Color representations for textual phrases

1. A method for color representation, comprising:

receiving an attribute-object pair including a first term comprising an attribute label and a second term comprising an object label;

encoding the attribute-object pair to produce encoded features using a neural network by generating attribute features and object features separately and combining the attribute features and the object features based on a predetermined ordering; and

generating a color profile for the attribute-object pair based on the encoded features, wherein the color profile is based on a compositional relationship between the first term and the second term.

2. The method of claim 1 , further comprising:

embedding the first term and the second term using a word embedding model to produce an attribute embedding and an object embedding.

3. The method of claim 2 , further comprising:

encoding the attribute embedding and the object embedding separately to produce the attribute features and the object features; and

concatenating the attribute features and the object features using an attribute-object ordering.

4. The method of claim 2 , further comprising:

encoding the attribute embedding and the object embedding together using a long short-term memory (LSTM) model.

5. The method of claim 1 , further comprising:

decoding the encoded features using the neural network, wherein the color profile is an output of the decoding.

6. The method of claim 1 , further comprising:

extracting image features and an image color profile from an image;

combining the color profile, an embedding of the first term and the second term, the image features, and the image color profile to produce a feature vector; and

generating a relevance score for the attribute-object pair and the image by decoding the feature vector.

7. The method of claim 1 , wherein:

the color profile comprises a human-interpretable color profile.

8. The method of claim 1 , further comprising:

searching for images based on the color profile.

9. An apparatus for learning color representation comprising a neural network, the neural network comprising:

an encoder configured to encode an attribute-object pair to produce encoded features by generating attribute features and object features separately and combining the attribute features and the object features based on a predetermined ordering of a first term comprising an attribute label and a second term comprising an object label; and

a decoder configured to generate a color profile for the attribute-object pair based on the encoded features, wherein the color profile is based on a compositional relationship between the first term and the second term.

10. The apparatus of claim 9 , wherein:

the encoder comprises a word embedding model, an attribute encoding layer, an object encoding layer, and a combined feature encoding layer.

11. The apparatus of claim 9 , wherein:

the encoder comprises a word embedding model and a long short-term memory (LSTM) layer.

12. The apparatus of claim 9 , further comprising:

a color classifier comprising a plurality of convolution layers, one or more attention layers, and a classification layer, wherein the color classifier is configured to identify an image color profile for images in an image set by attending to relevant portions of each image, and wherein attribute-object pairs associated with the image set are used together with the image color profile for training the neural network.

13. The apparatus of claim 9 , further comprising:

a ranking component comprising an image feature extractor, an image color profile component, a word embedding model, and a scoring component, wherein the ranking component is configured to generate a relevance score for the attribute-object pair and an image.

14. A non-transitory computer readable medium storing code for color representation, the code comprising instructions executable by at least one processor to:

receive an attribute-object pair including a first term comprising an attribute label and a second term comprising an object label;

encode the attribute-object pair to produce encoded features using a neural network that orders the first term and the second term based on the attribute label and the object label by generating attribute features and object features separately and combining the attribute features and the object features based on a predetermined ordering; and

generate a color profile for the attribute-object pair based on the encoded features, wherein the color profile is based on a compositional relationship between the first term and the second term.

15. The non-transitory computer readable medium of claim 14 , the code further comprising instructions executable by the at least one processor to:

embed the first term and the second term using a word embedding model to produce an attribute embedding and an object embedding.

16. The non-transitory computer readable medium of claim 15 , the code further comprising instructions executable by the at least one processor to:

encode the attribute embedding and the object embedding separately to produce the attribute features and the object features; and

concatenate the attribute features and the object features using an attribute-object ordering.

17. The non-transitory computer readable medium of claim 15 , the code further comprising instructions executable by the at least one processor to:

encode the attribute embedding and the object embedding together using a long short-term memory (LSTM) model.

18. The non-transitory computer readable medium of claim 14 , the code further comprising instructions executable by the at least one processor to:

decode the encoded features using the neural network, wherein the color profile is an output of the decoding.

19. The non-transitory computer readable medium of claim 14 , the code further comprising instructions executable by the at least one processor to:

extract image features and an image color profile from an image;

combine the color profile, an embedding of the first term and the second term, the image features, and the image color profile to produce a feature vector; and

generate a relevance score for the attribute-object pair and the image by decoding the feature vector.

20. The non-transitory computer readable medium of claim 14 , wherein:

the color profile comprises a human-interpretable color profile.