Generating descriptive text for images in documents using seed descriptors

Patent metadata

Field	Value
Publication number	US-9971790-B2
Application number	US-201414211487-A
Country	US
Kind code	B2
Filing date	Mar 14, 2014
Priority date	Mar 15, 2013
Publication date	May 15, 2018
Grant date	May 15, 2018

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating descriptive text for images. In one aspect, a method includes identifying a set of seed descriptors for an image in a document that is hosted on a website. For each seed descriptor, structure information is generated that specifies a structure of the document with respect to the image and the seed descriptor. One or more templates are generated for each seed descriptor using the structure information for the seed descriptor. Each template can include image location information, document structure information, image feature information, and a generative rule that generates descriptive text for other images in other documents. Descriptive text for other images is generated using the templates and the other documents. The descriptive text is associated with the images.

First claim

What is claimed is:

1. A method performed by data processing apparatus, the method comprising:
identifying a set of one or more seed descriptors for a given image in a given document that is hosted on a website, wherein each seed descriptor has been identified as being a correct description of the given image;
for each seed descriptor:
identifying a location of at least one word of the seed descriptor in the given document by comparing each word of the seed descriptor to text included in the given document;
in response to identifying the location of the at least one word of the seed descriptor in the given document, generating, based on the identified location of the at least one word of the seed descriptor in the given document, structure information that specifies a structure of the given document with respect to the given image and the seed descriptor, the structure specifying a location of a string of text that includes the at least one word of the seed descriptor within the given document with respect to a location of the given image within the given document; and
generating one or more templates using the structure information for the seed descriptor, each template including:
image location information specifying a location of the given image within the given document;
document structure information specifying the structure of the given document with respect to the given image and the seed descriptor, including the location of the string of text that includes the at least one word of the seed descriptor within the given document with respect to the location of the given image within the given document;
image feature information specifying one or more feature values of the given image, each feature value representing a respective visual characteristic of the given image or data regarding an image file in which the given image is stored; and
a generative rule that generates descriptive text for other images in other documents; and

for each generated template:
identifying a set of documents that have (i) document structure information that matches the document structure information specified by the generated template and (ii) an image that has image feature information that matches the image feature information of the given image; and
for each document in the set of documents:
generating descriptive text for the image of the document using the generative rule of the generated template and the document; and
associating the descriptive text with the image.

2. The method of claim 1 , wherein each template further includes text information that specifies a first portion of the string of text and a wildcard for a second portion of the string of text, the second portion of the string of text corresponding to at least one term of the string of text that matches at least one of the terms of the template's corresponding seed descriptor, and the first portion of the string of text being text that does not match a term of the seed descriptor.

3. The method of claim 1 , wherein the seed descriptors are text data including one or more terms, and generating the structure information for a particular seed descriptor comprises:
identifying at least one term of the particular seed descriptor within the given document;
identifying a string of text that includes the at least one term of the particular seed descriptor; and
generating the document structure information for the particular seed descriptor based on the location of the string of text that includes the at least one term of the particular seed descriptor within the given document.

4. The method of claim 3 , wherein generating descriptive text for a particular image in a particular document comprises:
determining that the particular document has a structure that matches a particular template by applying the particular template to the particular document;
identifying a particular string of text within the particular document, the particular string of text being located within the particular document at a location that corresponds to the text location information of the particular template; and
generating the descriptive text for the particular image using the particular string of text.

5. The method of claim 4 , wherein determining that the particular document has a structure that matches the particular template comprises:
determining that the particular image is located within the particular document at a location that matches the location of the given image within the given document;
determining that particular string of text is located within the particular document at a location that matches the location of the string of text within the given document; and
determining that the particular image includes one or more features that match the one or more features of the given image.

6. The method of claim 1 , wherein generating, for each seed descriptor, one or more templates using the structure information for the seed descriptor comprises:
generating a plurality of candidate templates using the structure information for the seed descriptors;
for each candidate template:
determining a respective number of documents hosted on the website that have document structure information that matches the structure information of the document; and
determining whether the respective number of documents meets a template threshold;

for each candidate template that has a respective number of documents that meets the template threshold:
designating the candidate template as a template; and
using the candidate template to generating descriptive text for one or more images included in the documents hosted on the website;

for each candidate template that has a respective number of documents that does not meet the template threshold, determining to not use the candidate template to generate descriptive text for images included in the documents hosted on the website.

7. The method of claim 1 , wherein the structure of the document includes at least one embedded coding fragment, each embedded coding fragment being a hypertext markup language (HTML) tag pair that encloses the given image or the string of text.

8. The method of claim 1 , wherein the one or more features of the image comprises at least one of an aspect ratio for the given image, a display size for the given image, a shape of the given image, or data identifying cropping of the given image.

9. The method of claim 1 , wherein the one or more features of the given image comprises at least one of a file name for the given image or a file type for the given image.

10. The method of claim 1 , wherein each seed descriptor is a seed query that is a query that has at least a threshold performance with respect to the given image.

11. The method of claim 1 , wherein the other documents include only documents that are hosted on the website.

12. The method of claim 1 , wherein each seed descriptor comprises a seed query for which the given image is selected at a frequency that satisfies a threshold frequency when the given image is referenced by a search result provided in the response to the seed query.

13. The method of claim 1 , further comprising generating at least one of the seed descriptors by applying visual features of the given image to an image classification model.

14. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by data processing apparatus cause the data processing apparatus to perform operations comprising:
identifying a set of one or more seed descriptors for a given image in a given document that is hosted on a website, wherei

Assignees

• Google Llc

Inventors

• Wu Yonghui
• Flaster Michael E
• Keller Randall G
• Haahr Paul

Classifications

• G06F16/21

  Design, administration or maintenance of databases · CPC title

• G06F3/04842

  Selection of displayed objects or displayed text elements (G06F3/0482 takes precedence) · CPC title

• G06F16/5846

  using extracted text · CPC title

• G06F16/93

  Document management systems · CPC title

• G06F16/583 Primary

  using metadata automatically derived from the content · CPC title