Cognitive analysis and classification of apparel images

Patent metadata

Field	Value
Publication number	US-10635952-B2
Application number	US-201916449601-A
Country	US
Kind code	B2
Filing date	Jun 24, 2019
Priority date	Apr 11, 2018
Publication date	Apr 28, 2020
Grant date	Apr 28, 2020

Abstract

As disclosed, f-scores can be generated for apparel items. Training images are identified, where each training image is associated with a corresponding set of tags including information about a plurality of attributes. A first convolutional neural network (CNN) is trained based on the plurality of training images and a first attribute. The first CNN is iteratively refined by, for each respective attribute, removing a set of neurons from the first CNN and retraining the first CNN based on the training images and the respective attribute. Upon determining that the first CNN has been trained based on each of the attributes, one or more CNNs are generated based on the first CNN. An image is received, where the image depicts an apparel item. The image is processed using the one or more CNNs, and an f-score for the apparel item is determined based on the output.

First claim

What is claimed is:

1. A method comprising:
identifying a plurality of training images, wherein each training image depicts a respective apparel item, and wherein each training image is associated with a corresponding set of tags, each tag including information about a respective attribute in a plurality of attributes;
training a first convolutional neural network (CNN) based on the plurality of training images and a first attribute in the plurality of attributes;
iteratively refining the first CNN by, for each respective attribute in the plurality of attributes:
removing a set of neurons from the first CNN, wherein the removed set of neurons correspond to a classifier, prediction, or multinomial-prediction layer; and
retraining the first CNN based on the plurality of training images and the respective attribute;

upon determining that the first CNN has been trained based on each of the plurality of attributes, generating a first set of one or more trained CNNs based on the first CNN;
receiving a first image to be classified based on the plurality of attributes, wherein the first image depicts a first apparel item;
processing the first image using the first set of one or more trained CNNs to generate one or more results vectors; and
determining, based on the one or more results vectors, an f-score for the first apparel item.

2. The method of claim 1 , the method further comprising:
processing the first image with one or more machine learning models to determine a first apparel category to which the first apparel item belongs; and
determining to process the first image using the first set of one or more trained CNNs based on the determined first apparel category.

3. The method of claim 2 , the method further comprising:
receiving a second image to be classified, wherein the second image includes a second apparel item;
processing the first image with the one or more machine learning models to determine a second apparel category to which the second apparel item belongs; and
determining to process the first image using a second set of one or more trained CNNs based on the determined second apparel category.

4. The method of claim 1 , wherein the plurality of attributes includes at least one of:
(i) a publication that featured the respective apparel item;
(ii) a designer of the respective apparel item;
(iii) a brand of the respective apparel item; and
(iv) a measure of contemporariness of the respective apparel item.

5. The method of claim 1 , the method further comprising:
determining that a public event has occurred;
retrieving at least one additional training image associated with the public event; and
refining at least one of the first set of one or more trained CNNs based on the at least one additional training image.

6. The method of claim 1 , the method further comprising determining a predicted popularity for the first apparel item based on the determined f-score and one or more trend vectors corresponding to one or more attributes in the plurality of attributes.

7. The method of claim 6 , wherein a first trend vector of the one or more trend vectors corresponds to the first attribute in the plurality of attributes, wherein the first attribute comprises a plurality of categories, and wherein the first trend vector comprises information about a popularity of each of the plurality of categories.

8. The method of claim 1 , wherein generating the first set of one or more trained CNNs comprises generating at least one CNN based on a combination of at least two of the plurality of attributes.

9. The method of claim 1 , wherein retraining the first CNN comprises modifying the weights of at least one middle layer of the first CNN and training a new classifier layer.

Assignees

• IBM

Inventors

• Sewak Mohit
• Hariharan Karthik P
• Fedulova Irina

Classifications

• G06Q30/0611

  Request for offers or quotes · CPC title

• G06N5/022

  Knowledge engineering; Knowledge acquisition · CPC title

• G06K2209/27

  Physics · mapped topic

• G06K9/6268

  Physics · mapped topic

• G06N3/08

  Learning methods · CPC title