Method and apparatus for generating text line classifier

• US 10,146,994 B2
• Filed: 03/24/2016
• Issued: 12/04/2018
• Est. Priority Date: 03/25/2015
• Status: Active Grant

• Current Assignee: Alibaba Group Holding Ltd.
• Original Assignee: Alibaba Group Holding Ltd.
• Inventors: Jin, Xuan, Wang, Tianzhou, Xue, Qin
• Primary Examiner(s): WOLDEMARIAM, AKLILU K
• Application Number:
  US15/080,047
• Publication Number:
  US 20160283814A1
• Time in Patent Office: 985 Days
• Field of Search: 382161, 382171, 382182, 382176, 382165, 382187, 382229, 382159
• CPC Class Codes: G06F 18/00 Pattern recognition
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06V 30/1914 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries, e.g. user dictionaries
• CPC Class Codes: G06V 30/2268 using stroke segmentation
• CPC Class Codes: G06V 30/2445 Alphabet recognition, e.g. Latin, Kanji or Katakana
• CPC Class Codes: G06V 30/293 of characters other than Kanji, Hiragana or Katakana
• CPC Class Codes: G06V 30/413 Classification of content, e.g. text, photographs or tables

Abstract

A method of generating a text line classifier including generating text line samples by use of a present terminal system font reservoir. The method also includes extracting features from the text line samples and pre-stored marked-up samples. The method further includes training models by use of the extracted features to generate a text line classifier for recognizing text regions. With the system font reservoir being utilized for generating text line samples, the generated text line classifiers can target different scenes or different requirements for text region recognition with a high degree of applicability and wide application in addition to ease of implementation. Together with the combinational use of the marked up samples for extracting features from the text line samples, the generated text line classifiers provide for enhanced classification efficiency and accuracy.

Claims

1. A method of generating a text line classifier for recognizing text regions in an image, the method comprising:

generating a plurality of lines of text characters, a number of text characters in a line of text characters being variations of text characters in a font reservoir, generating the plurality of lines of text characters to include;

selecting a plurality of text characters from the font reservoir;

varying an aspect of the plurality of text characters to form a plurality of character samples;

randomly arranging a number of character samples from the plurality of character samples to form a line of character samples; and

varying an aspect of the line of character samples to form a line of text characters; and

generating a plurality of pre-stored marked-up samples;

extracting a plurality of features from the plurality of lines of text characters and the plurality of pre-stored marked-up samples; and

training a plurality of models using the plurality of extracted features to generate the text line classifier.

2. The method of claim 1, wherein training the plurality of models includes:

generating type models corresponding to types of the line of text characters based on the extracted features; and

assigning weights to the type models based on the pre-stored marked-up samples to generate the text line classifier.

3. A method of recognizing text regions in an image, the method comprising:

selecting a plurality of characters from a font reservoir;

generating a line of text based on the plurality of characters, generating the line of text to include;

modifying the plurality of characters to form a plurality of modified characters; and

arranging a number of modified characters of the plurality of modified characters to form a line of modified characters;

extracting a plurality of features from the line of text;

representing the plurality of features extracted from the line of text as a first vector;

training a model utilizing the first vector to obtain a trained model;

detecting an image to be recognized;

determining a second vector from the image;

inputting the second vector into the trained model, the trained model generating a score;

determining that the image to be recognized is a text region if the score is greater than a pre-determined threshold; and

determining that the image to be recognized is a non-text region if the score is less than the pre-determined threshold.

4. A method of generating a text line classifier, the method comprising:

selecting a plurality of text characters from a font reservoir;

varying an aspect of the plurality of text characters to form a plurality of character samples;

randomly arranging a number of character samples from the plurality of character samples to form a line of character samples;

varying an aspect of the line of character samples to form a line of text characters; and

extracting from the line of text characters one or more of a gradient orientation histogram feature, a gradient magnitude histogram feature, a pixel histogram feature, and a pixel histogram change feature.

5. The method of claim 4, wherein:

a number of text characters of the plurality of text characters differ only in that the number of text characters has a different font; and

the font reservoir includes Asian characters.

6. The method of claim 4, wherein:

the text characters in the line of text characters have a same size, a same rotation angle, and a same font; and

more than half of the text characters in the line of text characters are commonly used characters.

7. The method of claim 4, wherein extracting includes:

obtaining continuous regions of the line of text characters; and

extracting features of the continuous regions.

8. The method of claim 4, further comprising, generating a model corresponding to a type of the line of text characters based on the extracted features.

9. A non-transitory computer-readable storage medium having embedded therein program instructions, which when executed by one or more processors of a device, causes the device to execute a process that generates a text line classifier for recognizing text regions in an image, the process comprising:

generating a plurality of lines of text characters, a number of text characters in a line of text characters being variations of text characters in a font reservoir, generating the plurality of lines of text characters to include;

selecting a plurality of text characters from the font reservoir;

varying an aspect of the plurality of text characters to form a plurality of character samples;

randomly arranging a number of character samples from the plurality of character samples to form a line of character samples; and

varying an aspect of the line of character samples to form a line of text characters; and

generating a plurality of pre-stored marked-up samples;

extracting a plurality of features from the plurality of lines of text characters and the pre-stored marked-up samples; and

training a plurality of models using the plurality of extracted features to generate the text line classifier.

10. The non-transitory computer-readable storage medium of claim 9, wherein training the plurality of models includes:

generating type models corresponding to types of the line of text characters based on the extracted features; and

assigning weights to the type models based on the pre-stored marked-up samples to generate the text line classifier.

11. A non-transitory computer-readable storage medium having embedded therein program instructions, which when executed by one or more processors of a device, causes the device to execute a process that recognizes text regions in an image, the process comprising:

selecting a plurality of characters from a font reservoir;

generating a line of text based on the plurality of characters, generating the line of text to include;

modifying the plurality of characters to form a plurality of modified characters; and

arranging a number of modified characters of the plurality of modified characters to form a line of modified characters;

extracting a plurality of features from the line of text;

representing the plurality of features extracted from the line of text as a first vector;

training a model utilizing the first vector to obtain a trained model;

detecting an image to be recognized;

determining a second vector from the image;

inputting the second vector into the trained model, the trained model generating a score;

determining that the image to be recognized is a text region if the score is greater than a pre-determined threshold; and

determining that the image to be recognized is a non-text region if the score is less than the pre-determined threshold.