JPH06119495A - Device and method for recognizing character - Google Patents

Abstract

PURPOSE:To reduce the number of times for comparison with a dictionary for recognition and to shorter a recognizing period of time by using a threshold value for stopping useless calculation for the comparison with a standard feature vector which takes the longest processing time among character recognition processing. CONSTITUTION:An input module 21 of a ROM device 20 performs processing for inputting images on the surface of paper containing plural handwritten or printed characters, and a segment module 22 segments character strings or character images from the inputted image data on the surface of paper. On the other hand, a feature vector is extracted from the character image segmented by a feature vector extraction module 23, and a threshold value calculation module 24 calculates the threshold value used for comparison with the dictionary. Then, a character decision module 25 compares the standard feature vector of each character stored in the dictionary for recognition with the feature vector extracted by the feature vector extraction module 23, and decides the segmented character.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method for accelerating comparison processing with a recognition dictionary in a character recognition device for recognizing characters from a paper surface containing handwritten or printed characters.

[0002]

2. Description of the Related Art The flow of character recognition processing in a character recognition device is as shown in FIG. Character recognition is performed by comparing the standard feature vector of each character stored in the dictionary with the feature vector extracted from the character image, and determining the character indicated by the closest standard feature vector as the character indicated by the character image (step 102). ). At the time of comparison with the standard feature vector stored in the recognition dictionary, the i-th standard feature vector registered in the recognition dictionary ai = (ai1, ai
2, ..., ain) and extracted feature vector α = (α1,
α2, ..., αn) is compared using the city block distance, the distance “add” is calculated by the equation shown in step 102.
And compare.

In the character recognition process, the comparison with the recognition dictionary (step 102) takes the longest processing time. Therefore, if a fixed threshold value is set and the standard feature vector exceeds the set threshold value during the comparison of each element between the standard feature vector and the extracted feature vector, the standard feature vector being compared is not appropriate. The decision was made to interrupt the comparison process, and the comparison with the next standard feature vector was started to speed up the process.

[0004]

However, the conventional method of comparing the recognition dictionary in the character recognition apparatus is fixed when the standard feature vector stored in the recognition dictionary is compared with the extracted feature vector. We tried to speed up by suspending the comparison using a threshold value, but it is not possible to set a threshold value suitable for the standard feature vector stored in the recognition dictionary or the extracted feature vector, which is a wasteful comparison. However, there is a problem in that the speed cannot be sufficiently increased.

Therefore, the character recognizing apparatus and the character recognizing method of the present invention are speeded up by setting an appropriate threshold value in the comparison processing with the dictionary in the character recognition processing and terminating unnecessary comparison with the recognition dictionary. It is intended to reduce the processing time of the entire character recognition process.

[0006]

In order to solve the above-mentioned problems, a character recognition device and a character recognition method of the present invention include an input means for inputting an image of a paper surface including handwritten or printed characters, and an input means. When comparing the cut-out means for cutting out a character string or a character from the image on the paper, the feature vector extraction means for extracting a feature vector from the character cut out by the cut-out means, and the standard feature vector stored in the recognition dictionary. Threshold value calculating means for calculating the threshold value used in the above, and the feature vector extracted by the feature vector extracting means using the standard feature vector stored in the recognition dictionary using the threshold value calculated by the threshold value calculating means. And character determining means for determining the character cut out by the cutting means.

[0007]

【Example】

(Embodiment 1) An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an example of the arrangement of the apparatus of this embodiment. From the figure, the character recognition device of the present embodiment is
A CPU 10 for controlling each processing, a ROM device 20 in which processing modules and the like are stored, a RAM device 30 for temporarily storing each processing result and the like, and an input / output device 40 for inputting and displaying characters and the like are provided. There is. In the ROM device 20,
From an input module 21 that is a process of inputting a paper surface image including a plurality of handwritten or printed characters, a cutout module 22 that cuts out a character string or a character image from the input paper surface image data, and a cutout character image A feature vector extraction module 23 for extracting a feature vector, a threshold value calculation module 24 for calculating a threshold value used for comparison with a dictionary, and standard feature vectors and features of each character stored in the recognition dictionary. A character determination module 25 that compares the feature vector extracted by the vector extraction module 23 and determines the cut-out character is provided. Further, the input / output device 40 includes an image scanner 41 for inputting an image of a paper surface including handwritten or printed characters, a display device 42 for displaying the input paper surface and a character recognition result, and a standard feature vector of each character. A storage device 43 for storing the stored recognition dictionary and the like is provided.

Next, the processing contents of this embodiment will be described. As shown in FIG. 2, the processing of the character recognition apparatus in this embodiment is an input unit 201 which is a data reading process by the input module 21, a cutout unit 202 by the cutout module 22, and a feature vector extraction module 2.
Feature vector extraction means 203 by 3 and threshold value calculation means (1) 204 by threshold value calculation module 24
And character determining means (1) by the character determining module 25
And 205. In the figure, the cutting means 20
When character recognition is performed every time one character image is cut out in 2, when the character code of the cut out character image is determined by the character determination means (1) 205, the cutout means moves to the cutout of the next character image. Return to 202. Also, in the cutting means 202,
When character recognition processing is performed after all character images have been cut out from the input paper image, feature vector extraction means 203 extracts a feature vector for each cut out character image, and threshold value calculation means (1) 204
It is assumed that the calculation of the threshold value by the above and the determination of the character code of the cut out character image by the character determining means (1) 205 are repeated.

First, the data input means 201 will be described. The data input unit 201 is the image scanner 41.
Use to read an image on paper that includes handwritten or printed characters. Specifically, the operator uses the CPU 10
Under the control of, the input module 21 is activated to read the image of the paper surface including the desired characters printed on the magazine or the like. By this reading, the image data of the paper surface is input. If necessary, the input image on the paper surface is displayed on the display device 42.

Next, the cutting means 202 will be described. The cutout unit 202 cuts out a character string from the paper image input by the input unit 201 and a character image for each character from the cut out character string. As a method for cutting out a character string and a character image for each character, for example, projection, which is a known method, is used. In the case of a horizontally written document, character string segmentation and character segmentation using projection are performed as follows. First, measure the horizontal projection (projection to the vertical axis) and cut out the range where the projection value exceeds a specific numerical value as a character string.
Next, the vertical projection (projection to the horizontal axis) is measured using the cut-out character string, and the range in which the projection value exceeds a specific numerical value is cut out as a character.

The feature vector extraction means 203 will be described. The feature vector extraction means 203 is a means for numerically expressing the features for character recognition from the cut out character image. A numerical representation of the features for character recognition is called a "feature vector". As an example of features for character recognition, peripheral features will be briefly described with reference to FIG. First, a circumscribing frame of a character pattern is obtained, and the circumscribing frame is divided into some four sides. In the figure, an example of dividing into four is shown. And
By counting the area of the divided circumscribing frame and the white area surrounded by the character part that first meets when seen from the circumscribing frame, and normalizing this with the area surrounded by the circumscribing frame, the feature vector α
{= (Α1, α2, ..., α16)} (16-dimensional feature vector) is obtained.

Next, the threshold value calculating means (1) 204 will be described. The threshold value calculation means (1) 204 finds a threshold value for speeding up the comparison between the standard feature vector stored in the recognition dictionary and the feature vector extracted by the feature vector extraction means 203. As the standard feature vector stored in the recognition dictionary, image data of a character to be a character recognition target is prepared in advance, and the feature vector extraction means 203 extracts the feature vector from the prepared image data of the character. A feature vector is obtained by the same method as the method, the obtained feature vector is registered together with the character code of the character to be recognized, and the feature image is used for character recognition of the character image cut out by the cutting means 203. "Standard feature vector" below
Is a standard feature vector stored in the recognition dictionary.

In this embodiment, a method of obtaining this threshold value from the feature vector obtained by the feature vector extraction means 203 will be described with reference to FIG. First, the feature vector extraction means 2
The sum of the elements of the n-dimensional feature vector obtained in 03 raised to the power of γ is obtained, and the sum is set as β (step 40).
1). It should be noted that γ may be set to “1” when the city block distance is used for comparison with the standard feature vector and “2” when the square of the Euclidean distance is used. Next, a threshold thα is obtained from a function f (β) having β as a variable obtained in advance by experiment (step 402), and this value is used in comparison with the standard feature vector. The method of obtaining the function f (β) shown here will be described later. Next, the character determining means (1) 205 will be described. here,
Threshold value thα obtained by threshold value calculation means (1) 204
Using the standard feature vector and feature vector extraction means 20
The character image cut out is recognized as a specific character by performing comparison with the feature vector obtained in 3. A method of speeding up the comparison with the dictionary using the threshold value thα in the case of comparing the M standard feature vectors to determine a character will be described with reference to FIG. In the figure, "i" is a variable indicating the number of the standard feature vector to be compared, and "min"
Is a variable for containing the minimum distance as a result of comparison with the standard feature vector, and “code” is a variable for containing the recognition result.

First, variables are initialized in steps 501 and 502. Here, the variable “min” has a threshold t
The initial character code is assigned to hα and the variable “code”. As the initial character code, a character code (for example, "■") to be displayed when the character cannot be recognized may be substituted.

Steps 503 to 511 in the figure are a comparison loop with each standard feature vector, and variables "code" and "min" are updated in the process of comparison with each standard feature vector. Steps 504 to 508 in the comparison loop with each standard feature vector are the feature vector α extracted by the feature vector extraction means 203 and the i-th standard feature vector a.
It is a comparison loop for each element of i. In this element-by-element comparison loop, unnecessary comparisons are reduced.

In step 503, the variable "j" indicating the element number of the feature vector is initialized to "1", and the variable "add" indicating the distance from the "i" th standard feature vector is set to "0".
Initialize with. The number of elements of the feature vector is n (n-dimensional). The j-th element aij of the i-th standard feature vector stored in the dictionary and the feature vector extraction means 20.
The distance from the feature vector element αj obtained in 3 is obtained and added to the variable “add” (step 504). Step 504
Then, I showed the method of comparison by city block distance,
The comparison with the standard feature vector may be obtained by another method, for example, the square of the Euclidean distance. Then, the variable "add" representing the distance to the standard feature vector and the variable "min" representing the minimum distance to the standard feature vector are compared (step 505). To the standard feature vector of (step 50
6), otherwise return to step 504 for comparison of next element of feature vector (steps 507, 5)
08).

In step 502, the variable "min" is initialized with a threshold value thα, and in the course of obtaining the distance from the standard feature vector in step 505, the distance "add" obtained up to that point and "min" representing the minimum distance are obtained. It is possible to reduce the number of comparisons by comparing with the standard feature vector and determining whether to continue the distance calculation with the standard feature vector. This is because the threshold value thα is a value indicating that if the distance is longer than this, comparison with an incorrect standard feature vector is performed, and this threshold value thα represents the minimum distance from the standard feature vector. This is because the variable "min" is initialized, and in the process, "min" is updated so as to be smaller than the threshold value thα and to indicate the closest distance among the compared standard feature vectors. Therefore, it is understood that the standard feature vector currently being compared at the time when the value of “min” is exceeded is not an appropriate comparison target, and it is not necessary to continue the comparison with the next element of the feature vector.

In the present embodiment, the feature vector is extracted from the cut-out character image, and the threshold value is set using the feature vector. Therefore, the threshold value suitable for the extracted feature vector can be set. Therefore, as compared with the conventional method using a fixed threshold value, a threshold value suitable for the extracted feature vector can be set, and the number of unnecessary comparisons with the standard feature vector can be reduced. This will be explained later.

When the comparison with all the elements of the feature vector is completed ("NO" in step 508), the determination in step 505 becomes "YES", and step 50
Therefore, the distance to the i-th standard feature vector is shorter (smaller) than the smallest distance “min” in the comparison so far. Therefore, "min" indicating the minimum distance from the standard feature vector is substituted with "add" indicating the distance from the i-th standard feature vector, and the variable "code" indicating the recognition result indicates the i-th standard feature vector. Substitute the character code "code (i)" (step 50)
9).

Comparison with each standard feature vector is performed by the above method. When the comparison with each standard feature vector is completed (step 512), the character code stored in the variable "code" becomes the recognition result of the character. In addition, step 5
If the update in 09 is not performed even once, the initial character code ('■') remains stored in "code". In such a case, there is no standard feature vector close to the feature vector α extracted by the feature vector extraction unit 203,
That is, it means that the character could not be recognized. As described above, by using the threshold value thα obtained by the threshold value calculating means (1) 204, the number of comparisons with the recognition dictionary can be reduced and the speed can be increased. FIG. 5 shows a method of accelerating the comparison with the dictionary using the threshold value thα in the case of comparing the M feature vectors stored in the recognition dictionary and determining the character. In the case of comparing only some of the feature vectors stored in the dictionary, the speed can be similarly increased by the threshold value thα.

In the threshold value calculation means (1) 203,
Although the function f (β) was used when calculating the threshold value, how to obtain this function will be described. Here, for each standard feature vector, a character image of the character indicated by the standard feature vector is prepared, and the prepared character image is used in step 40 of FIG.
1 is obtained, the distance from the standard feature vector is obtained, and the function f (β) is determined from the distribution map of β and the distance.

A detailed description will be given below with reference to FIG. In the figure, “i” is a variable indicating the number of the standard feature vector stored in the recognition dictionary, and it is assumed that there are M standard feature vectors.

In step 602, a character image of the same character as the character indicated by the i-th standard feature vector stored in the dictionary is prepared in advance, and the feature image is extracted from the prepared character image by the same method as the feature vector extraction means 203. The vector α is obtained, and β is obtained by the method shown in step 401 of FIG. At this time, when character images of a plurality of typefaces such as Mincho typeface or Gothic typeface with the same character as the i-th standard feature vector are prepared, β is obtained from each character image.

Next, the distance between the feature vector α obtained from the prepared character image and the i-th standard feature vector ai stored in the recognition dictionary is obtained (step 603).
Regarding the distance, all elements are compared using the same method as the distance calculating method used in the character determining means (1) 205 to obtain the distance. That is, when the city block distance is used in the character determining means (1) 205, step 1 in FIG.
The distance “add” is obtained by the method indicated by 02. At this time, if a plurality of character images are prepared in step 602, β and the distance are obtained from each character image.

When all the M standard feature vectors have been processed, a distribution map of the distance between β and the recognition dictionary can be drawn as shown in FIG. 6 (b). Therefore, the function f (b) is obtained from the distribution chart of FIG. 6 (b). As an example of how to obtain,
With the function f (β) as f (β) = c1 × β + c2, the constants c1 and c2 are determined by the method of least squares. Assuming that the function f (β) obtained here is a line represented by 611, 611 is a threshold value, and therefore points below this line can be recognized, but upper points are threshold values. Since it takes a larger value, it cannot be recognized. Therefore, if the highest recognition rate in the character image prepared in advance is to be set to ε% when obtaining the distribution map, the obtained function is moved like 612, and the points below 612 become ε% of the whole. It is good to do so. The function f is adjusted according to the capability required by the character recognition device by the above movement.
(Β) should be set. As described above, the threshold value t
A function f (β) for determining hα is obtained.

When the conventional fixed threshold value is used, 6
Similar to 12, if the points included under the line are ε% of the whole, the line becomes a dotted line 613. As can be seen from the figure, when the conventional fixed threshold value is used, the threshold value matching the distribution is not set. For example, since the effective threshold value has not been set in the area 614,
There was a wasteful comparison process until the threshold value was reached in comparison with the standard feature vector. However, by obtaining the function f (β) as shown in this embodiment, it is possible to set a threshold value suitable for the extracted feature vector, and therefore, as compared with the case of using a conventional fixed threshold value, Useless comparison processing can be reduced.

As shown in this embodiment, a threshold value is calculated from the feature vector obtained from the cut-out character image, and by using this threshold value, the recognition dictionary is more effective than when a fixed threshold value is used. It is possible to reduce the number of comparisons when obtaining the distance from the standard feature vector stored in.
Therefore, the comparison time with the recognition dictionary can be reduced and the character recognition speed can be increased.

(Embodiment 2) In Embodiment 1, a threshold value is calculated using the feature vector obtained from the cut-out character image, and the comparison time with the recognition dictionary is shortened using this threshold value. I explained how. In this embodiment, a method of calculating a threshold value used to shorten the comparison time with the recognition dictionary from the standard feature vector will be described.

The processing contents of this embodiment will be described. As shown in FIG. 7, the processing of the character recognition device in this embodiment is an input unit 201 which is a data reading process by the input module 21, a cutout unit 202 by the cutout module 22, and a feature vector extraction unit 203 by the feature vector extraction module 23. And threshold value calculating means (2) 206 by the threshold value calculating module 24 and character determining means (2) 207 by the character determining module 25. The input unit 201, the cutout unit 202, and the feature amount extraction unit 203 are the same as those in the first embodiment.

First, the threshold value calculating means (2) 206 in this embodiment will be described with reference to the drawings. Figure 8 (a)
FIG. 3 is a conceptual diagram showing the distribution status of each standard vector in the standard feature vector space. In the figure, the standard vector a1
When paying attention to, the closest vector in the comparison with the unknown feature vector α is judged to be a1 when it falls within the range surrounded by the broken line in the figure. The region surrounded by the broken line in the figure is drawn by obtaining a large number of points equidistant from each standard feature vector in the vicinity of a1. Therefore, when comparing the standard feature vector a1 and the unknown feature vector α, it suffices to determine whether they are within the broken line. Although the standard feature vector is conceptually shown in the two-dimensional space in FIG. 8, it is difficult to find the boundary indicated by the broken line in FIG. 8 because the standard feature vector actually exists in the n-dimensional space. Therefore, a method of approximating the broken line region by an n-dimensional sphere centered on a1 and using the radius of the sphere as a threshold value will be described below. For example, as a method of obtaining the radius of the sphere for the standard feature vector a1, m standard feature vectors (d1 to dm) near a1 are selected, the average of the distances between the selected standard feature vectors and a1 is calculated, and the value is calculated. Determine the radius of the sphere based on. How to obtain the M threshold values th (i) {1 ≦ i ≦ M} corresponding to the M standard feature vectors will be described in detail with reference to FIG. 9. In the figure, "i" is a variable indicating the number of the standard feature vector stored in the recognition dictionary. In step 702, the distance between the i-th standard feature vector and the other standard feature vectors is calculated, and m standard feature vectors are selected from the closest ones, and each selected standard feature vector and the i-th standard feature vector are selected. The average “A” of the distances from and is calculated. Here, the distance is obtained using the same method as the method for obtaining the distance used by the character determining means (2) 207. For example, assuming that the distance is a city block distance, the distance to the standard vectors ai and ak is obtained by the method shown in step 702. Then, the threshold value used when comparing the i-th feature vector ai with the feature vector obtained from the cut-out character is th (i) = κ × A. This th (i) is the i-th standard feature vector a
It indicates the radius of an n-dimensional sphere centered on ai which approximates the effective range of i (step 703).

When the threshold values have been calculated for all M standard feature vectors contained in the recognition dictionary, the process is terminated.

A method of setting κ in step 703 will be described. First, character images corresponding to the first to Mth standard feature vectors are prepared in advance, a feature vector is obtained from the prepared character images by the same method as the feature vector extraction means 203, and the distance from the corresponding standard feature vector is calculated. Ask. Regarding the distance, all elements are compared using the same method as the distance calculation method used by the character determining means (2) 207 to obtain the distance. The average A of this distance and the m distances obtained in step 702 is obtained for each standard feature vector,
The distribution chart shown in FIG. 8B is obtained. From the obtained distribution map,
If you want to set the highest recognition rate in the character image prepared in advance as ε '%, draw a line 621 that passes through the origin, and
The points below 21 should be ε '% of the whole. The slope of the line thus obtained is κ. Κ as above
The value of κ may be set according to the ability required by the character recognition device by the setting method of.

The threshold value is calculated in advance by the method described above, and only the calculated result is stored in the storage device 43 or the like as a table, and the character determining means (2) 2
At 07, the threshold is set by referring to the table to determine the character.

When the conventional fixed threshold value is used, 6
As in the case of 21, the dotted line 622 is obtained when the points included under the line are ε '% of the whole. As can be seen from the figure, when the conventional fixed threshold value is used, the threshold value matching the distribution is not set. For example, in the area indicated by 623, an effective threshold value has not been set, so that in comparison with the standard feature vector, useless comparison processing occurs until the threshold value is reached. However, since a threshold value suitable for each standard feature vector can be set by the threshold value calculation method as shown in the present embodiment, a wasteful comparison process can be performed as compared with the conventional fixed threshold value. Can be reduced.

Next, the character determining means (2) 207 for determining the character by comparing with the standard feature vector using the threshold value calculated by the threshold value calculating means (2) 206 will be described. Here, the threshold value t used in the comparison with each standard feature vector obtained by the threshold value calculation means (2) 206.
h () is compared with the minimum distance obtained in the process of comparison with each standard feature vector, and the smaller one is used as a threshold to reduce the number of comparisons and speed up. This will be described in detail with reference to FIG.

In the figure, "i" is a variable indicating the number of the standard feature vector, and "min" is a variable for accommodating the minimum distance as a result of comparison with the standard feature vector, "code".
Is a variable for storing the recognition result.

First, in step 901, each variable is initialized. The variable “min” is the threshold value calculation means (2) 2
Initialization is performed with the maximum of the M threshold values obtained in 06, and "code" is initialized with the initial character code.

Steps 902 to 915 in the figure are a comparison loop with each standard feature vector, and variables "code" and "min" are updated in the process of comparison with each standard feature vector. Steps 906 to 910 in the comparison loop with each feature vector are the feature vector α extracted by the feature vector extraction means 203 and the i-th standard feature vector ai.
Is a comparison loop for each element of. In this element-by-element comparison loop, the threshold value th is used to reduce unnecessary comparison.

In step 902, the threshold th (i) representing the effective range of the i-th standard feature vector and the variable "mi"
n ”, and the smaller one is used as the threshold value th for comparison with the standard feature vector (steps 903, 90).
4).

In step 905, the variable "j" indicating the element number of the feature vector is initialized to "1", and the variable "add" indicating the distance to the feature vector stored in the dictionary is initialized to "0". . And i in the dictionary
The distance between the j-th element aij of the th standard feature vector and the element αj of the feature vector obtained by the feature vector extraction means 203 is obtained and added to the variable "add" (step 90).
6). Although the method of obtaining the city block distance is shown here, the comparison with the standard feature vector stored in the recognition dictionary may be obtained by another method, for example, the square of the Euclidean distance. Then, the variable "add" representing the distance from the standard feature vector is compared with the threshold th (step 907), and if the condition is not satisfied, the process proceeds to the comparison with the next standard feature vector (step 908), If the condition is satisfied, the process returns to step 906 to compare the next element of the feature vector (steps 909 and 910).
In the middle of obtaining the distance to the standard feature vector in step 907, the distance “add” obtained up to that time is compared with the threshold th indicating the minimum effective range, and the distance calculation with the standard feature vector is continued. By making a judgment, the number of calculations can be reduced. This is because the threshold th is a value indicating that the standard feature vector to be compared is not appropriate if the distance is longer than this. Therefore,
By the determination in step 907, it becomes possible to shorten the time required for the comparison calculation with the useless standard feature vector.

After the comparison with all the elements of the feature vector, whether the obtained distance "add" is smaller than "min" showing the minimum distance in comparison with the standard feature vector up to the "i-1" th. The determination is made (step 911), and if smaller, the variable “min” and the variable “code” containing the recognition result are updated (step 913).

When the comparison with each standard feature vector is completed (step 915), the character determining means (2) 205 is ended (step 916). When the character determining means (2) 207 is terminated, the character code stored in the variable "code" becomes the character recognition result.

As described in the present embodiment, the effective range of each standard feature vector in the standard feature vector space is represented by a sphere centered on the standard feature vector, and the radius of the sphere is used as a threshold value for comparison. It is possible to set an appropriate threshold value for each standard feature vector to be used. In addition, 1) the threshold value “th ()” used in comparison with each standard feature vector, and 2) the smaller one of the minimum distance values “min” in the standard feature vectors compared up to that point. By interrupting the comparison with the feature vector currently being compared when the threshold value th exceeds this value, useless comparison can be omitted when comparing with each standard feature vector, and as a result, processing is performed. It becomes possible to shorten the time.

(Third Embodiment) In this embodiment, the threshold value thα calculated using the feature vector obtained by the feature vector extracting means 203 in the first embodiment and the threshold obtained from the standard feature vector in the second embodiment. A method of shortening the comparison time with the standard feature vector by using the value th () together will be described.

First, the processing contents of this embodiment will be described. As shown in FIG. 11, the processing of the character recognition device in the present embodiment is an input unit 201 which is a data reading process by the input module 21, a cutout unit 202 by the cutout module 22, and a feature vector extraction unit 203 by the feature vector extraction module 23. And threshold value calculation means (1) 20 by the threshold value calculation module 24
4 and (2) 206, and the character determination means (3) 208 by the character determination module 25. In addition,
The input unit 201, the cutout unit 202, the feature amount extraction unit 203, and the threshold value calculation unit (1) 204 are the same as those in the first embodiment, and the threshold value calculation unit (2) 206 is the same as that in the second embodiment. Therefore, in this embodiment, the character determining means (3)
208 will be described.

The character determining means (3) 208 is almost the same as the character determining means (2) 207 shown in FIG. The difference is that in step 901, the variable “min” representing the minimum distance from the standard feature vector is set to the threshold value calculation means (1) 20.
The point is that initialization is performed with thα obtained in 4.

A fixed threshold value 613 in FIG. 6B.
It can be considered that the fixed threshold value 622 in FIG. 8B is a value of the same level. Because the vertical axis in FIG. 6 (b) is the distance from the character image dictionary, and FIG.
This is because the vertical axis of is the distance from the character image in the dictionary, but it is the same in comparison between the dictionary and the character image.
Therefore, the maximum value max {th ()} of the threshold value obtained by the threshold value calculation means (2) 206 is almost the same as the fixed threshold value 622 in FIG. The same applies to the fixed threshold 613 in ().
From this, when the threshold value thα obtained by the threshold value calculation means (1) 204 and the maximum value max {th ()} of the threshold value obtained by the threshold value calculation means (2) 206 are compared,
It can be seen that thα is smaller than max {th ()}. Therefore, in step 901, a smaller value can be set as an initial value by initializing the variable “min” representing the minimum distance from the standard feature vector with thα obtained by the threshold value calculation means (1) 204. Therefore, the number of comparisons with each element of the standard feature vector can be reduced. Further, it is possible to reduce the time required for the comparison process with each standard feature vector as compared with the first and second embodiments.

As described above, the threshold value thα calculated using the feature vector obtained by the feature vector extraction unit 203 according to the first embodiment and the threshold value th () obtained from the standard feature vector according to the second embodiment are used together. By doing so, it is possible to set an appropriate threshold value obtained from two different directions, so that speedup can be expected.

[0050]

As described above, the character recognizing device and the character recognizing method of the present invention provide a threshold value for canceling unnecessary calculation in comparison with the standard feature vector which requires the longest processing time in the character recognizing process. By using it, the number of comparisons with the recognition dictionary is reduced, and as a result, the comparison processing time is shortened, and the character recognition time is shortened as a whole. As the threshold value calculation method used at this time, 1) calculation is performed using the feature vector extracted from the cut-out character image. 2) When comparing with each standard feature vector from the distribution state of the standard feature vector in the standard feature vector space. There are two ways to calculate the threshold.

As shown in FIG. 6 (b), 1) is a function f indicated by a solid line from the distribution map of the distance between β obtained from the extracted feature vector and the standard feature vector stored in the recognition dictionary.
Since (β) is obtained, it is possible to set an appropriate threshold value for the cut out character as compared with the conventional fixed threshold value shown by the dotted line 613, and it is possible to reduce unnecessary comparison in the area shown at 614.

As shown in FIG. 8 (b), 2) is a threshold 621 from the distribution map of the distance between the average A of m distances obtained from each standard feature vector and the character image prepared in advance. Since a value can be calculated and an appropriate threshold value can be set for each standard feature vector, useless comparison can be reduced in the area indicated by 623 compared with the conventional fixed threshold value indicated by the dotted line 622. it can.

By further using 1) and 2) together,
The initial value of the threshold value can be set to a smaller value, and unnecessary comparison can be reduced as compared with the case of only 1) or 2).

As compared with the method of performing comparison using a fixed threshold value as described above, in comparison with a standard feature vector that takes the longest processing time in character recognition processing, it is more flexible and faster than a recognition dictionary. Allows comparison.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration example of an apparatus necessary for the present invention.

2 is a block diagram of the present invention in Embodiment 1. FIG.

FIG. 3 is a diagram illustrating a feature vector extraction example for explaining a feature vector extraction unit.

FIG. 4 is a flow chart for explaining a threshold value calculation means in the first embodiment.

FIG. 5 is a flow chart for explaining a character determination means in the first embodiment.

FIG. 6 is a diagram for explaining a method of obtaining a function f (β) used by the threshold value calculation means in the first embodiment.

FIG. 7 is a block diagram of the present invention in a second embodiment.

FIG. 8 is a diagram for explaining a threshold value calculation means according to the second embodiment.

FIG. 9 is a flow chart for explaining a threshold value calculation means in the second embodiment.

FIG. 10 is a flow chart for explaining a character determination means in the second embodiment.

FIG. 11 is a block diagram of the present invention in a third embodiment.

FIG. 12 is a diagram for explaining a conventional character recognition method.

[Explanation of symbols]

 10 CPU 20 ROM device 21 Input module 22 Clipping module 23 Feature vector extraction module 24 Threshold calculation module 25 Character determination module 30 RAM device 40 Input device 41 Scanner 42 Display device 43 Storage device

Claims (6)

[Claims] 1. Input means for inputting an image of a paper surface including handwritten or printed characters, cutout means for cutting out a character string or characters from the image of the paper surface input by the input means, and cutout by the cutout means The feature vector extraction means for extracting the feature vector from the extracted character and the standard feature vector stored in the recognition dictionary are used to calculate the threshold value using the feature vector obtained by the feature vector extraction means. By using the threshold value calculation means and the threshold value calculated by the threshold value calculation means, the standard feature vector stored in the recognition dictionary and the feature vector extracted by the feature vector extraction means are compared. And a character deciding means for deciding the character cut out by the cutting means. 2. Input means for inputting an image of a paper surface including handwritten or printed characters, cutout means for cutting out a character string or a character from the image of the paper surface input by the input means, and cutout by the cutout means. The feature vector extracting means for extracting the feature vector from the character and the standard feature vector used in the comparison with the standard feature vector stored in the recognition dictionary are effective for each standard feature vector stored in the recognition dictionary. A threshold value calculating means for calculating the range, a threshold value calculated by the threshold value calculating means, a standard feature vector stored in the recognition dictionary, and a feature vector extracted by the feature vector extracting means, And a character determining means for determining the character cut out by the cutting means. Recognition device. 3. Input means for inputting a paper surface image including handwritten or printed characters, cutout means for cutting out a character string or characters from the paper surface image input by the input means, and cutout by the cutout means. The feature vector extracting means for extracting the feature vector from the extracted character and the standard feature vector stored in the recognition dictionary are used as a threshold value using the feature vector obtained by the feature vector extracting means. While calculating the value 1,
Threshold value calculating means for calculating the effective range of each standard feature vector stored in the recognition dictionary as the threshold value 2, and the threshold value 1 and the threshold value calculated by the threshold value calculating means. 2 in combination, the standard feature vector stored in the recognition dictionary is compared with the feature vector extracted by the feature vector extraction means, and the character determination means for determining the character cut out by the cutout means. A character recognition device characterized by comprising. 4. An input means for inputting an image on a paper surface including handwritten or printed characters, a cutting means for cutting out a character string or a character from the image on the paper surface input by the input means, and a cutout by the cutting means. Feature vector extracting means for extracting a feature vector from the extracted character, threshold value calculating means for calculating a threshold value used in comparison with the standard feature vector stored in the recognition dictionary, and the threshold value calculating means. Using the threshold value calculated by the above, the standard feature vector stored in the recognition dictionary is compared with the feature vector extracted by the feature vector extraction means, and the character cut out by the cutout means is determined. In the character recognition device including a determination means, the threshold value calculation means is characterized by the feature vector extraction means. Character recognition method characterized by calculating the threshold using torr. 5. Input means for inputting an image of a paper surface including handwritten or printed characters, cutout means for cutting out a character string or a character from the image of the paper surface input by the input means, and cutout by the cutout means. Feature vector extracting means for extracting a feature vector from the extracted character, threshold value calculating means for calculating a threshold value used in comparison with the standard feature vector stored in the recognition dictionary, and the threshold value calculating means. Using the threshold value calculated by the above, the standard feature vector stored in the recognition dictionary is compared with the feature vector extracted by the feature vector extraction means, and the character cut out by the cutout means is determined. In the character recognition device including a determination unit, the threshold value calculation unit includes each standard feature stored in the recognition dictionary. Character recognition method characterized by calculating the effective range of the vector as a threshold. 6. An input means for inputting an image of a paper surface including handwritten or printed characters, a cutting means for cutting out a character string or a character from the image of the paper surface input by the input means, and a cutout by the cutting means. Feature vector extracting means for extracting a feature vector from the extracted character, threshold value calculating means for calculating a threshold value used in comparison with the standard feature vector stored in the recognition dictionary, and the threshold value calculating means. Using the threshold value calculated by the above, the standard feature vector stored in the recognition dictionary is compared with the feature vector extracted by the feature vector extraction means, and the character cut out by the cutout means is determined. In the character recognition device including a determination means, the threshold value calculation means is characterized by the feature vector extraction means. The threshold value 1 is calculated by using Toru, and the effective range of each standard feature vector stored in the recognition dictionary is calculated as the threshold value 2. A character recognition method characterized by using the threshold value 2 together.