JPH0458073B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition method in which a dictionary is created from the characters of a character sample form, and when the form is read, the characters are recognized by comparison with the created dictionary.

Conventionally, with this type of character recognition method, even if a person writes messy characters, if the characters are limited to each individual, the letterforms will have a similar pattern. , the features of each character are extracted, and the range of feature values obtained for each character category is determined and used as a dictionary for form fillers.

FIG. 1 is a diagram showing an example of a guide character sample form for dictionary creation, and in this example, the category name is determined by the position on the form.

By the way, even with this method, the feature values extracted for different categories with similar shapes must be different because the dictionary is not created taking into account the feature values of other categories, and it is difficult to distinguish between macro features and micro features. It is difficult to create a dictionary because it is necessary to extract many at the same time and create a dictionary. Furthermore, creating a dictionary by taking into consideration the feature amounts of all other categories takes too much time.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character recognition method for creating a dictionary through division processing, which solves the above-mentioned problems.

In order to achieve the above object, the character recognition method of the present invention first inputs a character sample form, and generates a code value that encodes the category name given to each character and the feature quantities of a plurality of predetermined features. When the encoding of the characters on the character sample form is completed, the code value is calculated from the code value appearance frequency distribution for each category in one of the macro features (hereinafter referred to as feature H). Determine the code value range so that the number of categories within the range is small, and for each code value range, use the sequence of code values whose code values corresponding to the feature H are within the range, and calculate the code values with the same category name. Do not include the code value columns of other categories, combine the code values for each feature, find the lower limit code and upper limit code, make the code value range, and then calculate the category name and code value for each feature. A dictionary element of a divided dictionary is created using the range of , and a set of divided dictionaries created for each code value range is defined as a dictionary.

This method creates a dictionary by always considering the feature values of other categories. Therefore, this method automatically creates a dictionary that is suitable for recognizing the characters of the person filling out the form. Since the amount of data can be reduced, the creation time can be shortened.

FIG. 2 is a block diagram of a specific device for explaining the conventional character recognition method, in which a dictionary is stored in the dictionary section 5 from the auxiliary storage section 7 before reading the form.

The character pattern of one character on the form is photoelectrically converted in the scanning unit 1 and stored as image data in the pattern memory unit 2.
is memorized. The feature extraction section 3 extracts the feature amount of the feature necessary for recognition from the two-dimensional pattern in the pattern memory section 2, and the matching section 4 also matches the feature amount stored in the dictionary section 5 with the extracted feature amount. , outputs the reading result 6.

On the other hand, FIG. 3 is a block diagram showing an embodiment of a specific device for explaining the character recognition method according to the present invention. First, when a character sample form is input, the character pattern of one character on the form is scanned. It is photoelectrically converted in section 1 and stored as image data in pattern memory section 2, and feature extraction section 3 extracts and encodes the feature quantities of a plurality of features determined from the two-dimensional pattern in pattern memory section 2. It is stored in the code storage unit 8 together with the category name given as a value string. When the storage of the characters on the character sample form is completed, the dictionary generation section 9 uses the code value string in the code storage section 8 to create a code value appearance frequency distribution for each category for the feature H, and generates the obtained code. A plurality of code value ranges are determined, that is, divided, so that the number of categories within the code value range is reduced from the value appearance frequency distribution. Thereafter, the next divided dictionary is created for each code value string in the code storage unit 8 in which the code value corresponding to the feature H is within the code value range described above, and is stored in the dictionary unit 5. Combining the code values for each feature by making sure that the code value columns with the same category name do not include the code value columns of other categories, and then determining the lower limit code and upper limit value code as the range of code values, and specifying the category name. The dictionary element of the divided dictionary is expressed by the code value range for each feature.

Therefore, the dictionary is composed of a code value range sequence in which code values are arranged according to their features.

Also, when not using the character sample form,
A category name input section 10 gives a category name for the characters on the form.

FIG. 4 is a diagram showing an example of a code value string, in which category names and encoded code values of feature values for each feature are arranged. However, the number of features is assumed to be two for simplicity.

FIG. 5 is a diagram showing an example of the code value appearance frequency distribution obtained from the code value (front code value) for one feature of the code value sequence shown in FIG.

In the figure, 1 to 8 are code values Ci (i-1
~6) are category names, and the symbol 0 represents a frequency. One way to determine the code value range is to first select one code value that includes the largest number of categories, find code values that have categories that are included in the categories of the selected code value, and then This method uses a lower limit and an upper limit as a code value range.

First, code value 2, which has the largest number of categories, is selected, and code values including only categories C2, C3, C5, and C6 of code value 2 are successively determined to create a code value range. The category with code value 1 next to code value 2 is also included, and the category with code value 3 adjacent to one side is also included. Since the next category C4 with code value 4 is not included in the category with code value 2, a code value range of 1 to 3 is obtained.

The process is then repeated for the remaining code values.
If the next code value with the largest number of categories is the same, choose the one with the smaller code value, then code value 5 will be selected, and similarly the code value of one side that includes the category with code value 5 will be 4, other code value 6
Then, since category C2 is not included, the next code value range will be from 4 to 5. Furthermore, when the process is repeated for the remaining code values, code value 6 is selected first, and the code value range 6 to 8 is obtained from the inclusion relationship of code value 6 with the category.

The reading of the form is performed as follows.

The character pattern of one character on the form is photoelectrically converted by the scanning unit 1 and stored in the pattern memory unit 2 as image data, and the feature extraction unit 3 extracts the feature amount of the feature determined from the two-dimensional pattern in the pattern memory unit 2. The extracted and encoded code value range string of the divided dictionary read out corresponding to the feature code value of the feature H is compared with the code value string obtained by the feature extraction section 3, and a reading result 6 is output. The types of features extracted by the feature extractor 3 are roughly divided into two types: one is obtained by character line tracing, and the other is obtained by background analysis. The former converts characters into thin line patterns and traces the lines to detect the number of feature points such as end points, branch points, intersections, etc., their positional relationships, connections, curves between feature points, etc., and the latter converts the outline of characters. Track and recess,
Divided into convex parts, the degree of curvature of each part, the opening direction of each part,
The tracking length ratio of each part to the total length, the direction histogram of each part, etc. For example, the characteristic of the opening direction of the recess is used as the characteristic H.

FIG. 6 is a block diagram showing an embodiment of a character recognition device using a processor and memory to implement the character recognition method of the present invention corresponding to FIG. 11
12 is a scanning circuit that scans a predetermined pattern area;
13 is a pattern memory, 13 is a dictionary memory that stores a dictionary used for matching, 14 is a code memory that stores code value strings of category names and feature amounts used for dictionary creation, 15 is a program memory, and 16 is an output display for reading results. an output device; 17 is a key input circuit for modifying the output results; 18 is an auxiliary storage device that stores a feature extraction program, a collation program, a dictionary creation program, and a code value range creation program to be set in the program memory 15; 20
is a processor.

In order to perform the function shown in FIG. 3 with the character recognition device shown in FIG. 6, the following processing is required.

First, the processor 20 transfers the feature extraction program stored in the auxiliary storage device 18 to the program memory 15.
Set to . Next, when a character sample form is input, the characters on the form are scanned and quantized by a scanning circuit 11 and set in a pattern memory 12 as a binary pattern. Processor 20 is program memory 1
Run the feature extraction program set to 5.
Features are extracted from the binary pattern stored in the pattern memory 12, their feature amounts are determined and encoded, and the resulting code value string is stored in the mode memory 14 together with the category name given by the position on the form. When the characters on the character sample form are processed one after another and recorded in the code memory 14, a code value range determination process begins. When the processor 20 sets the code range creation program in the auxiliary storage device 18 into the program memory 15, it uses the code values in the code memory 14 corresponding to the specified feature (the feature H) to calculate the code value appearance frequency distribution. and find the code value range using the method described above. Next, divided dictionary creation processing begins for each code value range. The processor 20 sets the dictionary creation program in the auxiliary storage device 18 in the program memory 15, executes the program, takes out the code value string in the code memory 14 via the interface bus 19, and extracts the code range creation program specified by the code range creation program. A divided dictionary is generated and set in the dictionary memory 13 using only code value strings within the code value range for which code values corresponding to the same feature are found. After the above processing is completed for each code value range, actual reading of the form is performed.

When a form is input, the characters on the form are scanned and quantized by a scanning circuit 11, and set in a pattern memory 12 as a binary pattern. Processor 20
executes the feature extraction program stored in the program memory 15, extracts features from the binary pattern stored in the pattern memory 12, encodes each determined feature amount, converts it into a code value string, and at the same time converts the code of the feature H described above. The divided dictionary is read from the dictionary memory 13 based on the value.

Next, the processor 20 executes the collation program set in the program memory 15, collates the code value string of the determined feature material with the code value range string of the read divided dictionary, and outputs the result to the output device 16.
Output to.

FIG. 7 is a general symbol representation of FIG. 5 for determining the code value range, and the ○ marks in FIG. 5 are shown as "1", and the others are shown as "0". Also,
A sequence consisting of “0” and “1” for one code value i
When Vi is the number of categories and Ti is the number of categories, the code value range creation described above becomes the flowchart shown in FIG. In Figure 8, the process indicated by 10 is the process to detect the code value with the largest number of categories among the maximum or remaining code values, and MI is the detected code value, MT
represents the number of categories corresponding to the MI code value.
If the number of detected categories is 0, code value range creation ends. The process indicated by 20 is the code value described above.
The code value obtained by detecting a code value smaller than MI that has an inclusive relationship with category V of MI is LI, and the process of 30 described above detects a code value that is larger than MI that has an inclusive relationship with category V of code value MI. and create the UI. Code value range from the above processing
From LI to UI is determined, and multiple code value ranges can be obtained by repeating the process.

Figure 9 is a diagram of a code value string in which category names obtained from character samples and code values of several predetermined features are shown as symbols in order to create a divided dictionary. C1) 548
. . . 6, (C2) 826 . . . 5 code value strings are shown.

Here, it is assumed that all the code values for the feature H mentioned above are within the one code value range mentioned above.

In the figure, c is the category parameter that encodes the category name, k is the number of samples, and F(c,
Assuming that k) represents the code value of the feature amount, this means that the number of character samples is the same L for each category, the number of categories is N, and the number of features is M.

FIG. 10 is a flowchart for creating a divided dictionary using the symbols shown in FIG.

The process indicated by 110 is the process of clearing the memory location P (c, k) determined by the category parameter c and the sample number parameter k corresponding to the number of samples with the letter A, and clears a flag indicating whether it has already been used for dictionary creation. If P(c,k)=A, it means unprocessed.

The processing indicated by 120 is unprocessed, that is, P(c,
When k) = A, the process is to create the lower limit value F ₁ j and upper limit value F ₃ j of the feature value of feature Fj based on P (c, k), and P (c, k) - Y Represents something that has been processed.

The process indicated by 130 uses the same parameter value c as the category parameter value c specified in 120,
The raw P(c,
k), and the characteristics of the sample number parameter k
This is a process in which the special value of Fj is set to F ₂ j.

The process indicated by 140 is the process indicated by 170 of the feature values F ₁ j and F ₂ j, which is the process indicated by 130 and 14 described above.
This is a process for repeating the 0, 150, and 160 processes for a total number of samples L times by changing the sample number parameter k.

The process indicated by 180 is a process of creating one dictionary using the category parameter c and the lower limit value F ₁ j and upper limit value F ₃ j of the feature Fj.

The process indicated by 190 is a process for changing the sample number parameter h and repeating the above process L times for the total number of samples.

The process indicated by 200 is a process for changing the number of categories parameter c and repeating the above-mentioned dictionary creation process for each c a total of N times, the total number of categories.

Therefore, the created divided dictionary is composed of the code value c of the category name, the lower limit value code F ₁ j and the upper limit value code F ₃ j of the feature amount for each feature, as shown in FIG.

The above process is repeated for each code value range, and the set of divided dictionaries is the dictionary of this recognition method.

In this range creation method, when the number of data doubles, the number of repetitions increases approximately four times. Therefore, the amount of data that needs to be processed at the same time as the above-described division is reduced, and the dictionary creation time can be shortened. For example, if you divide it into four parts, the total dictionary creation time will be reduced to 1/4.

An example of the final verification processing method is shown below.

If the code value strings of the features obtained by executing the feature extraction program from the character pattern to be read are F _I1 , F _I2 . . . F _IM , then the code value range selected by the code value of the feature H is The smaller value of the divided dictionary in is Fjn, and the feature values F ₃ j and F ₂
This is a process in which the larger value of j is set to Fjm.

The process indicated by 150 calculates the difference Dal between the feature value Fj (a, l) at a position determined by the category parameter a other than c and the sample number parameter l and the Fjn, Fjm using the following formula, and This is a process in which D is the minimum difference amount obtained by changing a and the sample number parameter l.

Dal＝ _M 〓 ^j=1 Wj [Fjn − Fj (a, l)] + _M 〓 ^j=1 Wj [Fjn (a, l) − Fjm] However, [θ] = 0 (θ0), [θ] = θ(θ>0
) Here, Wj is the weight of feature Fj, which is assumed to be determined in advance through statistical processing.

In the process shown at 160, if the minimum difference amount D is equal to or greater than the threshold T, Fjn is set to the lower limit value F ₁ j of the feature Fj, Fjm is set to the upper limit value F ₃ j of the feature Fj, and the flag P(c, k) is set to Y.
Add it and consider it as processed.

A difference amount D(b) is calculated using the upper limit code F ₁ j(b) and the upper limit value F ₃ j(b).

D(b)= _M 〓 ^j=1 Wj [F ₁ j(b) − F _Ij ] + _M 〓 ^j=1 [F _Ij − F ₃ j(b)] However, [θ] = 0 (θ0), [θ]=θ(θ>
0), Wj is the weight of feature Fj.

The code value c of each category corresponding to b that has the minimum difference amount from b=1 to B is taken as the reading result of the character to be read.

The feature of the present invention is that multiple code value ranges are obtained based on the number of categories from the code value appearance frequency distribution obtained by encoding the feature values of macro features, and a divided dictionary is created for each code value range. By doing so, the number of data that needs to be considered at the same time is reduced, and the dictionary creation time is shortened. In the explanation so far, a divided dictionary is created using a single feature, but a divided dictionary can be created in the same way by using a set of multiple special features.

As explained above, according to the present invention, a dictionary can be created in the character reading device after the feature values are encoded and stored as a code string, and a dictionary for the characters of the form to be read can be generated, resulting in high performance. It becomes possible to obtain a character reading device, and its effects are significant.

[Brief explanation of drawings]

Fig. 1 is a diagram showing an example of a character sample form for dictionary creation, Fig. 2 is a block diagram of a conventional character recognition method, and Fig. 3 is a concrete implementation of the character recognition method according to the present invention. A block diagram showing an example, FIG. 4 is a diagram showing an example of a code value string, FIG. 5 is a diagram showing an example of the code value appearance frequency distribution, and FIG. Fig. 7 is a diagram showing the code value appearance frequency distribution of Fig. 5 using symbols, and Fig. 8 shows the code value range using the symbols shown in Fig. 7. A diagram showing an example of a flowchart to create,
Figure 9 shows the category names obtained from character samples and the code values of several predetermined features for making a dictionary. Figure 10 shows the symbols in Figure 9. FIG. 11 is a diagram showing an example of a dictionary format. In the figure, 1 is a scanning section, 2 is a pattern memory section, 3 is a feature extraction section, 4 is a matching section, 5 is a dictionary section,
6 is an output result, 7 is an auxiliary storage section, 8 is a code storage section, 9 is a dictionary generation section, 10 is a category name input section,
11 is a scanning section, 12 is a pattern memory section, 13 is a dictionary memory, 14 is a code memory, 15 is a program memory, 16 is an output device, 17 is a key input circuit, 18 is an auxiliary storage device, 19 is a bus line, 2
0 indicates a processor.

Claims (1)

[Claims] 1 A dictionary created from the feature quantities of features extracted from the characters on the form is stored in the character reading device in advance, and when reading the form, the feature quantities of the determined features are extracted from the characters on the form, and the dictionary is In character recognition, a character sample form is input before reading starts, and a string of code values is encoded with the category name given to each character and the feature quantities of multiple predetermined features. When the encoding of the characters on the character sample form is completed, first, the number of categories is calculated from the code value appearance frequency distribution for each category in one of the features (hereinafter referred to as feature H). Then, using the stored column of category names and code values in which the code value corresponding to the feature H is within the one code value range obtained, The code value column does not include the code value column of other categories, and the code values are combined for each feature to find the lower limit code and upper limit code, and the code value range is set as the category name and each feature. A character recognition method characterized in that dictionary elements of a divided dictionary are created based on code value ranges, and the dictionary is represented by a set of divided dictionaries created for each of the code value ranges.