Method of training an image classification model
A method of training a neural network for classifying an image into one of a plurality of classes, the method comprising: extracting, from the neural network, a plurality of subclass center vectors for each class; inputting an image into the neural network, wherein the image is associated with a predetermined class; generating, using the neural network, an embedding vector corresponding to the input image; determining a similarity score between the embedding vector and each of the plurality of subclass center vectors; updating parameters of the neural network in dependence on a plurality of the similarity scores using an objective function; extracting a plurality of updated parameters from the neural network; and updating each subclass center vector in dependence on the extracted updated parameters.
1. A method of training a neural network for classifying an image into one of a plurality of classes, wherein the method is applied to a computer device comprising at least one processor, and the method comprises:
extracting, from the neural network, a plurality of subclass center vectors for each class;
inputting an image into the neural network, wherein the image is associated with a predetermined class;
generating, using the neural network, an embedding vector corresponding to the input image;
determining a similarity score between the embedding vector and each of the plurality of subclass center vectors;
updating parameters of the neural network based on a plurality of the similarity scores using an objective function;
extracting a plurality of updated parameters from the neural network;
updating each subclass center vector based on the extracted updated parameters; and
determining a closest subclass center vector for each class using the similarity scores,
wherein the objective function comprises a multi-center loss term comparing the similarity score between the embedding vector and the closest subclass center vector from the predetermined class to the similarity scores between the embedding vector and the closest subclass center vectors from each of the other classes.
2. The method of claim 1 , further comprising: prior to updating the parameters of the neural network:
inputting a further image into the neural network, wherein the further image is associated with a predetermined class;
generating, using the neural network, a further embedding vector corresponding to the further input image;
determining a further similarity score between the further embedding vector and each of the plurality of subclass center vectors;
wherein updating the parameters of the neural network is further based on the further similarity scores.
3. The method of claim 1 , wherein the multi-center loss term is a margin-based softmax loss function.
4. The method of claim 1 , wherein the embedding vector and each subclass center vector are normalized, and wherein the similarity score is an angle between the embedding vector and one of the subclass center vectors.
5. The method of claim 1 , wherein each class comprises a dominant subclass, and the method further comprises:
for each class, determining an intra-class similarity score between a dominant subclass center vector and each of the other subclass center vectors in the class,
wherein the objective function comprises an intra-class compactness term that uses the intra-class similarity scores.
6. The method of claim 5 , wherein each subclass center vector is normalized, and the intra-class similarity score is an angle between the dominant subclass center vector and another subclass center vector in the class.
7. The method of claim 1 , wherein the neural network comprises a plurality of connected layers, and wherein each subclass center vector is updated using updated parameters extracted from a last fully connected layer of the neural network.
8. The method of claim 1 , wherein each class comprises a dominant subclass, and the method further comprises:
discarding a non-dominant subclass from a class based on the subclass center vector of the non-dominant subclass being above a threshold distance from the dominant subclass center vector of the class.
9. The method of claim 8 , wherein the discarding the non-dominant subclass from the class is performed based on a threshold condition being satisfied.
10. The method of claim 9 , wherein the threshold condition is a first threshold number of training epochs being exceeded.
11. The method of claim 1 , further comprising:
discarding all non-dominant subclasses based on a further threshold condition being satisfied.
12. The method of claim 11 , wherein the further threshold condition is a second threshold number of training epochs being exceeded.
13. The method of claim 1 , wherein the image is a face image.
14. The method of claim 1 , wherein the class corresponds to a classification condition of the image, and
wherein the image is from a batch that contains label noise such that the batch comprises at least one image labelled to relate to a class that does not correspond to the classification condition of the at least one image.
15. An apparatus comprising:
one or more processors; and
a memory, the memory comprising computer readable instructions that are executed by the one or more processors, and cause the apparatus to perform a method of training a neural network for classifying an image into one of a plurality of classes, the method including:
extracting, from the neural network, a plurality of subclass center vectors for each class;
inputting an image into the neural network, wherein the image is associated with a predetermined class;
generating, using the neural network, an embedding vector corresponding to the input image;
determining a similarity score between the embedding vector and each of the plurality of subclass center vectors;
updating parameters of the neural network based on a plurality of the similarity scores using an objective function;
extracting a plurality of updated parameters from the neural network;
updating each subclass center vector based on the extracted updated parameters; and
determining a closest subclass center vector for each class using the similarity scores,
wherein the objective function comprises a multi-center loss term comparing the similarity score between the embedding vector and the closest subclass center vector from the predetermined class to the similarity scores between the embedding vector and the closest subclass center vectors from each of the other classes.
16. A non-transitory computer-readable medium comprising computer readable instructions that are executed by a processor of a computing device, and cause the computing device to perform a method of training a neural network for classifying an image into one of a plurality of classes, the method including:
extracting, from the neural network, a plurality of subclass center vectors for each class;
inputting an image into the neural network, wherein the image is associated with a predetermined class;
generating, using the neural network, an embedding vector corresponding to the input image;
determining a similarity score between the embedding vector and each of the plurality of subclass center vectors;
updating parameters of the neural network based on a plurality of the similarity scores using an objective function;
extracting a plurality of updated parameters from the neural network;
updating each subclass center vector based on the extracted updated parameters; and
determining a closest subclass center vector for each class using the similarity scores,
wherein the objective function comprises a multi-center loss term comparing the similarity score between the embedding vector and the closest subclass center vector from the predetermined class to the similarity scores between the embedding vector and the closest subclass center vectors from each of the other classes.
17. The apparatus of claim 15 , wherein the method further comprises: prior to updating the parameters of the neural network,
inputting a further image into the neural network, wherein the further image is associated with a predetermined class;
generating, using the neural network, a further embedding vector corresponding to the further input image;
determining a further similarity score between the further embedding vector and each of the plurality of subclass center vectors;
wherein updating the parameters of the neural network is further based on the further similarity scores.
18. The non-transitory computer-readable medium of claim 16 , wherein the method further comprises: prior to updating the parameters of the neural network,
inputting a further image into the neural network, wherein the further image is associated with a predetermined class;
generating, using the neural network, a further embedding vector corresponding to the further input image;
determining a further similarity score between the further embedding vector and each of the plurality of subclass center vectors;
wherein updating the parameters of the neural network is further based on the further similarity scores.
19. The apparatus of claim 15 , wherein the multi-center loss term is a margin-based softmax loss function.
20. The non-transitory computer-readable medium of claim 16 , wherein the multi-center loss term is a margin-based softmax loss function.