Method And Apparatus For Image Character Recognition Model Generation, And Vertically-Oriented Character Image Recognition

• US 20180060704A1
• Filed: 12/29/2016
• Published: 03/01/2018
• Est. Priority Date: 08/30/2016
• Status: Active Grant

• Current Assignee: Baidu Online Network Technology (Beijing) Co., Ltd (Baidu Incorporated)
• Original Assignee: Baidu Online Network Technology (Beijing) Co., Ltd (Baidu Incorporated)
• Inventors: XIE, Shufu, XIAO, Hang
• Granted Patent: US 10,176,409 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/24133 Distances to prototypes
• CPC Class Codes: G06T 11/60 Editing figures and text; Combining figures or text
• CPC Class Codes: G06T 3/4046 using neural networks
• CPC Class Codes: G06T 3/60 Rotation of whole images or parts thereof
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 30/19173 Classification techniques
• CPC Class Codes: G06V 30/224 of printed characters having additional code marks or containing code marks
• CPC Class Codes: G06V 30/287 of Kanji, Hiragana or Katakana characters

Abstract

Embodiments of the present disclosure disclose an image character recognition model generation method and apparatus, and a vertically-oriented character image recognition method and apparatus. The image character recognition model generation method includes: generating a rotated line character training sample, wherein the rotated line character training sample includes a rotated line character image and an expected character recognition result corresponding to the rotated line character image, and there is a difference of 90 degrees between character units in the rotated line character image and character units in a standard line character image; and training a set neural network by using the rotated line character training sample, to generate an image character recognition model. The technical solutions of the embodiments of the present disclosure overcome the technical defect that existing line character images cannot be directly used to train a vertically-oriented character image recognition model, thereby implementing the efficient recognition of vertically-oriented characters.

Claims

1. A method for generating an image character recognition model, comprising:

generating a rotated line character training sample, the rotated line character training sample comprising a rotated line character image and an expected character recognition result corresponding to the rotated line character image, character units in the rotated line character image rotated by 90 degrees in relation to character units in a standard line character image; and

training a set neural network by using the rotated line character training sample, to generate an image character recognition model.

2. The method according to claim 1, wherein the character units include at least one of the following:

Chinese characters, Japanese characters, English letters, and numerals.

3. The method according to claim 1, wherein the generating a rotated line character training sample comprises:

acquiring a standard line character image from a standard line character image sample as a target operation image;

cropping the target operation image based on an indication position of each character unit in the target operation image, and generating a set of character unit pictures corresponding to the character units in the target operation image;

rotating each of the character unit pictures in the set of the character unit pictures by 90 degrees respectively;

splicing the rotated character unit pictures based on the cropping order, to generate a rotated line character image; and

generating the rotated line character training sample based on the rotated line character image generated through splicing and an expected character recognition result corresponding to the target operation image.

4. The method according to claim 1, wherein the generating a rotated line character training sample comprises:

entering set character units into a vertically-oriented character picture generation tool, to generate a standard vertically-oriented character image;

rotating the standard vertically-oriented character image as a whole by 90 degrees, to generate a rotated line character image as a target operation image; and

generating the rotated line character training sample by using the set character units as an expected character recognition result corresponding to the target operation image.

5. The method according to claim 3, wherein the 90-degree rotation comprises a 90-degree clockwise rotation and/or a 90-degree anticlockwise rotation;

and

the set neural network comprises a recursive neural network.

6. The method according to claim 4, wherein the 90-degree rotation comprises a 90-degree clockwise rotation and/or a 90-degree anticlockwise rotation;

and

the set neural network comprises a recursive neural network.

7. A vertically-oriented character image recognition method, comprising:

rotating a to-be-recognized vertically-oriented character image as a whole by 90 degrees, to generate a rotated to-be-recognized image;

entering the rotated to-be-recognized image into an image character recognition model, the image character recognition model generated by training a set neural network model by using a rotated line character image training sample, the rotated line character training sample comprising a rotated line character image and an expected character recognition result corresponding to the rotated line character image, character units in the rotated line character image rotated by 90 degrees in relation to character units in a standard line character image; and

determining a vertically-oriented character recognition result corresponding to the to-be-recognized vertically-oriented character image based on an output from the image character recognition model.

8. The method according to claim 7, wherein before the entering the rotated to-be-recognized image into an image character recognition model, the method further comprises:

performing size normalization on the rotated to-be-recognized image based on model parameters of the image character recognition model.

9. The method according to claim 7, wherein the determining a vertically-oriented character recognition result corresponding to the to-be-recognized vertically-oriented character image based on an output from the image character recognition model comprises:

acquiring an output result of the image character recognition model as a line character recognition result, and recognizing display positions of character units in the rotated to-be-recognized image;

determining a display position of the line character recognition result in the to-be-recognized vertically-oriented character image based on the display positions of the character units in the rotated to-be-recognized image and the to-be-recognized vertically-oriented character image; and

generating the vertically-oriented character recognition result based on the line character recognition result and the display position of the line character recognition result in the to-be-recognized vertically-oriented character image.

10. An image character recognition model generation apparatus, comprising:

a training sample generation module for generating a rotated line character training sample, the rotated line character training sample comprising a rotated line character image and an expected character recognition result corresponding to the rotated line character image, character units in the rotated line character image rotated by 90 degrees in relation to character units in a standard line character image; and

an image character recognition model generation module for training a set neural network by using the rotated line character training sample, to generate an image character recognition model.

11. The apparatus according to claim 10, wherein the character units include at least one of the following:

Chinese characters, Japanese characters, English letters, and numerals.

12. The apparatus according to claim 10, wherein the training sample generation module comprises:

a target operation image acquiring unit for acquiring a standard line character image from a standard line character image sample as a target operation image;

a character unit picture set generation unit for cropping the target operation image based on an indication position of each character unit in the target operation image, and generating a set of character unit pictures corresponding to the character units in the target operation image;

a character unit picture rotating unit for rotating each of the character unit pictures in the set of character unit pictures by 90 degrees respectively;

a rotated line character image generation unit for splicing the rotated character unit pictures based on the cropping order, to generate a rotated line character image; and

a rotated line character training sample generation unit for generating the rotated line character training sample based on the rotated line character image generated through splicing and an expected character recognition result corresponding to the target operation image.

13. The apparatus according to claim 10, wherein the training sample generation module comprises:

a standard vertically-oriented character image generation unit for entering set character units into a vertically-oriented character picture generation tool, to generate a standard vertically-oriented character image;

a vertically-oriented character target operation image generation unit for rotating the standard vertically-oriented character image as a whole by 90 degrees, to generate a rotated line character image as a target operation image; and

a vertically-oriented character training sample generation unit for generating the rotated line character training sample by using the set character units as an expected character recognition result corresponding to the target operation image.

14. The apparatus according to claim 12, wherein the 90-degree rotation comprises a 90-degree clockwise rotation and/or a 90-degree anticlockwise rotation;

and

the set neural network comprises a recursive neural network.

15. The apparatus according to claim 13, wherein the 90-degree rotation comprises a 90-degree clockwise rotation and/or a 90-degree anticlockwise rotation;

and

the set neural network comprises a recursive neural network.

16. A vertically-oriented character image recognition apparatus, comprising:

a rotated to-be-recognized image generation module for rotating a to-be-recognized vertically-oriented character image as a whole by 90 degrees, to generate a rotated to-be-recognized image;

a rotated to-be-recognized image entering module for entering the rotated to-be-recognized image into an image character recognition model, the image character recognition model generated by training a set neural network model by using a rotated line character image training sample, the rotated line character training sample comprising a rotated line character image and an expected character recognition result corresponding to the rotated line character image, character units in the rotated line character image rotated by 90 degrees in relation to character units in a standard line character image; and

a vertically-oriented character recognition result determining module for determining a vertically-oriented character recognition result corresponding to the to-be-recognized vertically-oriented character image based on an output from the image character recognition model.

17. The apparatus according to claim 16, wherein before the rotated to-be-recognized image entering module, the apparatus further comprises:

an image size normalization module for performing size normalization on the rotated to-be-recognized image based on model parameters of the image character recognition model.

18. The apparatus according to claim 16, wherein the vertically-oriented character recognition result determining module comprises:

a character unit display position recognition unit for acquiring an output result of the image character recognition model as a line character recognition result, and recognizing display positions of character units in the rotated to-be-recognized image;

a line character recognition result display position determining unit for determining a display position of the line character recognition result in the to-be-recognized vertically-oriented character image based on the display positions of the character units in the rotated to-be-recognized image and the to-be-recognized vertically-oriented character image; and

a vertically-oriented character recognition result generation unit for generating the vertically-oriented character recognition result based on the line character recognition result and the display position of the line character recognition result in the to-be-recognized vertically-oriented character image.

19. A non-transitory computer storage medium storing a computer program, which when executed by one or more processors, cause the one or more processors to perform operations, the operations comprising:

generating a rotated line character training sample, the rotated line character training sample comprising a rotated line character image and an expected character recognition result corresponding to the rotated line character image, character units in the rotated line character image rotated by 90 degrees in relation to character units in a standard line character image; and

training a set neural network by using the rotated line character training sample, to generate an image character recognition model.