Determination of font similarity

Patent metadata

Field	Value
Publication number	US-9824304-B2
Application number	US-201514876660-A
Country	US
Kind code	B2
Filing date	Oct 6, 2015
Priority date	Oct 6, 2015
Publication date	Nov 21, 2017
Grant date	Nov 21, 2017

Abstract

Font recognition and similarity determination techniques and systems are described. In a first example, localization techniques are described to train a model using machine learning (e.g., a convolutional neural network) using training images. The model is then used to localize text in a subsequently received image, and may do so automatically and without user intervention, e.g., without specifying any of the edges of a bounding box. In a second example, a deep neural network is directly learned as an embedding function of a model that is usable to determine font similarity. In a third example, techniques are described that leverage attributes described in metadata associated with fonts as part of font recognition and similarity determinations.

First claim

What is claimed is:

1. In a digital medium environment to determine which fonts are similar to rendered text in an image, a method implemented by one or more computing devices comprising:
obtaining a model, by the one or more computing devices, that is trained using machine learning that is applied to a plurality of training images that include:
an anchor image having text using a font type;
a positive image having:
text that is different than the text of the anchor image; or
text having one or more applied perturbations; and

a negative image having text that is not in the font type;

determining similarity, by the one or more computing devices, of a font used for the rendered text in the image to respective ones of a plurality of fonts using the obtained model; and
outputting, by the one or more computing devices, the fonts that are similar to the rendered text in the image in a user interface based on the determining.

2. The method as described in claim 1 , wherein the text in the positive image that is different from the text in the anchor image has a font type that matches the font type used for the text in the anchor image.

3. The method as described in claim 1 , wherein the text of the positive image having the one or more applied perturbations matches the text of the anchor image.

4. The method as described in claim 1 , wherein the machine learning is performed by the one or more computing devices using a neural network that includes at least three columns to independently process the anchor image, the positive image, and the negative image, respectively.

5. The method as described in claim 4 , wherein the three columns of the neural network are constrained to match in both structure and parameter such that a single set of parameters is learned through use of a loss function for the neural network to train the model.

6. The method as described in claim 1 , wherein the machine learning is performed such that an embedded function of the model is automatically learned when a neural network that performs the machine learning is optimized using back propagation.

7. The method as described in claim 1 , wherein the anchor image, the positive image, and the negative image are selected through use of font metadata associated with respective said fonts.

8. The method as described in claim 7 , wherein the selecting is performed based on a Hamming distance computed between metadata vectors that describe respective said font metadata.

9. The method as described in claim 7 , wherein the font metadata includes family, weight, regular or italic, recommended use, or calligraphy style.

10. In a digital medium environment to generate a model that is suitable to determine which fonts are similar to rendered text in an image, a method implemented by one or more computing devices comprising:
selecting training images, by the one or more computing devices, using font metadata associated with respective said fonts used to render text included in respective said training images, the selected training images including:
an anchor image having text rendered using a font type;
a positive image having:
text that is different than the text of the anchor image; or
text having one or more applied perturbations; and

a negative image having text that is not in the font type;

controlling training of the model as an embedding function for font similarity, by the one or more computing devices, using machine learning using the anchor image, the positive image, and the negative image; and
exposing the model, by the one or more computing devices, to one or more applications such that the applications can determine the fonts that are similar to the rendered text in the image.

11. The method as described in claim 10 , wherein the selecting is performed based on a Hamming distance computed between metadata vectors that describe respective said font metadata.

12. The method as described in claim 10 , wherein the font metadata specifies a family, weight, regular or italic, recommended use, or calligraphy style.

13. The method as described in claim 10 , wherein the text in the positive image that is different from the text in the anchor image has a font type that matches the font type used for the text in the anchor image.

14. The method as described in claim 10 , wherein the text of the positive image having the one or more applied perturbations matches the text of the anchor image.

15. The method as described in claim 10 , further comprising determining similarity of a font used to render text in an image, by one or more computing devices, to one or more of a plurality of fonts using the trained model.

16. In a digital medium environment to determine which fonts are similar to rendered text in an image, a system comprising one or more computing devices including a processing system and memory having instructions stored thereon that are executable by the processing system to perform operations comprising:
selecting training images using font metadata associated with respective said fonts used to render text included in respective said training images, the selected training images including:
an anchor image having text using a font type;
a positive image having:
text that is different than the text of the anchor image; or
text having one or more applied perturbations; and

a negative image having text that is not in the font type;

controlling training of the model using machine learning using the anchor image, the positive image, and the negative image;
determining similarity of a font used for the rendered text in the image to respective ones of a plurality of fonts using the model; and
outputting one or more fonts that are similar to the rendered text in the image in a user interface based on the determining.

17. The system as described in claim 16 , wherein the selecting is performed based on a Hamming distance computed between metadata vectors that describe respective said font metadata.

18. The system as described in claim 16 , wherein the font metadata specifies a family, weight, regular or italic, recommended use, or calligraphy style.

19. The system as described in claim 16 , wherein the text in the positive image that is different from the text in the anchor image has a font type that matches the font type used for the text in the anchor image.

20. The system as described in claim 16 , wherein the text of the positive image having the one or more applied perturbations matches the text of the anchor image.

Assignees

• Adobe Systems Inc

Inventors

• Wang Zhaowen
• Jin Hailin

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/2413

  based on distances to training or reference patterns · CPC title

• G06V30/245 Primary

  Font recognition · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06N3/045

  Combinations of networks · CPC title