Word segmentation system, method and device

1. In an optical character recognition (“OCR”) system, a word segmentation method, comprising:

acquiring a sample image comprising a word spacing marker or a non-word spacing marker;

processing the sample image with a convolutional neural network to obtain a first eigenvector corresponding to the sample image, a word spacing probability value and/or a non-word spacing probability value corresponding to the first eigenvector;

acquiring a to-be-tested image, and processing the to-be-tested image with the convolutional neural network to obtain a second eigenvector corresponding to the to-be-tested image, a word spacing probability value or a non-word spacing probability value corresponding to the second eigenvector; and

performing word segmentation on the to-be-tested image by using the just obtained word spacing probability value or the non-word spacing probability value.

2. The OCR system according to claim 1 , wherein acquiring the sample image comprises:

extracting at least two letter images from a font library, splicing the at least two letter images together to obtain a spliced image, and setting a word spacing marker or a non-word spacing marker for the spliced image; and

marking the spliced image as the sample image or performing one of the following operations or any combination thereof on the spliced image: a translation operation, a rotation operation, a stretching operation, adding Gaussian noise, and marking the image on which the operation is performed as the sample image.

3. The OCR system according to claim 1 , wherein the method further comprises:

before processing the sample image with the convolutional neural network, converting the sample image into a gray scale sample image, converting the gray scale sample image into a sample image with a size of M*N, and normalizing the sample image with the size of M*N; and

before processing the to-be-tested sample image with the convolutional neural network, converting the to-be-tested image into a gray scale, to-be-tested image, converting the gray scale, to-be-tested image into a to-be-tested image with a size of M*L, and normalizing the to-be-tested image with the size of M*L,

wherein the number M is a positive integer, the number N is a positive integer, and L is a positive integer; and a ratio of the number M and the number L is the same as an aspect ratio of the to-be-tested image.

4. The OCR system according to claim 1 , wherein processing the to-be-tested image with the convolutional neural network to obtain a second eigenvector corresponding to the to-be-tested image comprises:

translating the to-be-tested image to the left K times or to the right K times, wherein each translation has a different translational position so that K+1 different to-be-tested images are obtained, and the number K is a positive integer; and

processing the different K+1 to-be-tested images with the convolutional neural network to obtain K+1 eigenvectors corresponding to the to-be-tested images, and combining the obtained K+1 eigenvectors respectively corresponding to the to-be-tested images to obtain a second eigenvector corresponding to the to-be-tested image.

5. The OCR system according to claim 1 or 4 , wherein obtaining the word spacing probability value or the non-word spacing probability value corresponding to the second eigenvector comprises:

obtaining a plurality of first eigenvectors matching the second eigenvector by using the convolutional neural network, and taking the word spacing probability value or the non-word spacing probability value corresponding to the plurality of first eigenvectors as the word spacing probability value or the non-word spacing probability value corresponding to the second eigenvector.

6. The OCR system according to claim 1 or 4 , wherein performing word segmentation on the to-be-tested image by using the just obtained word spacing probability value or the non-word spacing probability value comprises:

obtaining a probability value result graph by using a position of the to-be-tested image corresponding to the second eigenvector as an X coordinate and using the word spacing probability value corresponding to the second eigenvector as a Y coordinate;

finding a position of a maximum value of the word spacing probability value from the probability value result graph, and performing word segmentation on the to-be-tested image by using the maximum value position; alternatively, obtaining the probability value result graph by taking the position of the to-be-tested image corresponding to the second eigenvector as the X coordinate and taking the non-word spacing probability value corresponding to the second eigenvector as the Y coordinate; and

finding a position of a minimum value of the non-word spacing probability value from the probability value result graph, and performing word segmentation on the to-be-tested image by using the minimum value position.

7. A word segmentation device comprising a processor adaptable for reading computer instructions and executing program commands, wherein the processor is configured to

execute a training module comprising software instructions configured to acquire a sample image, the sample image comprising a word spacing marker or a non-word spacing marker, and process the sample image with a convolutional neural network to obtain a first eigenvector corresponding to the sample image, a word spacing probability value and/or a non-word spacing probability value corresponding to the first eigenvector; and

execute a predicting module comprising software instructions, configured to acquire a to-be-tested image and process the to-be-tested image with the convolutional neural network to obtain a second eigenvector corresponding to the to-be-tested image, a word spacing probability value or a non-word spacing probability value corresponding to the second eigenvector; and

perform word segmentation on the to-be-tested image by using the just obtained word spacing probability value or the non-word spacing probability value.

8. The device according to claim 7 , wherein

the training module is configured to extract at least two letter images from a font library in a process of acquiring the sample image, splice the at least two letter images together to obtain a spliced image, and set a word spacing marker or a non-word spacing marker for the spliced image; and

mark the spliced image as the sample image or perform one of the following operations or any combination thereof on the spliced image: a translation operation, a rotation operation, a stretching operation, adding Gaussian noise, and mark the image on which the operation is performed as the sample image.

9. The device according to claim 7 , wherein

the training module is further configured to do the following: before the sample image is processed with the convolutional neural network, convert the sample image into a gray scale sample image, convert the gray scale sample image into a sample image with a size of M*N, and normalize the sample image with the size of M*N; and

the predicting module is further configured to do the following: before the to-be-tested sample image is processed with the convolutional neural network, convert the to-be-tested image into a gray scale, to-be-tested image, convert the gray scale, to-be-tested image into a to-be-tested image with a size of M*L, and normalize the to-be-tested image with the size of M*L,

wherein the number M is a positive integer, the number N is a positive integer, and L is a positive integer; and a ratio of the number M and the number L is the same as an aspect ratio of the to-be-tested image.

10. The device according to claim 7 , wherein

the predicting module is configured to do the following: in a process of processing the to-be-tested image with the convolutional neural network to obtain a second eigenvector corresponding to the to-be-tested image, translate the to-be-tested image to the left K times or to the right K times, wherein each translation has a different translational position so that K+1 different to-be-tested images are obtained, and the number K is a positive integer; and

process the different K+1 to-be-tested images with the convolutional neural network to obtain K+1 eigenvectors corresponding to the to-be-tested images, and combine the obtained K+1 eigenvectors respectively corresponding to the to-be-tested images to obtain a second eigenvector corresponding to the to-be-tested image.

11. The device according to claim 7 or 10 , wherein

the predicting module is configured to do the following: in a process of obtaining the word spacing probability value or the non-word spacing probability value corresponding to the second eigenvector, obtain a plurality of first eigenvectors matching the second eigenvector by using the convolutional neural network, and take the word spacing probability value or the non-word spacing probability value corresponding to the plurality of first eigenvectors as the word spacing probability value or the non-word spacing probability value corresponding to the second eigenvector.

12. The device according to claim 7 or 10 , wherein

the predicting module is configured to do the following: in a process of performing word segmentation on the to-be-tested image by using the just obtained word spacing probability value or non-word spacing probability value, obtain a probability value result graph by using a position of the to-be-tested image corresponding to the second eigenvector as an X coordinate and using the word spacing probability value corresponding to the second eigenvector as a Y coordinate;

find a position of a maximum value of the word spacing probability value from the probability value result graph, and perform word segmentation on the to-be-tested image by using the maximum value position;

alternatively, obtain the probability value result graph by taking the position of the to-be-tested image corresponding to the second eigenvector as the X coordinate and taking the non-word spacing probability value corresponding to the second eigenvector as the Y coordinate; and find a position of a minimum value of the non-word spacing probability value from the probability value result graph, and perform word segmentation on the to-be-tested image by using the minimum value position.