JPS63103392A - Character recognizing system - Google Patents

Abstract

PURPOSE:To recognize characters with high precision in a relatively simple constitution by dividing a character image into areas most suitably for the character image to extract stable features. CONSTITUTION:The character image on a document is read by a scanner 1 and is binarized by a preprocessing part 2 and is subjected to normalization and smoothing processings. A feature extracting part 3 adds direction codes to the character outline part of the binary character image preprocessed by the preprocessing part 2. The feature extracting part 3 divides the character image into mesh areas. This division is so determined that direction code giving picture elements are distributed to respective mesh areas as uniformly as possible. The histogram of each direction code in respective mesh areas is obtained by the feature extracting part 3. A recognizing part 4 outputs a recognized character code in accordance with the histograms of respective direction codes extracted from the input character and histograms registered in a dictionary 5.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field] The present invention relates to a character recognition system, and more particularly to an improvement of a character recognition system that uses a histogram of direction codes attached to the contours of character lines for recognition.

[Prior art]

As a recognition method for printed characters, a 5-character image is divided into areas of 4×4 mesh size, the number of meshes that make up the character part in each area is calculated, and if the count value is 1 or more, 1 is subtracted, and the subsequent There is a method that performs matching with a dictionary using values (Japanese Patent Laid-Open No. 110191/1983). Although this method has the advantage of simplifying the recognition device, it cannot be applied to the recognition of handwritten characters that are highly deformed because the size of the region and the division position are fixed.

On the other hand, the handwritten letter L has a large deformation! Various weaving methods have been devised, including ones that utilize stroke characteristics, but all of them have the problem of complicating the recognition device.

Furthermore, in character recognition, features are often extracted by dividing a character image into regions, as in the conventional example. As methods for such area division, there are two methods: (1) fixing the division points as in the conventional example, and (2) the document "Oki Electric Research and Development" December 1980, No. 121, vol. 1.5 ONa3, pp. '77. There is a known method of determining dividing points for each character by using the center of gravity of the peripheral distribution of a character image, as shown in No. 1-82.

The above method (2) is effective in stabilizing features when applied to recognition of handwritten characters that are highly deformed.

However, since the center of gravity of the peripheral distribution is used, the amount of calculation required for region division becomes large, making high-speed processing difficult.

On the other hand, method ■ does not require any special calculations.

If the characters are deformed, the area division becomes inappropriate.
There is a problem that character features cannot be extracted correctly. That is, even if a divided area includes a character line when the deformation is small, if the character is deformed, the character line may no longer fit into the divided area, making the characteristics unstable. for example,
The only clue to identifying the kana characters ``tsu'' and ``wa'' is the presence or absence of a protruding part at the top of the character, and this information must be stably extracted based on the domain and reflected in the feature vector. However, with fixed division, when the ``tsu'' is deformed, the upper protrusion may be included or not included in the specific area, making it easy to misrecognize it as a ``wa''.

〔the purpose〕

The present invention has been made in view of the above-mentioned problems, and provides a character L2m method that can recognize printed and handwritten alphabetic characters, numbers, symbols, kana, etc. with high precision, and can be implemented using a device with a relatively simple configuration. The purpose is to provide.

〔composition〕

The character recognition method of the present invention attaches a direction code to the character line outline of a binary character image, and at the same time and after that, a histogram H and HYj, divides the character image in the X direction and Y direction into a plurality of mesh areas including direction coded pixels almost evenly based on the total number and the histogram HXi, HYj, and assigns a direction code to each mesh area. This method is characterized by obtaining another histogram and calculating the distance between the histogram and the dictionary's histogram to recognize characters.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a schematic flowchart of character recognition processing according to the present invention, and FIG. 2 is a schematic block diagram of an optical character recognition apparatus (OCR) that performs character recognition according to the same flowchart. Contents common to each of the embodiments to be described later will be described here.

A character image of a document is read by a scanner 1, binarized by a preprocessing section 2, cut out into individual characters, and subjected to normalization and smoothing processing. Feature extraction part 3
adds a direction code to the character line contour of the binary character image preprocessed by the preprocessing unit 2. This direction code is 1
There are two methods: one method is to attach it to black pixels of the character line outline, and the other method is to attach it to white pixels. When attaching a direction code to a black pixel, a pattern of five pixels in total, including the black pixel of interest and the pixels above, below, left and right of it, is observed, and the direction code shown in FIG. 3 is attached. When attaching a direction code to a white pixel, observe a pattern of five pixels in total, including the target white pixel and the pixels above, below, left, and right of the target white pixel, and attach the direction code shown in FIG. 4. Third
In each of the 5-pixel patterns in the figure and FIG. 4, the diagonal grid indicates black pixels, and the open grid indicates white pixels.

In addition, in the method of attaching direction codes to black pixels.

Character line parts with a line width of 2 or more are given a direction code on each edge, whereas character line parts with a line width of 1 are given a common direction code for each edge, and the latter is different from the former. The number of direction codes will be halved. This means that in the case of a "faint" character in which a character line portion with a line width of 1 occurs, the feature amount becomes unstable. Such a problem can be avoided by adding a direction code to a white pixel.

Therefore, if a method of attaching direction codes for white pixels is adopted, characters will be more resistant to "fading". On the other hand, if a method of attaching direction codes to black pixels is adopted, the characters will be "squashed".
Strong against

Figure 5 shows the preprocessed image of the kana character "A". Figure 6 shows the result of adding direction codes to the outline black pixels of the character image.
In addition, FIG. 7 shows the result of attaching a direction code to one outline white pixel.

As will be described later, at the same time as the direction coding process, a histogram of the total number of direction coded pixels and the projection of the X direction axis and/or the Y direction axis is obtained and registered in the table. Of course, such counting of the total number of pixels and creation of a histogram table can be performed independently after adding direction codes, but in addition to simultaneous execution, the number of scans for one character image is reduced, which reduces processing time. It has the advantage of being able to shorten the time.

Next, in the feature extraction section 3, the character image is changed to "X".

It is divided into N×M mesh regions in the Y direction. The dividing position is N×
The direction code-applied pixels are determined to be distributed as evenly as possible in the M mesh areas. By doing this, it is possible to divide the mesh area in a way that is compatible with the deformation of the characters, and it is possible to solve the above-mentioned problem when dividing at a fixed position.

Next, in the feature extraction unit 3, each mesh region (II
A histogram Hkij for each direction code is obtained for each direction code. Here, k is a direction code (1 to 8).

The recognition unit 4 uses the direction code-specific histogram extracted from the input character by the feature extraction unit 3 and similar histograms registered in the dictionary 5 to calculate the distance between the input character and each dictionary registered character, Output the character code with the smallest distance as the recognized character code 0 For example, considering simple Euclidean distance, the distance dQ between the character Ω, the input character histogram DkijQ, and the input character histogram Hkij is dQ=ΣΣΣ1DkijQ−Hkij lik expressed.

Hereinafter, the unique contents of each embodiment will be explained in order.

Embodiment 1 After pre-processing, a binary input character image is subjected to direction coding processing (characteristic processing) according to a certain algorithm.

At the same time as this process, the total number (PE) of direction coded pixels (pixels with characteristics) is counted, and histograms HXi and HYj of the projection of the direction coded pixels onto the X and Y axes are calculated. It will be done.

A specific example of this histogram is shown in FIG. 8. In this figure, 20 is the processed image with a direction code, the number string at the top end indicates the lower one digit of the X address, and the number string at the left end shows the Y address.
It shows the address, and the other numbers are direction codes.

In this embodiment and the embodiments to be described later, the horizontal direction is the X direction and the vertical direction is the Y direction, but the relationship may be reversed.

A histogram HXi of the projection of the direction coded pixels of this image 20 onto the X-direction axis and a histogram HYj of the projection onto the Y-direction axis are each stored in a table 2 of the type shown in the figure.
Registered on 1.22.

After such processing, the character image is divided into N parts in the X direction,
The area is divided into M areas in the Y direction, resulting in N×N mesh areas.

FIG. 9 is a schematic flowchart of the process of determining dividing points in the X direction. This process will be explained in order.

After initializing counter nyi and register P (step 100), counter i is incremented (step 104), and the histogram H corresponding to its value (i) is
The value of Xi is read from the table, added to the value of register P, and reset to register P (step 106).
).

It is checked whether the value (P) of register P exceeds (PE/N)Xn (step 108).

If the result of this determination is No, similar processing is repeated from step 104.

This process essentially involves raster scanning the image while incrementing the
It is compared with n. However, it is inefficient to repeat such a raster scan, and furthermore, a raster scan for determining the total number PE.

Therefore, in this embodiment, the histograms HXi, HYj and the total number PE are obtained at the same time in the processing stage for the direction coding accompanying the raster scanning of the image.
By using this, raster scanning is not performed and processing efficiency is improved.

Now, when P≧(PE/N) The X address obtained by subtracting is stored as the X address of the starting point of the (n+1)th X direction divided area (
Step 110).

Note that the starting point of the first X-direction divided area is the left end (X=0) of the character image, and the end point of the last, ie, Nth, X-direction divided area is the right end of the character image.

After step 110, the counter n is incremented (step 112), and processing for finding the next division point is performed.

Such processing ends when n=N in step 102.

Dividing points in the Y direction are found in the same way. That is, in this case, the counter j corresponding to the Y address is used instead of the counter i, and the histogram HYj of the projection onto the Y direction axis is referred to.

Further, the starting point of the first Y-direction divided area is the upper end of the character image, and the ending point of the M-th Y-direction divided area is the lower end of the character image.

In this way, dividing points in the X direction and Y direction are found, and the character image is divided into X by these dividing points.

It is divided into N×M mesh regions in the Y direction.

FIG. 10 shows an example of mesh area division of an image of handwritten kana characters 1-J, assuming that the overlapping amount T of the mesh areas is 1. The key to distinguish between "nuj" and "su" is the downward-sloping stroke to the right, and it is important to correctly reflect this feature in the mesh area. According to this embodiment, such conditions are satisfied and confusing characters such as "nu" and "su" can be correctly identified and recognized.

Example 2 In this example, direction coding is performed at the same time as processing.
The total number PE of direction coded pixels and a histogram HYj of the projection of the direction coded pixels onto the Y-direction axis are determined.

Then, by the same processing as in Example 1, the character image is Y
It is divided into M parts in the direction.

Next, each Y-direction divided area is raster-scanned while the area is divided into N parts in the X direction.6 FIG. 11 is a schematic flowchart of the process of dividing one Y-direction divided area into N parts.

To explain this process, after initializing counters n, i and register P (step 200), counter i is incremented (step 204), and the character image Y is The raster scan is performed in the direction, the number of direction coded pixels in that line is counted, and the value Pi is added to the register P (
Step 208). The value CP of register P is (PE/N
×M) It is checked whether Xn is exceeded (step 210).

If the result of this determination is NO, similar processing is repeated from step 204.

When P≧(PE/NxM) The address is stored as the X address of the starting point of the (n+1)th X-direction divided area (step 212).

Note that the starting point of the first X-direction divided area is the left end (X=O) of the character image, and the end point of the last, ie, Nth, X-direction divided area is the right end of the character image.

After step 212, the counter n is incremented (step 214), and processing for finding the next division point is performed.

Such processing ends when n=N in step 202.

In this way, dividing points in the X direction and Y direction are found, and the character image is divided into X by these dividing points.

It is divided into N×M mesh regions in the Y direction.

FIG. 12 shows the result of dividing the image of "nu" according to this embodiment. However, only the results of the X-direction division for the middle Y-direction region are shown. Also, T=0.

In the above explanation, the division in the Y direction was performed using the method of Example 1, but after the division in the X direction, the Y direction was
Directional division may also be performed.

Example 3 When direction coded pixels are concentrated in the overlapping area, X,
The variation in the number of pixels within the divided regions in each Y direction increases. In that case, PE/(N×M
), the region division may become inappropriate.

According to this embodiment, such a problem can be solved.

In this embodiment, simultaneously with the direction code assignment process, the total number PE of direction code assignment pixels and the histogram HYj of the projection of the direction code assignment pixels onto the Y axis are obtained, and then the same processing as in the first embodiment is performed. The character image is divided into M parts in the Y direction.

During this Y-direction division, the number Pm of direction code-applied pixels in each divided area is determined with reference to the histogram HYj.

Next, while raster scanning the inside of each Y-direction divided area, the area is divided into N parts in the X-direction.

FIG. 13 is a schematic flowchart of N-dividing processing for one Y-direction divided region.

To explain this process, after initializing counters n, i and register P (step 300), counter i is incremented (step 304), and the character image is moved in the Y direction for the X address corresponding to the value (i). The line is raster scanned, and the number of direction coded pixels in that line is counted.

The value Pi is added to register P (step 308
). Then, the value CP of register P is (Pm/N)Xn
It is checked whether the value exceeds (step 310).

If the result of this determination is NO, step 304 or similar processing is repeated.

When P≧(Pm/N) The address is stored as the X address of the starting point of the (n+1)th X-direction divided area (step 312).

Note that the starting point of the first X-direction divided area is the left end (X=O) of the character image, and the end point of the final, ie, Nth, X-direction divided area is the right end of the character image.

After step 312, the counter n is incremented (step 314), and processing for finding the next division point is performed.

Such processing ends when n=N in step 302.

At the division points in the X direction thus obtained, the Y direction divided area is divided into N parts in the X direction, and the image is divided into N×M mesh areas.

FIG. 14 shows the result of region segmentation of the handwritten "nu" image with overlap T=1. However, only the result of dividing the middle Y-direction divided area in the X-direction is shown.

Example 4 Add a direction code to the black pixels of the character line outline.

Other than this, the same processing as in the first embodiment is performed.

Example 5 Add a direction code to the black pixels of the character line outline.

Other than this, the same processing as in the second embodiment is performed.

Example 6 Add a direction code to the black pixels of the character line outline.

Other than this, the same processing as in the third embodiment is performed.

Each of the processing steps described above can be easily performed by hardware or software. Such processing execution means are
Those skilled in the art will be able to easily realize this from the above explanation, so a specific example thereof will be omitted.

〔effect〕

According to the character recognition method of the present invention described above, it is possible to perform optimal region segmentation according to the character image and extract stable features (orientation code and histogram) that are not easily affected by character deformation. letters, numbers, symbols,
Since characters such as kana can be recognized with high precision and the processing content is simple, character recognition can be performed with a relatively simple configuration.

[Brief explanation of the drawing]

FIG. 1 is a schematic flowchart showing the processing flow of the character recognition method according to the present invention, FIG. 2 is a schematic block diagram of an optical character recognition device for character recognition according to the present invention, and FIG. Figure 4 is a pattern correspondence diagram of the direction code added to the black pixels in the line axis outline. Figure 4 is a diagram showing the pattern correspondence of the direction code added to the white pixels in the character line outline. Figure 5 is the handwritten kana character "A" after pre-processing. Figure 6 is a diagram showing the result of processing to attach a direction code to the black pixels of the character line outline of the same pre-processed image, and Figure 7 is a diagram showing the result of processing to add direction codes to the black pixels of the character line outline of the same pre-processed image. FIG. 8 is a diagram showing an example of a histogram of direction coded pixels; FIG. 9 is a flowchart for explaining region division processing in the first embodiment; FIG. FIG. 10 is a diagram showing the results of region division in Example 1, FIG. 11 is a flowchart for explaining region division processing in Example 2, and FIG. 12 is a diagram showing the results of region division processing in Example 2.
FIG. 13 is a flowchart for explaining region division processing in the third embodiment, and FIG. 14 is a diagram showing the results of the region division processing according to the third embodiment. 1... Scanner, 2... Preprocessing unit, 3... Feature extraction unit, 4... Recognition unit, 5... Dictionary. :21 [! 1 Figure 2 Figure 3 Code 0125306 (1 Code 4 8 00 7 0
00 Figure 4 Code 01253060 Coat=4 8 0 0 7
0 0 0 Fig. 5 Fig. 6 Fig. 7 Fist - Lee - # - - - 222 - - - Italy - meea
eamam*Figure 11"222mgmmmm*eeemee*aa meetingee*a
asleepFigure 13

Claims (2)

[Claims] (1) Attach a direction code to the character line outline of a binary character image, and at the same time or after that, calculate the total number of direction coded pixels and histograms HXi and HYj of the projections onto the X-direction axis and the Y-direction axis, and dividing the character image in the X direction and the Y direction into a plurality of mesh regions including direction code-applied pixels almost equally based on the total number and the histograms HXi and HYj, and obtaining a histogram for each direction code for each mesh region; A character recognition method that recognizes characters by calculating the distance between the histogram and the histogram of a dictionary. (2) Attach a direction code to the character line contour of a binary character image, and at the same time or after that, calculate the total number of direction coded pixels and the histogram HXi of the projection onto the X direction axis;
The character image is divided in the X direction into N regions including direction coded pixels almost equally based on the total number and the histogram HXi, and at the same time or after that, each of the X
The number of pixels to which a direction code is applied in each direction divided area is calculated, and the number of pixels to which a direction code is applied is calculated by raster scanning in the X direction while increasing the Y address by one in each of the X direction divided areas. Divide the direction divided region into M regions including direction code-applied pixels almost equally, and obtain a histogram for each direction code for each of the N×M mesh regions of the character image thus divided; A character recognition method that recognizes characters by calculating the distance between the histogram and the histogram of a dictionary.