JP2784004B2 - Character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] The present invention relates to a character recognition device for recognizing input image data obtained by optically reading a character string in units of a character frame having a size corresponding to one character. The purpose of this method is to enable accurate character recognition even if it is out of the character frame of the size corresponding to, or is connected to adjacent characters. A dictionary in which candidate data including information on character candidates and attributes of the candidates are stored in advance, matching means for reading candidate data for input image data by matching the dictionary with the dictionary, and an attribute of two adjacent candidates A data table storing in advance validity data representing the validity of each candidate from the correlation, and a validity data corresponding to the data table indexed by the candidate data from the matching means. Data determination means for determining candidates for each character by reading data.

[Industrial applications]

The present invention relates to a character recognition apparatus, and more particularly to a character recognition apparatus that recognizes input image data obtained by optically reading a character string in units of a character frame having a size corresponding to one character.

[Conventional technology]

2. Description of the Related Art Conventionally, when a handwritten character is optically read and recognized by a character recognition device, each handwritten character needs to be entered in a character frame having a size corresponding to one character. Fig. 8 (a)
Fig. 8B shows an example in which handwritten characters are entered in a character frame, and Fig. 8B shows an example in which a handwritten character protrudes greatly outside the character frame and is connected to an adjacent handwritten character. In the former case, three handwritten characters are correctly recognized as "5", "6", and "7". However, in the latter case, since the first two handwritten characters protrude greatly from the character frame and are connected to the adjacent handwritten characters, they cannot be recognized. Be recognized.

[Problems to be solved by the invention]

Therefore, conventionally, when a read character protrudes from a character frame of a size corresponding to one character or is connected to an adjacent character, there has been a problem that the character cannot be recognized.

Therefore, the present invention provides a character recognition device that enables accurate character recognition even when a read character protrudes from a character frame of a size corresponding to one character or is connected to an adjacent character. With the goal.

[Means for solving the problem]

FIG. 1 is a diagram illustrating the principle of the present invention. In the figure, reference numeral 1 denotes a dictionary, which stores in advance candidate data including information on one or more character candidates and attributes of the candidates for image data in a character frame unit having a size corresponding to one character. I have. Reference numeral 2 denotes a collation unit, which reads candidate data for input image data obtained by optically reading a character string by collating the candidate data with the dictionary 1. Reference numeral 3 denotes a data table in which validity data indicating the validity of each candidate is stored in advance from the correlation between the attributes of two adjacent candidates. Reference numeral 4 denotes a candidate determining unit.

[Action]

The candidate determining means 4 determines a candidate for each character by indexing the data table with the candidate data from the matching means 2 and reading out the corresponding validity data.

Therefore, accurate character recognition can be performed even if a read character greatly extends from a character frame of a size corresponding to one character or is connected to an adjacent character.

FIG. 2 shows an embodiment of the present invention. For example, the handwritten character string shown in FIG. 3 is optically read by the scanning device 10 and the read image data is stored in the image memory 11. The character cutout unit 12 cuts out image data from the image memory 11 in units of a character frame having a size corresponding to one character,
It is stored in the cutout result memory 13 as character data in character frame units. The feature extracting unit 14 extracts features such as loops, end points, and dents used for character recognition from the character data from the cutout result memory 13 and stores the feature data in the feature memory 15. The matching unit 16 reads candidate data for the feature data from the feature memory 15 by checking it against the dictionary 17. The candidate data includes information on character candidates and attributes of the candidates.

Here, the candidates and the attributes of the candidates will be described. Third
When the character strings in the figure are read and cut out in character frame units, the cut out characters are as shown in FIG. 4, and these character data are stored in the cutout result memory 13. In the case of the character data of FIG. 4 (a), there are “3” and “left part of“ 5 ”as candidates, and in the case of the character data of FIG. There is a connected '6'. In the case of the character data shown in FIG. 4 (c), "'7'"
And "'1' connected to the left". These candidate features, such as "left part" and "connected to the left"
It can be considered as a candidate attribute.

The dictionary 17 stores candidate data as shown in FIG. 5 in advance. FIG. 5 (a) shows candidate data for the candidate "left part of '5'", the answer is "5" and the attribute is "left part". Similarly, FIG. 5 (b) shows candidate data relating to the candidate "6 having a connection to the left". The answer is "6" and the attribute is "left connection". (C) shown in FIG. 5 shows candidate data relating to the candidate "'1' connected to the left". The answer is "1" and the attribute is "left connected". FIG. 5 (d) shows candidate data relating to the candidate "'3'", and "normal" in the figure indicates that there is no particular attribute information such as "left connection". Therefore, the matching unit 16 matches the candidate data for the feature data from the feature memory 15 with the dictionary 17, and for example, for the leftmost character in FIG. Left part of
Is read out. In the present embodiment, candidate data “6” connected to the left is read from the dictionary 17 for the middle character in FIG. 3, and “7” is read for the rightmost character in FIG. The candidate data “” and the candidate data “1” connected to the left ”are read from the dictionary 17.

FIG. 6 shows an example of the format of candidate data stored in the dictionary 17 in association with FIG. In FIG.
a is an answer (or category) code, b is an attribute code, c is a bit representing an attribute "left part" in the attribute code b,
“d” indicates a bit representing an attribute “left connection” in the attribute code b. In the drawing, "0" indicates that the bit has no attribute represented, and "1" indicates that the bit has the attribute represented. As is clear from FIG. 6, the attribute is not limited to “left” or “left connection”.

The candidate data output from the matching unit 16 is stored in a candidate memory 18.
Is stored in The editing unit 19 determines the validity of the candidate data from the candidate memory 18 using the adjacency matrix 20, and determines a final candidate. The determined candidate is the answer memory
Stored in 21. Since the subsequent processing of the determined candidate stored in the answer memory 21 is irrelevant to the gist of the present invention,
The description is omitted.

The adjacency matrix 20 is composed of two adjacent candidates (characters)
The validity data indicating the validity of each candidate is stored in a table format in advance based on the relative relationship between the attributes. FIG. 7 shows the concept of the adjacency matrix 20. For example, when determining the final candidates for these characters from the candidate data for the leftmost character and the candidate data for the middle character in FIG. 3, the candidate data includes "3" and "5". "Part" is obtained from the candidate memory 18. As a candidate for the right character adjacent to the leftmost character, “6 connected to the left” is obtained from the candidate memory 18. Therefore,
In FIG. 7, when the character on the left is “3”, the validity data indicates that it is impossible that the character on the left is “normal” and the character on the right is “left connected”. "X"
Is determined. Further, it is determined from the validity data “○” that the left character is “left part” and the right character is “left connection”. Thereby, the editing unit 19 determines and outputs the “left part of '5'” as the final candidate among the two candidate data of the leftmost character.

Similarly, looking at the case where the adjacent characters are the middle character and the rightmost character in FIG. 3, and the right character is “1” connected to the left, the attribute is that the left character is “1”. It is determined from the validity data “x” that it is impossible that the character on the right is “left connected” and the character on the right is “left connected”.
Further, it is determined from the validity data “O” that the left character is “left connected” and the right character is “normal”. Thereby, the editing unit 10 determines and outputs “'7'” as the final candidate among the two candidate data of the rightmost character. In the case of the present embodiment, the third
Since only one candidate data is obtained for the character at the center in the figure, "'6' connected to the left" is determined as the final candidate. Thus, the character strings in FIG. 3 are correctly recognized as "5", "6", and "7".

Even when three or more candidate data are obtained for one character, one candidate can be finally determined by repeating the determination of the validity of each candidate using the adjacent matrix 20 a plurality of times in the same manner. The determination of the validity of each candidate using the adjacent matrix 20 may be performed, for example, on a line-by-line basis. In this case, as a result of the first determination of the validity of the one-line unit candidate, if there are still a plurality of candidates for the character, the validity of the one-line unit candidate is determined at least once again. The number of candidates for each character can be reduced to one.

In the above embodiment, the characters to be recognized are numbers, but the present invention is not limited to this. In addition, adjacent characters may be vertically adjacent to each other in addition to horizontally adjacent. Therefore, when indexing the adjacent matrix 20, it is also possible to use an attribute indicating a connection with upper and lower characters.
The determination of the validity of a candidate is more accurate as the information on the attribute is more. Therefore, the validity data representing the validity of each candidate is determined in advance by the information amount as necessary based on the relative relationship between the attributes of the adjacent candidates. What is necessary is just to store it in the matrix 20.

Further, the character frame in the above embodiment is a frame actually printed on paper or the like indicating an area in which handwritten characters are to be written, for convenience of showing a reading area for one character when reading characters. Frame.

Although the present invention has been described with reference to the embodiments, the present invention can be variously modified in accordance with the gist of the present invention, and these are not excluded from the present invention.

〔The invention's effect〕

According to the present invention, the candidate for each character is determined using the validity data indicating the validity of each character candidate from the relative relationship between the attributes of the adjacent character candidates, so that the read character corresponds to one character. Accurate character recognition can be performed even if the character is out of the character frame of the size or is connected to an adjacent character, which is extremely useful in practice.

[Brief description of the drawings]

 FIG. 1 is a diagram for explaining the principle of the present invention, FIG. 2 is a block diagram showing one embodiment of the present invention, FIG. 3 is a diagram showing a read character string, and FIG. 4 is a cut-out character. FIG. 5, FIG. 5 shows candidate data, FIG. 6 shows an example of the format of candidate data, FIG. 7 shows a diagram for explaining an adjacent matrix, and FIG. 8 shows recognition of handwritten characters. FIG. 1 to 7, 1 is a dictionary, 2 is a collating unit, 3 is a data table, 4 is a candidate determining unit, 10 is a scanning device, 11 is an image memory, 12 is a character extracting unit, and 13 is a extracting result memory. Reference numeral 14 denotes a feature extracting unit, 15 denotes a feature memory, 16 denotes a matching unit, 17 denotes a dictionary, 18 denotes a candidate memory, 19 denotes an editing unit, 20 denotes an adjacency matrix, and 21 denotes an answer memory.

Claims (1)

(57) [Claims] 1. A character recognition device for recognizing input image data obtained by optically reading a character string in units of a character frame having a size corresponding to one character. A dictionary storing candidate data including information about one or more character candidates and attributes indicating characteristics of the candidates in advance, and a collation for reading candidate data for the input image data by collating the dictionary with the dictionary. Means, a data table preliminarily storing validity data representing the validity of each candidate from the correlation between the attributes of two adjacent candidates, and indexing the data table by the candidate data from the matching means. And a candidate determining means for determining a candidate for each character by reading validity data to be performed.