JPH07152862A - Character segmenting device for character recognition - Google Patents

Abstract

PURPOSE:To correctly segment a character without being affected by noise. CONSTITUTION:A character segmenting device is provided with a picture input part 10 inputting an analog picture (b) including the character a1, binary picture data generating part 12 generating binary image data (c) by binalizing the analog picture (b) every picture element, an input data generating part 14 generating input data X1-Xn and Y1-Ym by dividing binary picture data (c) into picture elements every line in a horizontal direction X and the longitudinal direction Y, a neural network processing part 16 recognizing whether or not a part of the character is included by executing the neural network processing to the respective input data X1-Xn and Y1-Ym and a circumscribed frame output part 18 outputting the position (d) of the circumscribed frame a2 of the character a1 based on recognition results x1-xn and y1-ym.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character slicing device used in a character recognizing technique for optically reading a character, and more specifically, a character slicing device for prescribing a character in an input image for slicing out the character Regarding a character slicing device.

[0002]

2. Description of the Related Art In character recognition technology, for example, as shown in FIG.
21, 822 and 823 are respectively attached to the circumscribing frames 802 and 82.
After cutting out at 4,825,826, those characters 8
01, ... are recognized.

In a conventional character cutting device, for example, FIG.
As shown in, the pixels of the input images 80 and 82 are added in the vertical direction and the horizontal direction to project the black portion, and projection histograms 803, 804, 827 and 828 are created.
In the projection histograms 803, ..., a portion above a certain threshold is cut out as an area in which the characters 801 ,. Such a character cutting device is disclosed in, for example, Japanese Patent Laid-Open No. 1-11.
It is described in Japanese Patent No. 4992.

[0004]

However, the conventional character slicing device has the following problems.

If the characters 801, ... Are broken or broken, or if noise is mixed in, the characters 801 ,.
It was difficult to accurately cut out 1, ... this is,
Projection histogram 8 for scratches, scratches, noise, etc.
This is because it is reflected in 03, ....

In order to judge whether or not the characters 801, ... Are included using the projection histograms 803, ..., the operator has to determine the threshold value in the projection histograms 803 ,. Therefore, the operation was very troublesome.

[0007]

SUMMARY OF THE INVENTION Therefore, the object of the present invention is
Another object of the present invention is to provide a character slicing device capable of accurately slicing a character without being affected by noise or the like and improving operability.

[0008]

SUMMARY OF THE INVENTION Therefore, a character cutting device according to the present invention is made to achieve the above-mentioned object, and an image input section for inputting an analog image including characters and this image input section. The binarized image data creating unit that creates binarized image data by binarizing the analog image input in step S1 and the binarized image data created by the binarized image data creating unit. By dividing the input data creation unit that creates input data by dividing each of the horizontal and vertical directions into pixels for each column, and performing neural network processing on each of the input data created by this input data creation unit, Neural network processing unit that recognizes whether a part of a character is included and outputs the position of the circumscribing frame of the character based on the recognition result by this neural network processing unit It is obtained by a that circumscribing frame output unit.

The neural network processing unit includes an input layer having a number of units equal to the number of pixels of the input data created by the input data creating unit, an intermediate layer having an arbitrary number of units, and a part of a character. And an output layer composed of one unit number for outputting information on whether or not to include. Further, the neural network processing unit may be composed of a recurrent neural network.

The term "pixel" as used herein is a broad concept from a pixel having the minimum resolution to a unit in which a plurality of these pixels are collected.

[0011]

The analog image containing characters is input from the image input unit and binarized for each pixel in the binarized image data creation unit to be binarized image data. This binarized image data is divided into pixels for each column in the horizontal and vertical directions.
Input data is composed of the pixels for each column.

On the other hand, in the neural network processing unit, the neural network is made to learn in advance so that it can be accurately recognized whether or not a part of the character is included corresponding to the input data. Therefore, when the input data is input to the neural network processing unit, it is recognized whether or not the input data includes a part of the character, and the circumscribing frame output unit outputs the position of the circumscribing frame of the character based on the recognition result. It

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing an embodiment of a character slicing device according to the present invention. FIG. 2 is an explanatory diagram showing the configuration and operation of the neural network processing unit in this embodiment. Hereinafter, description will be given based on these figures. Note that, in FIG. 2, for convenience of illustration, the input data X ₁
It described only when processing to X _n, the input data Y ₁ to Y
_{The m} is omitted.

The character segmentation device according to the present invention is a character a
An image input unit 10 for inputting an analog image b including 1;
Created by a binarized image data creation unit 12 that creates binarized image data c by binarizing the analog image b input by the image input unit 10 for each pixel, and by the binarized image data creation unit 12. An input data creation unit 14 for creating input data X _{1 to} X _n , Y _{1 to} Y _m by dividing the binarized image data c thus obtained into pixels for each column in the horizontal direction X and the vertical direction Y; A neural network processing unit 16 that recognizes whether or not a part of a character is included by performing neural network processing on each of the input data X _{1 to} X _n and Y _{1 to} Y _m created by the data creating unit 14. And the recognition results x _{1 to} x in the neural network processing unit 16.
_n, the position d of the bounding box a2 characters a1 on the basis of the y ₁ ~y _m
And a circumscribing frame output unit 18 for outputting

The image input unit 10 is composed of, for example, an image scanner, a CCD camera, etc., and photoelectrically converts an image and outputs it as an analog value. The binarized image data creation unit 12 is composed of, for example, an A / D converter or the like, and converts the analog value input from the image input unit 10 into a digital value.

The input data creating means 14, the neural network processing unit 16 and the circumscribing frame output unit 18 can be realized by, for example, a computer and a program for operating the computer. In this case, the computer may be configured to control the image input unit 10, the binarized image data creation unit 12, and the like. The neural network processing unit 16 can also be realized by a neurochip.

Further, the neural network processing unit 16
Is the input data X ₁ created by the input data creation unit 14.
To X _n and Y _{1 to} Y _m, the input layer 20 has the number of units equal to the number of pixels, and the intermediate layer 22 has an arbitrary number of units.
And information indicating whether or not a part of the character a1 is included (“1”,
Output layer 2 consisting of one unit that outputs "0")
4 and. Further, the neural network processing unit 16 is composed of a recurrent neural network.

The units of the input layer 20, the intermediate layer 22, and the output layer 24 are not restricted to symmetric coupling or unidirectional coupling,
They are asymmetric and recurrent. The recurrent neural network having such characteristics can accurately recognize input data containing noise. Note that the recurrent neural network itself is a known technique, for example, "Measurement and control" Vol.30.No.
4 (April 1991) pp.296-301. Also,
The number of units of the input layer 20 is the input data X _{1 to} X _n , Y ₁
It is a number corresponding to the number of Y _m and can be set arbitrarily. In this embodiment, n units are provided as shown in FIG.

In such a configuration, the input data X _{1 to} X _n , Y _{1 to} Y _m input to the input layer 20 and the output layer 2 are input.
Recognition result from 4 x ₁ ~x _n, in order to obtain the relationship between y ₁ ~y _m, determined in advance by learning a coupling strength between the units. That is, certain input data is input to the input layer 20.
When the input data is a large lump (a set of “1”), the output layer 24 outputs “1”. When the input data is a small lump, the output layer 24 outputs “0”. To learn to output.

By using the learned neural network processing unit 16, it is possible to accurately determine "0" or "1" from the output layer 24 when given unknown input data. . Since the neural network processing unit 16 learns the shape of the pattern, correct pattern recognition is possible even when some noise is mixed in the input pattern (“input data” in this embodiment).

Next, the operation of this embodiment will be described with reference to the flow chart of FIG. 3 and the drawings indicated below.

As shown in FIG. 4, the binarized image data c created by the binarized image data creating unit 12 is composed of n horizontal pixels × m vertical pixels. Input data creation unit 14
Creates the input data X _{1 to} X _n by dividing the binarized image data c into pixels for each column in the horizontal direction X (step 101).

Next, the neural network processing unit 1
Input data X _{1 to} X _n are sequentially input to the input layer 20 of _No. 6 (step 102). Then, the recognition result x ₁ ~x _n corresponding to the input data X ₁ to X _n are sequentially outputted from the output layer 24 (step 103). For example, the recognition result x ₁ with respect to the input data X ₁ as shown in FIG. 5 (b), FIG. 5
As shown in (b), the recognition result x for the input data X _2
2 , the recognition result x ₃ is output for each input data X ₃ as shown in FIG.

Next, similarly, input data Y ₁ in the vertical direction Y ₁
˜Y _m are created (step 104), and input data Y _{1 to} Y _m are sequentially input to the input layer 20 (step 10).
5), the recognition result y ₁ ~ corresponding to the input data Y ₁ to Y _n
y _m are sequentially output from the output layer 24 (step 10).
6).

[0025] Finally, circumscribing frame output unit 18, vertical and horizontal recognition result x ₁ ~x _n, circumscribing frame a character a1 from y ₁ ~y _m
The position d of 2 is obtained, and the character a1 is cut out in the circumscribing frame a2 (step 107). That is, as shown in FIG. 6, the recognition result x ₁ ~x _n, among the y ₁ ~y _m, respectively SX the position of the beginning portion of the output is "1" (indicated by hatching in the figure), and SY , The number of consecutive "1" is L
X and LY. As a result, the left end and the upper end (SX, S
Y) and a rectangle of size (LX, LY) are recognized as the circumscribing frame a2, and the character a1 is cut out in the circumscribing frame a2. As shown in FIG. 1, the cut-out character a1 is pattern-recognized by a character recognition unit 30 outside the character cutting device.

The neural network processing unit 16
It is most preferable to use a recurrent neural network as in this embodiment, but the present invention is not limited to this, and a feedforward network or the like may be used.

[0027]

According to the present invention, the input image data is processed by the neural network processing unit by dividing the binarized image data into pixels for each column in the horizontal direction and the vertical direction and creating the input data. can do. As a result, it is possible to recognize whether or not the input data includes a part of the character by removing the influence of noise, and it is possible to cut out the character with an accurate circumscribing frame based on the recognition result.

Further, conventionally, the operator had to determine the threshold value in the projection histogram, but in the present invention, such an effort is not required when cutting out a character, so that the operability can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing the configuration and operation of a neural network processing unit according to an embodiment of the present invention.

FIG. 3 is a flowchart showing the operation of the embodiment of the present invention.

FIG. 4 is a plan view showing binary image data including characters.

[Figure 5] is an explanatory diagram showing the operation of the neural network processing unit according to an embodiment of the present invention, FIG. 5 (b) is a recognition result for the input data X ₁ x _1, FIG. 5 (b) is the input data X ₂ recognition result x ₂ for 5 (c) shows a recognition result x ₃ with respect to the input data X _3.

FIG. 6 is an explanatory diagram showing a recognition result of the neural network processing unit in the embodiment of the present invention.

FIG. 7 is an explanatory diagram showing an example of cutting out characters.

FIG. 8 is an explanatory diagram showing a character cutting method in a conventional example.

[Explanation of symbols]

 10 image input unit 12 binarized image data creation unit 14 input data creation unit 16 neural network processing unit 18 circumscribing frame output unit 20 input layer 22 intermediate layer 24 output layer

Claims (3)

[Claims] 1. An image input unit for inputting an analog image including characters, and a binarized image for creating binarized image data by binarizing the analog image input by the image input unit for each pixel. A data creation unit, an input data creation unit that creates input data by dividing the binarized image data created by the binarized image data creation unit into pixels for each row in the horizontal and vertical directions, and A neural network processing unit that recognizes whether or not a part of a character is included by performing neural network processing on each of the input data created by the input data creation unit and the recognition result by this neural network processing unit. And a circumscribing frame output unit that outputs the position of the circumscribing frame of the character based on the character slicing device for character recognition. 2. The neural network processing unit,
An input layer composed of a number of units equal to the number of pixels of the input data created by the input data creation unit, an intermediate layer composed of an arbitrary number of units, and a single unit for outputting information as to whether or not a part of a character is included. The character slicing device for character recognition according to claim 1, further comprising an output layer composed of a number of units. 3. The neural network processing unit,
The character slicing device for character recognition according to claim 1 or 2, comprising a recurrent neural network.