the alternative bi-stage strategy includes fixing the feature extractor and applying the feature generator to generate new transferred features that are more diverse and violate a decision boundary.

11. The computer-implemented method as recited in claim 2 , wherein one stage of the alternative bi-stage strategy includes fixing the fully connected classifier and updating the feature extractor and the feature generator.

12. A computer program product for facial recognition, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:

receiving, by a processor device, a plurality of images;

extracting, by the processor device with a feature extractor utilizing a convolutional neural network (CNN) with an enlarged intra-class variance of long-tail classes, feature vectors for each of the plurality of images;

generating, by the processor device with a feature generator, discriminative feature vectors for each of the feature vectors;

classifying, by the processor device utilizing a fully connected classifier, an identity from the discriminative feature vector; and

controlling an operation of a processor-based machine to react in accordance with the identity.

13. A facial recognition system, comprising:

a camera;

a processing system including a processor device and memory receiving input from the camera, the processing system programmed to:

receive a plurality of images;

extract, with a feature extractor utilizing a convolutional neural network (CNN) with an enlarged intra-class variance of long-tail classes, feature vectors from each of the plurality of images;

generate, with a feature generator, discriminative feature vectors for each of the feature vectors;

classify, with a fully connected classifier, an identity from the discriminative feature vectors.

14. The facial recognition system as recited in claim 13 , further programmed to train the feature extractor, the feature generator, and the fully connected classifier with an alternative bi-stage strategy.

15. The facial recognition system as recited in claim 13 , wherein the feature extractor shares covariance matrices across all classes to transfer intra-class variance from regular classes to the long-tail classes.

16. The facial recognition system as recited in claim 13 , wherein the feature generator optimizes a softmax loss by joint regularization of weights and features through a magnitude of an inner product of the weights and features.

17. The facial recognition system as recited in claim 13 , wherein the feature extractor averages the feature vector with a flipped feature vector, the flipped feature vector being generated from a horizontally flipped frame from one of the plurality of images.

18. The system as recited in claim 13 , further programmed to control an operation of a processor-based machine to react in accordance with the identity.

19. The facial recognition system as recited in claim 14 , wherein one stage of the alternative bi-stage strategy fixes the feature extractor and applies the feature generator to generate new transferred features that are more diverse and violate a decision boundary.

20. The facial recognition system as recited in claim 14 , wherein one stage of the alternative bi-stage strategy fixes the fully connected classifier and updates the feature extractor and the feature generator.