JPS63208181A - Character recognizing device - Google Patents

Abstract

PURPOSE:To recognize character only with a general microprocessor by dividing a pattern to partial patterns based on the continuity of the optically read pattern and recognizing a character in accordance with feature and connection of partial pattern. CONSTITUTION:The document read by a scanner 15 is stored in a RAM 13 as black-and-white binary image data. In this case, character pattern data of the whole of a character pattern consisting of gathering of run data indicating mutual connection relations is obtained when one raster scanning is terminated. This character pattern data is divided into partial patterns and is converted to a node data structure of connections of node data. The code of this node data structure is compared with codes of standard patterns in a dictionary part 28, and a character or the like of the standard pattern whose code coincides with the code of the node data structure is defined as the recognition result.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character recognition device, and particularly to a character recognition device that reads and recognizes printed or handwritten characters, symbols, etc.

[Prior Art] Conventionally, character recognition devices generally read characters optically and extract a plurality of characteristic quantities from a binary character image.
Recognition was performed by calculating similarity and distance from standard patterns in a dictionary that stored feature quantities extracted from learning data in advance.

However, in most cases, these methods require processing to scan almost every pixel of a character image many times for feature extraction, or require multiple sum-of-products operations for similarity and distance calculations. Therefore, dedicated hardware circuits such as a feature extraction circuit and a similarity/distance calculation circuit are required to shorten processing time, resulting in an expensive device.

[Problems to be Solved by the Invention] The present invention solves the above-mentioned drawbacks and is a low-cost system that can be implemented using only a general-purpose microprocessor without requiring dedicated hardware circuits for feature extraction or distance calculation. Provides a character recognition device. Further, the present invention provides a character recognition device whose recognition algorithm is easy to improve and maintain, and whose recognition rate is improved.

Furthermore, a character recognition device capable of dealing with character deformations is provided.

[Means for Solving the Problem] As a means for solving this problem, the character recognition device of the present invention extracts the characteristics of an optically read pattern and extracts the characteristics of the character stored in advance. A character recognition device that recognizes a character by comparing it with a character, comprising: a continuity extraction means for extracting continuity in a predetermined direction of the pattern; and a character recognition device that recognizes the pattern based on the continuity extracted by the continuity extraction means. a pattern dividing means for dividing into partial patterns; a feature/connection storage means for storing the pattern and storing characteristics and connections of the partial patterns; character recognition based on the characteristics and connections of the partial patterns; and character recognition means.

Also, a character recognition device that extracts the characteristics of an optically read pattern and recognizes the character by comparing it with the characteristics of the character stored in advance, the character recognition device extracting the continuity of the pattern in a predetermined direction an extracting means, a pattern dividing means for dividing the pattern into partial patterns based on the continuity extracted by the continuity extracting means, and a pattern dividing means for dividing the pattern into partial patterns based on the thickness of the partial pattern. division correction means for correcting; feature/connection storage means for storing the pattern and storing the characteristics and connections of the partial patterns; and character recognition based on the characteristics and connections of the partial patterns. recognition means.

[Operation] In such a configuration, based on the continuity of the optically read pattern extracted by the continuity extraction means, the pattern division means divides the pattern into partial patterns, and the divided patterns are stored in the feature/connection storage means. Characters are recognized by comparing the characteristics and connections of each of the partial patterns thus obtained as pattern characteristics with character characteristics stored in advance by a character recognition means.

Further, based on the continuity of the optically read pattern extracted by the continuity extraction means, the pattern division means divides the pattern into partial patterns, and each partial pattern is stored in the feature/connection storage means. Characters are recognized by comparing the characteristics and connections of the partial pattern with character characteristics stored in advance in a character recognition means. Based on this, the division of the pattern dividing means is corrected.

[Example] Hereinafter, the present invention will be described in detail with reference to the drawings. Note that the term "characters" in this embodiment also includes other patterns such as symbols.

FIG. 1 is a hardware configuration diagram of the character recognition device of this embodiment. 10 is a CPU for executing processing procedures, and 11 is a ROM for storing a dictionary.

12 is a ROM for storing processing procedures; 13 is a RAM for working in processing procedures and as a buffer for storing image data for one page of a manuscript; 14 is a scanner interface;
5 is a scanner that optically reads the original and converts it into an electrical signal to create black and white binary image data; 16 is an interface for outputting the recognition result to the outside.

FIG. 2 is a schematic logical block diagram showing the processing of this embodiment. Reference numeral 20 denotes a manuscript on which characters are written, 21 a photoelectric conversion unit that converts optical information into electrical information, 22 a binarization processing unit that binarizes multivalued information, 23 a character cutting unit that cuts out character frames, and 24 25 is a run extraction unit that extracts a continuous column of black pixels (called a run) for each row; 25 is a node data structure creation unit that analyzes the structure of a character from the continuity of the runs; 26 is a code for the node data structure. A coding unit 27 is an identification unit that compares the code created by the coding unit 26 with a code stored in advance in the dictionary unit 28 to identify the corresponding character; 28 is a dictionary that stores codes and characters in correspondence; Department.

10(a) and 11 stored in the ROM 12.
The processing procedure will be explained along the processing flowchart of FIG. When a document is read by the scanner 15, photoelectric conversion and binary processing are performed within the scanner 15, and black and white binary image data is transferred to a buffer within the RAM 13. The flow up to this point is not illustrated.

In step 5100 of FIG. 10(a), a character image of one character is cut out and stored in the RAM 13 using a known technique to form a character.

Figure 3 shows an example of the formed character image of ``A''. Here, black and white binary data is stored at each address in the RAM 13, and Figure 3 schematically shows this. The following description will be made assuming that the horizontal direction is the x@ axis and the vertical direction is the y axis in FIG. Detects the change point and extracts the run for each row.R in Figure 3
1 to R20 are run numbers assigned to each run.

Furthermore, when extracting a run, if the oblique shadows of the run in the previous row and the run in the current row on the X axis overlap, these runs are defined as being connected in the vertical direction, and the connection status of all runs is checked. investigate. Here, one run will be expressed using the data structure (referred to as run data) shown in FIG. In FIG. 4, Pl.

P2. P3 is a pointer to the connected run data in the upper row, and Pi, P2 . This corresponds to a run from left to right in the order of P3, and if the pointer value is O, there is no subsequent run data to connect.

Similarly, R4, R5, and R6 are pointers to connected run data in the lower row. At this point, the work flag 1 is OFF.

When one rask scan of a character image is completed in this way, the data structure of the entire character pattern (referred to as character pattern data) consisting of a collection of run data showing mutual connection relationships as shown in Fig. 5 is created. can get. The character pattern data is data that can faithfully reproduce the original character image.

In FIG. 5, 50 is an array of pointers to run data (R1 and R7 in this example, referred to as head runs) that do not have pointers to runs connected to the previous row.

Steps 5102 and 5103 are loop steps in which node generation processing is performed for each pointer to each head run in the array 50, and step 5103 calls a recursive function node generation subroutine using the run data of Heradran as an argument.

Here, the recursive function node generation process will be explained with reference to FIG. 11(a).

First, in step 5200, it is checked whether the work flag 1 of the run data passed as an argument is ON, and if it is ON, the equivalent number is returned with the function value set to "0". If the work flag 1 is OFF, the process advances to step 5210, where the node data area shown in FIG. 6 is secured, and the work flag 2 of the node data is cleared to ``0''. In step 5220, the run length (end point coordinates - starting point coordinates + 1) of the run data is compared with a predetermined threshold for run length determination (for example, 4), and if it is greater than or equal to the threshold for run length determination, the node data Set the node type to 1" (indicating long), and if it is less than the threshold for run length judgment, set it to 0" (
) to indicate that it is short.

In step 5230, step 5240→5250a
→5260a→Set the address of the run data in the run pointer of the loop counter 5280, and assign the run length of the run data to variable A.

Next, in step 5240, the work flag 1 of the run data pointed to by the run pointer is turned ON. Step 5250
In a, the run length change amount V is calculated using the following equation.

Assuming that the run length of the run data pointed to by the run pointer is 1', if (42>1') then v= (j2* too) /statement' 1se V= (j2'*100) /It In step 5260a, ■ is predetermined. It is checked whether the run length change rate threshold (for example, 11o) is exceeded, and if it is not, it is assumed that the run lengths match, and the process proceeds to step 5280, where the run pointer is set to the address of the next run data, i.e. The value of R4 of the run data pointed to by the run pointer is set, and the process returns to step 5240 and loops. If V exceeds the run length change rate threshold in step 5260, it is assumed that the continuity of the run data has been interrupted, and step 5
Proceeding to 270, the value of the run pointer is set to the last run data address of the node data, the work flag 2 of the node data is cleared to 0'0', and the node data is completed.

Next, in step 5290, it is checked whether the run data pointed to by the run pointer is child run data (that is, whether P4 to P6 of the run data pointed to by the run pointer are not 0''), and if there is no child run data. Return with the address of the node data as the function value. If there is child run data, use the child run data as an argument and return to step 5300.
Call the node generation function recursively. Step 53
In step 10, the node data pointed to by the function value of the function read in step 5300 is set as a child, and the node data is linked to each other using a pointer as a parent. Next, the process returns to step 5290 to check whether there is next child run data, and if not, returns the address of the node data as the function value.

In this way, the character pattern data of FIG. 5 is converted into a node data structure as shown in FIG. 7, which is a connection of node data shown in FIG. 6(a). Figure 6(a)
Ql, Q2. Q3 is a pointer to the parent node, Q4.
Q5. Q6 is a pointer to a child node. The square blocks in FIG. 7 are node data, and N1 to N5 are node names and correspond to N1 to N5 in FIG. 5. It is written in the lower half of the square block, and S is the node type.

L stands for long and S stands for short.

Returning to FIG. 10(a), if there are no unprocessed heradrans in step 5102, the process proceeds to step 5104, where each note data in the node data structure is numbered. The numbering is done by scanning the headruns in order from left to right in the headrun pointer array 50, tracing the child nodes of Q4 downward from the headrun, and from where they arrive, the sister nodes,
Work backwards in order of parent nodes. FIG. 8 is a numbered representation of the node data structure of FIG. Next, in step 5105, the node data structure is encoded by converting each node into a 32-bit code in the order of the node number. As shown in FIG. 9, the nodes are encoded by representing the node number, node type, and node numbers Q1 to Q6 with 4 bits each.

Finally, in step 5106, the code of the node data structure determined in step 5105 is compared with the code of the standard pattern in the dictionary section 28, and the characters of the matched standard pattern are taken as the recognition result.

In the embodiment described above, a change in run length is detected by determining whether the ratio of the run length to the previous run exceeds a predetermined threshold, but other embodiments include the following.

Assume that the nth previous run is R1, and the run length of Rn is 2°. The function f(Ml, 2...dan,) is one previous run R,
~The run length of the next run (referred to as RO) of R1 from the run length of R8. A function that predicts the run length of the run RO. and f(
Dan 1. Cross 2. . . . A change in run length is detected based on whether the difference between the two values (r) is larger than a constant value d. Examples of the function f include the following.

if C1t >1+ ) thenf(dan 11
Sentence 2) Once, - (dan, once,) 1se f (intersection 8.dan,)=dan, + (dan, once,) This embodiment judges whether the run length is longer or shorter than a fixed threshold value. Rather than assuming that it is short or short, we have determined that it is relatively long or short relative to the run length of the previous run, so the short run below the long run resulting from the slope of the horizontal stroke is the same as the long run above. By making it belong to a node, the recognition rate is greatly improved because it takes measures against the slope deformation of the horizontal stroke and absorbs noise near the stroke, and it also eliminates the need to perform pixel-by-pixel tracking processing on all character images. To enable recognition by processing only run data. Therefore, a character recognition device can be realized without requiring any special dedicated hardware circuit.

Next, an example will be shown in which the node data has its attribute value. FIG. 6(b) is a flowchart of node data having attribute values, and FIG. 11(b) is a flowchart of a node generation subroutine to have attribute values. Note that in the flow of FIG. 11(b), parts that are the same as those in FIG. 11(a) are not shown.

In step 5250b, the number of black pixels of the run data pointed to by the run pointer (i.e., run length =
Add the ending point coordinates - starting point coordinates + 1).

In step 3260a, the run length of the run data pointed to by the run pointer is compared with a run length determination threshold to determine whether it is long or short, and it is determined whether it matches the node type of the node data. If they match, the address of the next run data (that is, the value of P4 of the run data pointed to by the run pointer) is set in the run pointer and the process returns to step 5240; if they do not match, the vertical continuity of the run has been interrupted. It is regarded as such and proceeds to step 3270. In step 5270, the value of the run pointer is set in the run data address at the end of the node data.

In FIG. 6(b), Ql, Q2. Q3 is a pointer to the parent node, Q4. Q5. Q6 is a pointer to a child node, and the attribute value includes the number of black pixels included in the node.

In this case, the last step 5106 in FIG. 10(a)
For identification, the code of the node data structure determined in step 5105 is compared with the code of the standard pattern in the dictionary section 28, and characters of the matched standard pattern are selected as recognition candidates. The dictionary unit 28 stores a set of a code of a node data structure, an attribute value of each node (average number of black pixels included in a node), and a character code for the number of types of characters.

Furthermore, characters that were candidates for recognition (i=1.2°...,
For n), the distance of the observed character pattern is calculated using the following formula.

DI=Σlxk ai+'1 kg+ Here, Xk (k=1...m) is the attribute value of the second node in the node data structure of the observed character pattern, a,
l is the attribute value of the second node of the first candidate.

(D+ l i = 1, ---n), the minimum distance 1I
Characters having iItDI are recognized as recognition results. Thus,
Candidates were narrowed down using the code of the node data structure, and then identified by comparing the number of black pixels included in the nodes of the node data structure.

In the embodiment described above, the number of black pixels included in a node is used as the attribute value of the node, but other similar embodiments can be considered in which the following quantities are used as the attribute value.

(1) Set the stroke direction as an attribute value.

Since each node is a collection of runs connected vertically,
When the node type is short (run length is short), the most approximate straight line to the collection of midpoints of each run is found, and the slope of this straight line is used as the attribute value. A collection of runs constitutes a stroke, so the slope of the straight line is in the direction of the stroke.

(2) Set the position of the contact point as an attribute value.

In the case of a run where the parent node type is long and the child node type is short, both nodes are approximated by line segments, the parent node line segment is divided into three parts: left, center, and right, and the child node line is Find out which three parts of the line segment of the parent node the portion is in contact with, and use the position of this contact point as the attribute value.

In this example, the node data structure and node attribute values can be determined by scanning the character image only once, which is easier than other methods that trace the character image pixel by pixel and extract features. Since character recognition can be performed simply by matching codes and comparing node attribute values, character recognition is possible without the need for special hardware circuits and with a standard configuration of a general-purpose microprocessor. Since the device can be realized, a low-cost character recognition device can be provided.

Furthermore, as shown in FIG. 11(C), if the process jumps from step 5240 to step 5260 and compares the run lengths using Hayabusa and S, the process becomes simple. In addition, we found points of change in run length and turned each run into a node as a break in the continuity of the run, but as another example, we can use the merging point, branching point, or inflection point of the strokes of a character pattern as a node. Since these points can be found from the connection state, it is conceivable to node these points as breaks in the continuity of the run.

As shown in the above embodiment, by scanning the character image once, character pattern data consisting of a set of run data is created, and thereafter the character pattern data is
Feature extraction can be performed simply by processing. NxN character images
(pixels), then in other general character recognition methods, N2
In contrast, the amount of processing required is proportional to the number of pixels, whereas in the present invention, only runs are processed, so the amount of processing is significantly reduced.

Further, recognition does not require distance calculation, and only simple code matching is required. Therefore, a character recognition device can be realized with a standard configuration of a general-purpose microprocessor without the need for a special dedicated hardware circuit, so that a low-cost character recognition device can be provided.

FIG. 10(b) is a flowchart in which a step 5110 for correcting the node data structure is inserted before node numbering in step 5104.

If there are no unprocessed heradrans in step 5102, the process advances to step 5ilo, and a node data structure correction subroutine is called.

Next, the node data structure correction process will be described with reference to FIG. In FIG. 7, the head nodes that have no upwardly connected nodes are N1 and N3.

Steps 5400 and 5410 are loop steps that process all Herad nodes. If there is an unprocessed Herad node in step 5400, the process proceeds to step 34''10, where the nodes are traced in order from the head node downwards, and for each node, the sum of the number of black pixels of all runs included in the node is calculated. can be calculated by tracing the run data included in the node, since the node data includes the first run data address and the last run data address as shown in FIGS. 6(a) and ('b).

If the sum of the number of black pixels is smaller than a certain value, this node is caused by noise or splashes and should not occur in the first place, so the node data is changed from the node data structure as shown in an example in Figure 13. delete. Then, pointers Q1 to Q6 of node data to be connected to this node data
Correct. However, even if the sum of the number of black pixels is smaller than a certain value, if the node data of interest is connected to two or more node data above or below (i.e., node data that is a confluence or branch point) is removed from the node data structure,
Not excluded.

If there is no unprocessed Herad node in step 5400,
Proceed to step 5420. Steps 3420 and 5430 are also loop steps that process all Herad nodes. If there is an unprocessed herad node in step 5420, proceed to step 5430 and trace all the nodes downward from the head node. Processing is performed to integrate separated node data into one node data. If there is no unprocessed Herad node in step 5420, the subroutine ends and returns.

In the above embodiment, the node is deleted when the sum of the number of black pixels of the runs included in the node is less than a certain value, but in another embodiment, when the number of runs included in the node is less than a certain number, the node is deleted. There is also a way to delete the .

According to this embodiment, a character recognition device capable of dealing with character deformations can be realized using only a general-purpose microprocessor without requiring special hardware circuits for feature extraction or distance calculation.

Note that in this example, character recognition was performed using black dots.
Black and white is relative, and recognition may be based on a white point, or a predetermined color or the like may be considered. Further, in this embodiment, character recognition is performed based on vertical and horizontal continuity, and in particular, the extraction result of vertical continuity is made more effective and simple. However, the vertical and horizontal directions are also relative, and the direction in this embodiment is not limited at all, and may be an oblique direction, a concentric direction, a radial direction, or the like.

[Effects of the Invention] According to the present invention, it is possible to provide a low-cost character recognition device that does not require dedicated hardware circuits for feature extraction or distance calculation and can be implemented using only a general-purpose microprocessor. In addition, it is easy to improve and maintain the recognition algorithm.
A character recognition device with improved recognition rate can be provided.

Furthermore, it is possible to provide a character recognition device that can deal with character deformations.

[Brief explanation of the drawing]

Fig. 1 is a hardware configuration diagram of the character recognition device of the embodiment, Fig. 2 is a logical block diagram of the character recognition device of the embodiment, Fig. 3 is a diagram showing character images, and Fig. 4 is run data. Figure 5 is a diagram showing character pattern data, Figure 6 (a),
(b) is a diagram showing node data; FIG. 7 is a diagram showing a node data structure; FIG. 8 is a diagram showing a numbered node data structure; FIG. 9 is a diagram showing coded nodes; Figure 10 (a
), (b) are flowcharts for character recognition, Figures 11 (a), (b), and (c) are flowcharts for node generation subroutines, Figure 12 is a flowchart for node data structure correction subroutine, and Figure 13 is a node diagram. It is a figure which shows the deletion and integration of. In the figure, 10.CPU, 11.. ROM for dictionary, 12..
...Processing procedure ROM, 13...RAM, 14...
Scanner interface, 15...Scanner, 16
...output interface, 20...manuscript, 21.
...Photoelectric conversion section, 22...Binarization processing section, 23...
Character extraction unit, 24... Run extraction unit, 25... Node data structure creation unit, 26... Encoding unit, 27...
- Identification section, 28... Dictionary section. Figure 1 □X Figure 3 Figure 4 Figure 6 (0) Figure 6 (b) Figure 10 (0) Figure 7 Figure 9 Figure 10 (b) Figure 11 (b) Figure 11 (C) Figure 12

Claims (15)

[Claims] (1) Extract the characteristics of the optically read pattern,
A character recognition device that recognizes characters by comparing them with pre-stored character features, comprising: continuity extraction means for extracting continuity of the pattern in a predetermined direction; and continuity extraction means for extracting continuity in a predetermined direction of the pattern; hand,
a pattern dividing means for dividing the pattern into partial patterns; a feature/connection storage means for storing the pattern and storing the characteristics and connections of the partial patterns; based on the characteristics and connections of the partial patterns; A character recognition device comprising: character recognition means for recognizing characters. (2) The continuity extraction means includes horizontal extraction means for extracting continuity in the horizontal direction, and vertical extraction for extracting continuity in the vertical direction based on the continuity in the horizontal direction extracted by the horizontal extraction means. 2. A character recognition device according to claim 1, further comprising means. (3) The character recognition device according to claim 2, wherein the horizontal extraction means extracts columns that also have horizontal continuity within one line as runs. (4) The character recognition device according to claim 3, wherein the vertical extraction means determines continuity based on whether the length of the run is longer or shorter than a predetermined value. (5) The character recognition device according to claim 3, wherein the vertical extraction means determines continuity based on a rate of change in run length. (6) The character recognition device according to claim 1 or 3, wherein the feature/connection storage means stores a run length as a feature of the partial pattern. (7) The character recognition device according to claim 1, wherein the feature/connection storage means stores the sum of black dots in the partial pattern as a feature of the partial pattern. (8) Extracting the characteristics of the optically read pattern,
A character recognition device that recognizes characters by comparing them with pre-stored character features, comprising: continuity extraction means for extracting continuity of the pattern in a predetermined direction; and continuity extraction means for extracting continuity in a predetermined direction of the pattern; hand,
a pattern division means for dividing the pattern into partial patterns; a division correction means for correcting the division of the pattern division means based on the size of the partial pattern; and a division correction means for storing the pattern and determining the characteristics of the partial pattern. A character recognition device comprising: a feature/connection storage means for storing connections; and a character recognition means for recognizing characters based on the features and connections of the partial patterns. (9) The continuity extraction means includes horizontal extraction means for extracting continuity in the horizontal direction, and vertical extraction for extracting continuity in the vertical direction based on the continuity in the horizontal direction extracted by the horizontal extraction means. 9. The character recognition device according to claim 8, further comprising means. (10) The character recognition device according to claim 9, wherein the horizontal extraction means extracts columns that also have horizontal continuity within one line as runs. (11) The character recognition device according to claim 10, wherein the vertical extraction means determines continuity based on whether the length of the run is longer or shorter than a predetermined value. (12) The character recognition device according to claim 10, wherein the vertical extraction means determines continuity based on a rate of change in run length. (13) The character recognition device according to claim 8 or 10, wherein the feature/connection storage means stores a run length as a feature of the partial pattern. (14) The character recognition device according to claim 8, wherein the feature/connection storage means stores the sum of black dots in the partial pattern as the feature of the partial pattern. (15) The division correction means includes partial pattern deletion means for deleting the partial pattern when the sum of black points of the partial pattern resulting from division by the pattern division means is less than a predetermined value; 9. The character recognition device according to claim 8, further comprising partial pattern reconstruction means for reconstructing a pattern.