JPS60164882A - Character recognition device - Google Patents

Abstract

PURPOSE:To shorten a process time for dictionary collating by extracting character pattern features by inducing character groups common in features as candidate character and by storing start point information showing a search direction. CONSTITUTION:Characters on a form P are read optically by an A/D conversion circuit 1, converted into digital data and are stored in a picture memory 2. A CPU3 and stroke detection circuit 4 are connected to the circuit 1. The CPU3 performs character recognition process including collating process based on a program, and the circuit 4 converts the read characters into four direction stroke. An ROM5 is a memory for a dictionary storing standard patterns of the features to be recognized and for a program. An RAM6 is a memory for a work area used in storing a four direction stroke of input characters and eight direction conversion data and for program execution. The storage capacity of the RAM is stored corresponding to an image after stroke extraction.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field of the Invention The present invention relates to a character recognition device that automatically reads and recognizes unknown characters such as kana, alphabets, and numbers.

<Background of the Invention> Conventionally, in this type of character recognition device, as shown in the flowchart of FIG. Next, the above binary pattern is subjected to pre-processing to make the subsequent processing more effective, i.e. noise processing such as black dots on paper P on which characters are recorded, and character graphics. A series of processing is performed including smoothing of the boundary surface. Next, feature extraction processing is performed to extract several features (intersections, branch points, number of loops, stroke length information, etc.) necessary for character recognition. Depending on this extraction result, some of the characters that have common features are narrowed down as candidate characters from a large number of characters. This narrowing down and refinement completes the first stage recognition, but if there are a plurality of candidate characters, detailed identification processing is further performed to select the only character from among them. This detailed identification process is generally called dictionary matching process. The dictionary here stores the characteristics of each character in a fixed series, and is usually constructed from a ROM. Therefore, the above-mentioned narrowing down of candidate characters is called dictionary guidance processing, which refers to selecting a specific group of characters from the dictionary based on the feature group obtained by feature extraction.

As described above, following this dictionary guidance process, the degree of matching between the standard feature series of each character (hereinafter referred to as standard pattern) stored in advance in the dictionary and the corresponding feature series of the recognition target character is detected. Verification processing is performed. Based on the result of this collation process, the read character is specified, or the recognition result is output as unrecognizable.

Next, an example of a dictionary will be explained.

FIG. 2(A) is a graphic representation of the standard pattern [)1 of the kana character "a" in FIG. In other words, in this dictionary, the character "A" has three end points ■ and 1
It is made up of 1 branch points (2) and 10 direction components (vectors) represented by number values. The vector number values shown here are for distinguishing which of the eight directions from 0 to 7, not shown in FIG. 3, the stroke direction of each character is. Therefore, in the above standard pattern D1, the letter "A" starts from the end point ■ and is 2, 1, 3, 4.
・An element that has vectors 5 in this order and then ends at branch point 0, and an element that starts from end point ■ and has vectors 2, 3, 4,
5 and an element having the end point ■ in this order. A diagram illustrating each of these elements is shown in Figure 2.
This is shown in the lower part of η.

Here, the more specific processing procedure of the character recognition device explained in the flowchart of FIG. 1 will be described in FIGS.
This will be explained using FIGS. 8 through 8.

Characters read from paper P are binarized and preprocessed, and then stored in the storage area of the system. The character pattern stored here is a thick one, as shown in FIG. 4, just like the handwriting.

However, this thick character pattern is subjected to so-called thinning processing and converted into information indicating the skeleton of the character as shown in FIG. In this conversion, the skeleton is normalized in four directions: AC (left and right), B (up and down), C (diagonally upward to the right), and D (diagonally downward to the right) as shown in FIG. The normalization carried out here in □ is to facilitate differentiation into the eight directional components shown in FIG. 3. Following this conversion process, a process to obtain a feature sequence of the read character is performed. This feature series is expressed in the same way as the standard pattern D1 in the dictionary (see FIG. 2), and as a result, the feature series 134, 34, and 34 as shown in FIG. 6 are obtained. Since this feature series indicates a pattern of the character to be read, that is, an unknown character, it is hereinafter referred to as an unknown pattern.

The unknown pattern obtained as described above is compared with the □ standard pattern of candidate characters obtained as a result of the dictionary guidance process for its degree of matching. Assuming that the unknown pattern shown in Figure 6 is compared with the seed yarn pattern of "A" shown in Figure 2 09, this matching process is performed by comparing the □ direction component in the standard pattern as shown in Figure 8. This is done according to certain regulations, such as whether the order matches the order of the direction components of the unknown pattern. In this example, the unknown pattern's directional components are all included in the standard pattern's directional components, and the number 111B also matches, so the unknown pattern is recognized and output as a kana character "A".

In general character recognition devices, character recognition is performed as described above, but as can be understood from the above explanation, the order of the directional components that make up a character is important in the matching process in character recognition. I understand. It can also be understood that the order of these directional components changes depending on which position of the character is used as the starting point.

By the way, in the conventional character recognition device, the position of the character to be used as the starting point for the directional component search is determined according to Fig. 9 (A
), it was fixed and □ as shown in (B). In FIG. 9(A), the search starts from the top of the character,
FIG. 9 shows an example in which the search starts from the left side of the character. Searching from fixed directions like these is intuitive and easy to understand for humans, but it is not necessarily appropriate when performing mechanical recognition.

For example, as shown in Figures 10 (A) and (11), if the starting points for both the characters "ya" and "nu" are searched from the upper side of the characters and the pattern elements of both are extracted, similar patterns of both characters can be extracted. The elements are transposed and both characters can be reliably identified.

However, as shown in Figures 11 (A) and (B), when the unknown pattern of the characters "YU" and "WO" is searched from above the characters and the elements of both characters are extracted, the elements of both characters and The order is the same, that is, the patterns of both characters are the same, and in such a case, it is not possible to distinguish between the two characters only by dictionary matching. To deal with this problem, conventional character recognition devices extract character features from other viewpoints and then add another identification process. For this reason, in conventional character recognition devices, additional processing time is required for the additional identification processing, which hinders speeding up of the processing. Furthermore, for example, if the character pattern shown in FIG. 20 is searched in the search direction 1; It is necessary to store the two types of search patterns that this accumulates in the dictionary,
The processing time for dictionary matching increases, and the capacity of the dictionary increases, resulting in an impediment to speeding up processing and reducing common prices.

<Objective of the Invention> An object of the present invention is to eliminate the drawbacks of the conventional character recognition device described above, shorten processing time for dictionary comparison, and provide a high-speed and highly accurate character recognition device.

<Configuration and Effects of the Invention> In order to achieve the object described above, the present invention stores, in a dictionary storage means for storing standard patterns, starting point side information indicating the scanning direction and stroke end point position information for each character. In this way, the unknown pattern is scanned during the matching process based on the starting point side information and the stroke end point position information.

Thus, by implementing the present invention, for example, when searching for the starting point of both characters in FIG. 12 from the bottom side (see FIG. 13), the extraction order of the directional components is the character "Y" first, then ■1■, then ■ 05
■, but in the case of the character ``wo'', the opposite is true,
Both characters can be reliably identified. Therefore, compared to the conventional method, which always scans from a fixed direction and starts searching from the end point of the first character encountered, characters can be identified more reliably, and multiple types of standard patterns are prepared for each character. There is no need for this method, and the invention achieves remarkable effects such as contributing to shortening of collation processing time and dictionary capacity.

<Description of Embodiments> FIG. 15 shows a character recognition device according to the present invention.
The character ``-'' is read by the A/D transformation circuit, converted into digital data, and stored in the image memory 2. A
A CPU 3 and a stroke detection circuit 4 are coupled to the /D conversion circuit 1, and the CPU 3 executes character recognition processing including collation processing according to a program, and the stroke detection circuit 4 detects read characters, for example, as the first character. It is converted into a four-direction stroke as shown in Figure 16. Also in the figure,
ROM 5 is a memory for a dictionary and program that stores standard patterns of characters to be recognized; The stored contents of the λAM 13 are stored corresponding to the image after stroke extraction shown in FIG. 16, and the specific contents are shown in FIG. 17.

In FIG. 17, TERM is a storage area for end point information of characters and figures, T1 represents one end point, SA represents a stroke in the incoming direction, and Sl represents its serial number in the A direction. Similarly, T2~
T4 indicates another end point, and SC indicates a stroke in the C direction. Also, R indicates the position of the end point of the character relative to the stroke; 1. indicates that the end point position is on the left side of the stroke, and similarly, k indicates that the end point position is on the right side. CIf KN is an area where connection information (inflection points) between strokes is stored, and in the illustrated example, the R (right side) of the A direction stroke of serial number S1 and the R (right side) of the C direction stroke of serial number S2. This means that they are connected as an inflection point. ASTM is an area where two-dimensional coordinate address information of the end point of the A stroke is stored, and C8TM is an area where two-dimensional coordinate address information of the end point of the C stroke is stored. Note that the area for storing the two-dimensional coordinate address information of the end point of the stroke is as follows:
Areas RSTM and DS'FM for storing end point address information of the B stroke and the D stroke are also provided, but are not shown here because the character examples in FIG. 16 do not include the B stroke and the D stroke. Also F ON
T is an area for storing stroke information in the stroke tracking process of an unknown pattern, and SCM is an area for storing data converted in eight directions of the stroke of the recognition target character, that is, the unknown character.

FIG. 18 shows an example of the dictionary portion of the contents stored in the ROM 5, and the illustrated example shows the character "U" in FIG. This dictionary first stores scanning starting point side information for each character storage area. In the figure, the starting point consecutive number ■ indicates that scanning is performed using the upper side of the character as the starting point, and similarly, the starting point consecutive number ■■■ indicates that scanning is performed using the left side, lower side, and right side, respectively. (See Figure 21). Therefore, the first
In the example shown in FIG. 8, the first scan is performed using the lower edge as the starting edge, and the second scanning is performed using the "-" side edge as the starting edge.

Next to this starting point side information, tuna information at the end point of the stroke is stored. In the figure, AL indicates tracking from the left end point of stroke A, similarly AR at the right end point of stroke A, BU and BD at the upper and lower end points of stroke Is, and C at the left and right end points of stroke C.
L.

Ctt, DL at the left and right end points of stroke D
, DR. In the example of FIG. 18, tracing is performed from the left end point of stroke A in both the first and second scans. Therefore, when an unknown character is scanned from the starting point side, if the first character end point encountered is the left end point of stroke A, the stroke of the unknown pattern is traced from that character end point.

FIG. 19 shows the collation processing operation of the apparatus according to the embodiment described in -.
We will explain the matching process when the results are narrowed down to .

First, in step 11 (hereinafter referred to as "5T11"), "Y" in ROM 5 is selected from the narrowed down candidate characters "YU".
Refer to the dictionary (FIG. 18) and load the starting point sequence number ■ of the first scan. Next, in the same way with 5T12, ROM5
After referring to the dictionary and loading the stroke end point position information AL of the first scan, determine the scanning mode from the lower side using the starting point sequence number ■ (STI 3''), which corresponds to the left end of the lower side. The address is stored in the scanning counter and initialized (sT14). Next, the image shown in FIG. 16 is scanned on the RAM 5 according to the scanning mode. In the illustrated example, as shown in FIG. At this time, for each scanning point, a comparison is made to see if the two-dimensional address of the stroke end point in the □ASTM area and C5TM area of RAM5 matches the coordinates of the scanning counter (ST15).Result of comparison Unless both coordinates match, the scan counter is updated and this coordinate address comparison is repeated (ST15-17).The contents of the scan counter and the end point two-dimensional address of RAM6 are, for example, coordinates (2, 28
), this stroke end point is used as a temporary starting point, and it is then determined whether this stroke end point is a character end point (sTIB). This coordinate (2, 28) is ASTM
It corresponds to the left end point of stroke number 2 of the incoming stroke. Information on this endpoint exists as information on the character endpoint T3 in the TERM area. Therefore, this stroke end point is set as the character end point, and this information (S
A, 82. I-) is stored.

Next, in ST]9, this character end point is compared with the stroke end point position information loaded from the dictionary to determine whether or not the stroke type and end point position match. In the case of this embodiment, the information loaded from the dictionary is A L, which means the left end point of the A stroke, so both of them match. Therefore, this character end point is used as the starting point for scanning, and the character end point mark ■ is stored in the RAM. (Stored in the SCM area of i.

By the way, if the Y coordinate address of the end point T3 in FIG. 16 is smaller than the end point T4, the end point T4 becomes the temporary starting point in scanning from the lower side. In this case, the end point T4 corresponds to the right end point of the stroke in the A direction,
This does not match the stroke end point position information loaded from the dictionary. In this case, the judgment of 5T19 is ・NO・,
The determination of 5T20 is “YES” and the scan counter is
Updates will be repeated until the contents reach the Y coordinate address of the end point T3.

If the determination at 5T20 is "No", it is assumed that a completely unintended character endpoint has been extracted, and the process is terminated, and the process is rejected as a match mismatch.

On the other hand, if the data stored in the FONT area is not found in the character endpoint storage area TERM, this indicates that the detected coordinates are not character endpoints, so control is passed to ST□ to update the scan counter. be transferred. .

Next, if a character endpoint is detected, the FO
Using the stroke information in the NT area, convert it into the eight-direction data shown in FIG. 3 (incoming stroke from left to right is direction 1), and store this direction data "1" in
8-way conversion data storage area SCM (sT2
1).

Next, in order to further search for connection information from the stroke stored in the FONT area, the FONT area is changed to the other stroke end point information (SA, S2°R) of this stroke. Then, it is determined whether the right end (6) of stroke number 2 of the detected stroke in the A direction is a character end point or an inflection point by referring to the data in the FONT area, and the character end point storage area TERM and the inflection point are checked. Check the registered contents in the storage area CHKN and check. In the illustrated example, from the TERM area (SA.

52, R), and it can be seen that the stroke end point after the change is the character end point (T4).

This result is written to the 8-direction conversion data storage area SCM.
(sT22). Next, it is determined whether there is the next starting point designation by referring to the dictionary of the ROM 50 character "Yu" (ST23), but in the illustrated example, as the second starting point information exists as shown in FIG. 18, After loading the next starting point serial number ■ at 5T24 and loading the second stroke end point position information AL at 5T25, control is transferred to 5T13 to perform the same scanning as in ``1''. In addition, in the operation described above, the strokes that have been converted in eight directions are deleted as appropriate after the conversion (the stroke numbers in the ASTM and C8TM areas of RAM 5J2 are converted to space codes). by this,
A stroke that has been detected once will no longer be detected twice.

In the second search, as in the first search, the scanning mode is determined based on the data specifying the starting point, that is, the starting point serial number (ST13), and the scanning counter is initialized (STI).
4). In the above example, the second starting sequence number is ■, so the second
As shown in FIG. 1, scanning proceeds from the left side to the right side of the upper side. Then, as in the first case described above, scanning is performed while updating the scanning counter, and R
The two-dimensional address data of the stroke stored in the AM6 (DASTM, C5TM area) and the scan counter are compared (ST15 to 5T17).

In the example shown in FIG. 16, a match is established at address (3, 6). This address (3,6) is.

It is stored in the ASTM area. Therefore, the end point 9 is stroke number 1 of the incoming stroke (SA) (
It can be seen that it is on the left side (I,) of St). Next, it is determined whether this stroke end point is a character end point, and if it is a character end point, that information is stored in the FONT area of the RAM 5 as a temporary starting point during scanning. Next, this character end point is compared with the second stroke end point information, and when they match, this character end point is used as the starting point for scanning, and a character end point mark 2 is stored in the SCM area of the RAM 5. Next, similarly to the first time, the stroke information in the FONT area is used to convert the 8-direction data direction 1, and this direction data pattern 1'' is stored as "1" in the 8-direction conversion data storage area SCM (
ST21).

Next, in order to further search for connection information from the stroke stored in the FONT area, the other stroke end point information of this stroke indicates the right end point of the PONT area (S
A, Sl, R). To check if there is a connection, use this F
TERM area and CHKN by referring to the data in the ONT area.
This is done by checking the area, but in the illustrated example, the stroke end point (SA, Sl, R) after the change is (sc).

It can be seen in the CHK N area that it is connected to S1, R). Therefore, the judgment of ST22's connection information check is "
Yes. Along with this, the stroke information of the FONT area is changed to this (SC).

After changing to 5T21 (Sl, R), control is transferred to 5T21.

After that, 5T21 and ST2 until the end point of the character is detected.
2 is executed repeatedly. As a result, in this illustrated example, ■1■
■14■ is stored in the eight-direction stroke conversion data storage area SCM of the RAM 6. Should I specify the next starting point again in ST23? However, in the illustrated example, only the starting point edges up to the second time are stored in the ROM 5, so the determination is "NO" here, and then the contents of the 8-direction conversion data storage area SCM in the RAM 5 are stored. The character “YU” in the dictionary of ROM5
The standard patterns of are compared (ST26). As a result of the comparison, if a match is found at a certain rate or higher, the input character is recognized as "Yu".

[Brief explanation of drawings]

FIG. 1 is a flowchart for explaining the outline of a general character recognition device, FIG. Figure 3 is a diagram showing the direction numbers when the directional components of the stroke of a character are normalized into eight directions, and Figure 4 is an example of the character "A".
Figure 5 is a diagram showing only the skeleton of the stroke by thinning the letter "A" in Figure 4, and Figure 6 is a diagram showing the pattern as it is optically read. ” is expressed by end points, directional components, and refraction points, 7th figure
The figure is a diagram for explaining the four-directional components, and Figure 8 is a diagram for explaining the character recognition operation by matching the standard pattern of the character "A" shown in Figure 2 with the characters shown in Figures 5 and 6. Figures 9(c) and 9(b) are diagrams explaining the starting point edge for starting literature search ° scanning, and Figures 10(a) and (b) are the characters "ya".
A diagram showing the eight-direction component pattern of each stroke when scanning starts from the upper side of "nu", Fig.
Figures 12(A) and 12(B) show the 8-direction component pattern of each stroke when scanning starts from the upper side of the characters "YU" and "WO". Figure 13 shows the 8-direction component pattern of each stroke in 181 when scanning is started from
14 is a diagram for explaining the case where scanning starts from the lower side of a character, FIG. A circuit block diagram of a character recognition device in which the invention is implemented, FIG. 16 is a diagram showing an example of characters input to the character recognition device shown in FIG. 15, and FIG. 17 is a diagram of a RAM of the character recognition device shown in FIG. A diagram showing an example of storage area arrangement, FIG.
A diagram showing an example of data arrangement for one character in the dictionary storage area of OM, FIG. 19 is a flow chart for explaining the operation of the character recognition device shown in FIG. 15, and FIG. 20 is a flow chart for explaining the disadvantages of the conventional method. 21 are diagrams showing the side on which scanning is started based on the starting point side information shown in FIG. 18. -1- B]I ``■ JP-A Showa GO-164882 (8) li-L-Sacrificial package 1 2I To I JP-A Showa GO-164882 (11) Minute 27 Figure■

Claims (1)

[Claims] A means for optically reading unknown characters to form a character pattern, a means for extracting features of the character pattern, a means for storing the extracted features, and a group of characters that share the same extracted features as candidate characters. a dictionary storage means for storing a unique standard pattern for each character, and predefining and storing starting point side information indicating a search direction for each character and end point position information of a stroke during the search; , a character recognition device that searches for features of an unknown pattern and compares each character of the derived character group with a standard pattern based on the starting point side information and the stroke end point position information 1. Device.