JPH0314081A - Character recognition device - Google Patents

Abstract

PURPOSE:To attain discrimination which is close to a human intuitional observation and to enable high-accuracy recognition by selecting and matching a character recognition result among selected candidate characters according contour information. CONSTITUTION:This device includes a character discriminating means 5, a detecting means 7 which detects contour feature information, a storage means 9 for storing the standard contour feature information, and a selecting means 8. Then the length or quantity of contours constituting a character is determined, character by character, which is utilized to select plural candidate characters to be recognized by calculating the distances between the character to be recognized and feature patterns extracted from a standard character, thereby selecting and matching the character recognition result among the selected candidate characters according to the contour feature information. Consequently, when character recognition is performed by using a Glucksman method, the discrimination which is close to a human intuitional observation is possible and the high- accuracy recognition is enabled.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character recognition device that converts character images into character codes.

[Conventional technology]

Conventionally, a character recognition device calculates the distance between a feature pattern extracted from a character image and a standard feature pattern whose character coat is known, and outputs a character code that is most similar to both feature patterns as a recognition result of the character image. Document 11. is one of the character recognition methods used in character recognition devices.
8. Glucksman + C1 assifica mouth
n of m1xed font alphabet
icsby characteristic 1o
ci'' Digestof 1st 8nn.

IEEE' Computer Conf, pp.
The character recognition method shown in 137-141.1967 (hereinafter referred to as Glucksmann's method) captures the topological characteristics of characters well, so it can recognize alphanumeric and katakana characters with simple recognition processing. It is known that this can be done with relatively good results.

Glucksman's method focuses on the background of the character image,
A half-line is extended in the vertical and horizontal directions from the white pixels forming the background part, and the number of times it intersects with the character line is counted. Then, the count value is divided into three stages: 0 times, 1 time, and 2 or more times, and each value is “0”.
, "1", and code this code for each direction for one white pixel.Next, for each white pixel:L' = 81 combinations of the above codes for character lines (envelopment state information) is created as a table in advance, the encirclement information code is extracted from the character image to be recognized, and the number of white pixels having the same encirclement information code as the encirclement information code in the above table is calculated as an 81-dimensional feature vector. Character recognition was performed by comparing this first feature vector, given as , with a standard feature vector whose character code was known.

(Problem to be Solved by the Invention) However, the above method is not limited to Gluxman's method, and distance calculation between a feature pattern (vector) extracted from a character to be recognized and a standard feature pattern extracted from a character whose character code is known. When performing character recognition using this method, characters with similar shapes may be incorrectly recognized. For example, in Glucksman's method mentioned above, J is turned into n, 9 is turned into ? I sometimes mistook shi for n, n for shi, and mi for 8.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a character recognition device that can eliminate the above-mentioned problems and further improve character recognition accuracy without significantly changing the configuration of conventional character recognition devices. .

(Means for Solving the Problems) In order to achieve such an object, the first form of the present invention is as follows:
A character that calculates the distance 1ffl between a feature pattern extracted from a character image to be recognized and a standard feature pattern prepared for each character code, and outputs character codes of multiple standard patterns that approximate the feature pattern as character recognition candidates. an identification means; a detection means for detecting, from a character image to be recognized, contour feature information of at least one of the length and number of contour lines constituting the character image using an image recognition method; Marks set in advance for each character code to check contour feature information! A storage means storing one-town contour feature information and standard contour feature information corresponding to a plurality of character codes outputted from the character identification means are extracted from the storage means, and the extracted standard outline feature information and a detection means The present invention is characterized by comprising a selection means for selectively outputting a character code that matches the contour feature information detected by the method.

Furthermore, the second form of the present invention is characterized in that the number included in the second contour information has an upper limit value and a lower limit value.

[For production]

In the present invention, the length or the number of contour lines constituting the character is determined for each character, and the distance from the feature pattern extracted from each of the recognition target character and the standard character is 1.
f! Since multiple recognition candidate characters are selected by IiI calculation and character recognition results are selected and collated from among the selected candidate characters based on contour feature information, character recognition using, for example, Glucksman's method can be performed. In this case, it is possible to achieve identification closer to human intuitive observation and to enable highly accurate recognition.

〔Example〕

The present invention will be described in detail below with reference to the drawings.

FIG. 1 shows the circuit configuration of a first embodiment of the present invention.

In FIG. 1, 1 is a photoelectric transformer l! that reads a document and converts it into binary image data. ! ! This is a reading section using an element.

2 is a binary 1ilUii@ data, which is an EIf elephant effect reminiscent of 9self. 3 is a character cutting circuit using a central processing unit ((:Pt1) for cutting out the binary image data stored in the image memory 2 into the size of one character; 4 is a character cutting circuit using a central processing unit ((:Pt1)) for cutting out the binary image data stored in the image memory 2; This is a feature extraction circuit that extracts character feature vectors using a method and sends the extracted feature vectors to the first identification circuit 5.
Use u.

5 receives the above feature vector and calculates the distance between the feature vector and the average vector of each character in the dictionary section 6.
This is a first identification circuit using u, which selects, for example, three candidate characters from the one with the smallest distance, and sends the character codes of the three candidate characters to the matching circuit 8.

Reference numeral 7 denotes a contour line number counting circuit using CI'U which extracts ring 'AS lines from the extracted character image data in the image memory 2 and calculates coefficients of the number of inner contours and the number of outer contours. 8 is a matching circuit using CPU, which calculates the number of inner contours and outer contours determined by the contour counting circuit 7 and the ring 5i31!
Number of inner contour lines and outer ring 9 for each character in the information table
15 lines, selects the most matching character among the three candidate characters obtained by the first recognition circuit 5, and outputs its character code. 1G is an output buffer in which character codes are written.

FIG. 2 shows the contents of the contour information table.

In FIG. 2, the contour information table 9 stores the character code of the character to be recognized and the upper limit of variation in the number of inner contour lines and the upper limit of variation in the number of outer contour lines corresponding to the character code. Here, the upper limit of variation means the upper limit of the number of contour lines that is statistically possible due to variations in handwritten characters. For example, E“
In normal calligraphy, the number of outer contour lines for the character ``1'' is 1 degree, but when the character lines are separated, the number of outer contour lines is 2 degrees.
Therefore, the upper limit of variation is .degree. 2". Note that if the lower limit of variation is used in addition to the upper limit of variation, character recognition accuracy will be increased.

Returning to FIG. 1, the circuit operation of the embodiment of the present invention is shown in FIG. 3 (1).
, (2) will be explained with reference to the flowchart.

When the reading unit 1 inputs a document image from an original, the document information is converted into a binary image and stored in the image memory 2 (step St in FIG. 3). Next, the character cutting circuit 3 operates, and the binary image data of each document is cut out into character image data for each character and written into the margin of the image memory 2 (step S2).

Next, the feature extraction circuit 4 operates to create a feature vector from the character image data using Gluxman's method, and outputs this feature vector to the discrimination circuit 5 (step 511). The Euclidean distance between the average vector of each character and the feature vector is calculated, and three candidate character codes are searched (sorted) in descending order of distance value (steps 512 to 516). Up to this point, the operation is similar to that of the conventional character recognition circuit, and the three candidate character codes of order 0 are passed to the matching circuit 8 in order of decreasing distance value.

On the other hand, at the same time as the feature extraction circuit 4 starts operating, the contour number counting circuit 7 starts operating, calculates the number of inner contours and the number of outer contours of the character image data using a well-known contour tracking method, and outputs the calculation results. It is sent to the matching circuit 8 (step 520). When the verification circuit 8 completes the input of the information from the contour number counting circuit 7 and the identification circuit 5, it performs the next verification process.
The number of manually inputted inner contours is 01, and the number of inputted outer contours is 00.
, the inner limb number upper limit and the outer limb number upper limit of the contour information table 9 corresponding to the first candidate character, the second candidate character, and the third candidate character (the upper limit of the S number is Nl l , NI., N21 .
N2°, and N.

1.N. That is, matching process (1) The above numerical values are compared, and the comparison result is 1≦Nth and n. ≦N
,. In this case, the character code of the first candidate character is directly output as the final recognition result to the output buffer IO (step S22→526).

The comparison result is n I > N + + or n, > N. When this happens, the following verification processes (2) and (3) are performed.

(2) If the size relationship of the number of contours is n1≦N21 and n0
≦N,. If so, the first candidate character and the second candidate character are exchanged. If not, the next verification process (3) is performed (step S23→525).

(3) If the size relationship of the number of contours is n1≦N31 and n. ≦
If N3°, the first candidate character and the third candidate character are exchanged (step S24→525).

After performing the above-mentioned matching processes (1) to (3), the matching circuit 8 outputs the character code of the first candidate character to the output buffer IO (step 52B).

In this embodiment, the number of recognition candidates is limited to three, but the number may be set to a fixed number as desired.Also, the distance calculation results performed by the identification circuit 5 are compared with a threshold value, and the number of candidates below the threshold value is All character codes having distance calculation results may be used as recognition candidates.

[Other Examples] Next, a second example will be described. In this example, the inner contour length and the outer contour length are used as the contour information instead of the number of contours.

In this case, as shown in FIG. 4, the contour number counting circuit 7 is replaced by the contour length calculation circuit 7', and the contents of the contour information table are changed from the contour to the contour length.

In this example, as in the first embodiment, the identification circuit 5 calculates the distance between the average vector of each character and the input feature vector, and calculates the three character codes having the average vector with a small distance and the single value of the distance. It is sent to the verification circuit 8. These distance values are represented by dt, dt, and d* in order, and the character codes selected as recognition candidates are represented by C1, C2, and C3 in order. The inner contour length and outer contour length of the human character image calculated by the contour length calculation circuit 7 are respectively indicated by sentences, ,X,.

In the matching circuit 8, the distance regarding the outline length for each of the three candidate characters is determined using the following formula.

eb''(fl+-Ll(k')2+ (-9° 1o(kl)2) where 61( is the distance N[, 1, lkl, L
olkl is the average length of the inner contour and the average length of the outer contour in the contour information table (see FIG. 5) for candidate character C3. Then, a new distance 1llIfo+ is calculated using a predetermined weight ω.

is calculated using the following formula.

[1b= dk+ (Lle k(1, -1, 2, 3
) Then, the character code Ct corresponding to the minimum value among Db(m-1, 2, 3) is sent to the output buffer 7.

(Effects of the Invention) As explained above, according to the present invention, the length or number of outlines that make up each character is determined, and the outlines are extracted from each of the recognition target character and the standard character. Recognition candidate characters are selected by calculating the distance Ht from the feature pattern, and character recognition results are selected and compared based on contour information for each of the selected candidate characters, so for example, characters using Glucksman's method When performing recognition, it is possible to achieve identification that is closer to human intuitive observation, enabling highly accurate recognition.

Furthermore, by providing a lower limit value in addition to an upper limit value for the number of contour lines, accurate character recognition can be performed even when the contour lines constituting a character are spaced apart from each other.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the circuit configuration of the first embodiment of the present invention, FIG. 2 is an explanatory diagram showing the contents of the contour information table 9 of the first embodiment of the present invention, and FIGS. ) is a flowchart showing the operation procedure of the first embodiment of the present invention, FIG. 4 is a block diagram showing the circuit configuration of the TtS2 embodiment of the present invention, and FIG. 5 shows the contents of the contour information table of the second embodiment of the present invention. FIG. DESCRIPTION OF SYMBOLS 1...Reading part, 2...Image memory, 3...Character cutting circuit, 4...Feature extraction circuit, 5...Identification circuit, 6...Dictionary part, 7...Contour line Counting circuit, 8... Verification circuit, 9... Wheel 915 a +if fFJ Tef/l/
, 1θ...output buffer 7°/. Outline of Honfei No. 1 Tora Shii 1 M1 News table 9 inner g and 1 explanatory time Figure 2 X4I! : Ming? +1 actual] Flowchart of column f Fig. 1 Flowchart of column B Monthly JE 4th column Fig. 3 (2) 10 $4esll Fluff diagram of column f of 2nd package Fig. 4 Main date/month'$ Contents of 21 criminals 1's transport and 1 present table 9! L, zi explanatory diagram Figure 5

Claims (1)

[Scope of Claims] 1) Distance calculation between a feature pattern extracted from a character image to be recognized and a standard feature pattern prepared for each character code is performed to determine character codes of a plurality of standard patterns that approximate the feature pattern. character recognition means for outputting a character image as a character recognition candidate; and contour line feature information on at least one of the length and number of contour lines constituting the character image using an image recognition method from the character image to be recognized. a detection means for detecting a plurality of character codes outputted from the character identification means; a storage means for storing standard outline characteristic information set in advance for each character code in order to collate the contour line characteristic information; a selection means for extracting the corresponding standard outline feature information from the storage means and selectively outputting a character code in which the extracted standard outline feature information matches the outline feature information detected by the detection means; A character recognition device characterized by comprising: 2) The character recognition device according to claim 1, wherein the number of contour lines included in the standard contour feature information has an upper limit value and a lower limit value.