JPH0927014A - Dictionary data learning system of handy optical character recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate new dictionary data in a short time from many sample characters by outputting feature data extracted from read data on plural sample characters to a host computer. SOLUTION: An OCR 10 reads sample characters printed by >=2 different printing means as to one reference character from an image sensor 1 and outputs feature data extracted from read data on respective sample characters to the host computer 6. The host computer 6 generates dictionary data on the reference character from the feature data on >=2 different sample characters and outputs them to a microcomputer 4 together with educational data corresponding to the reference character. Consequently, the host computer 6 which can perform fast processing can performs calculation for the dictionary data generation. The new dictionary data are generated by processing feature data together for >=2 sample data, thus it is hardly affected by the read order of sample characters is little.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to dictionary data for creating dictionary data stored in a dictionary data storage unit of a handy type optical character recognition device for recognizing an optically read character from a plurality of sample characters. Regarding learning methods.

[0002]

[Prior art]

A handy type optical character recognition device (hereinafter referred to as OC
In R) 100, dictionary data for comparison with characters printed by a printer, for example, is stored in the dictionary data storage unit, and character recognition is performed by comparing this dictionary data with the read characters. .

That is, as shown in FIG. 7, a minute analog signal of a character to be read, which is read by an image sensor 101 such as a CCD (charge coupled device), is amplified by an amplification / binarization circuit 102 and then amplified. , And is converted into a binary signal (a signal of 0 or 1) representing light and dark, and the read data is written in the image data storage unit 103 formed of a RAM (random access memory).

On the other hand, an electrically rewritable ROM (EE
A dictionary data storage unit 104 including a ROM) stores dictionary data created from read data of a specific typeface character set, for example, a set of Century typeface print characters, based on the characteristics of the character image. .

The microcomputer 105 connected to the image data storage unit 103 and the dictionary data storage unit 104 operates in two modes, a character recognition mode and a dictionary learning mode. In the character recognition mode, the image data storage unit 103 is operated. The read data read from the dictionary data is compared with the dictionary data stored in the dictionary data storage unit 104 to perform character recognition of the read character.

That is, the two are compared for each feature of each part,
When the degree of disagreement k between the two is within a predetermined range, it is determined that the read character can be recognized, and the recognition data of the read character is converted into a predetermined character code, for example, an ASCII code.
The data is output to the host computer 106 connected to the OCR 100.

This character recognition will be described in more detail.
The read data is divided into m masks as shown in 6 (a),
Assigns the read data value for each mask to m components
Read data as an m-dimensional vectory= (Y₁, Y_Two, Y
_Three... y_m), And the feature data
M-dimensional vector corresponding to the dataf= (F₁, F_Two, F
_Three... f_m) Dictionary data consisting of
You.

[0009]

(Equation 3)

The degree of disagreement k with a certain dictionary data obtained in step S1 is smaller than a constant C ₁ , that is, k <C ₁ , and the degree of disagreement between this degree of disagreement k and the second smallest dictionary data. The difference from k ₂ is C ₂ or more, that is, k ₂ −k>
When C ₂ is satisfied, it is determined that the read data is a character related to the dictionary data.

On the other hand, in the dictionary learning mode, when new dictionary data is created in the dictionary data storage unit 104 or dictionary data matching the font of the character to be recognized in order to increase the character recognition rate is created. The dictionary data storage unit 104 stores these dictionaries.

The operation of the OCR 100 and the host computer 106 in this dictionary learning mode will be described with reference to the flowchart of FIG.

First, in order to set the microcomputer 105 of the OCR 100 to the dictionary learning mode, the host computer 106 outputs a learning mode designating signal requesting the setting to the dictionary learning mode (step S101).

Upon receipt of this signal, the microcomputer 105 sets itself in the dictionary learning mode, and then, in response to the response signal, the host computer 1
06 is notified (step S102).

The operator confirms this notification on the display of the host computer 106 and then reads from the image sensor 101 the sample character of the font to be stored as dictionary data as the read character. The microcomputer 105 outputs the read data of this sample character to the host computer 106 (step S).
At the same time as 103), the read data is divided into m masks and the feature data x consisting of an m-dimensional vector is extracted from the read data, as in the character recognition mode.

While checking the read data displayed on the display of the host computer 106, the operator inputs the corresponding correct character code as educational data, and the host computer 106 outputs this educational data to the microcomputer 105. Yes (step S10
4).

The microcomputer 106 reads out the dictionary data f corresponding to the educational data output from the host computer 106 from the dictionary data storage unit 104,
The dictionary data is rewritten by creating new dictionary data f ′ from the extracted feature data x and the dictionary data f , writing the new dictionary data f ′ in the dictionary data storage unit 104 in association with the education data.

By the way, generally, even if the same font is used, different images are formed due to differences in printer models and the like, and when creating dictionary data of a new font, at least one reference character (for example, A) is used. The sample characters printed by several types (for example, n types) of printers are read, and each of these characteristics is incorporated to create dictionary data.

Therefore, if the average value of the feature data x and the dictionary data f is taken every time the sample character is simply read to obtain new dictionary data f ', the feature of the sample character read later has a greater effect on the new dictionary data. give.

Therefore, in Japanese Patent Laid-Open No. 6-96283, new dictionary data is created by adding a constant learning speed coefficient α (eg, = 0.9) to the feature data of sample characters.
This problem is solved by creating f ′ by f ′ = α f + (1-α) x and weighting the existing dictionary data.

[0021] Thus, when stored in the dictionary storage unit 104 in association with the new dictionary data f 'to the education data, the microcomputer 105 outputs a learning completion signal to the host computer 106 (step S10
5).

The learning end signal is sent to the host computer 106.
When the display is confirmed, the operator continuously reads the sample characters printed by another printer from the image sensor 101.

The same processing from step S103 to step S105 is repeated until all n kinds of sample characters are read and stored as new dictionary data.

When the last sample character is read and the learning completion signal is confirmed on the display of the host computer 106, the operator finds the microcomputer 105
Host computer 1 to return the character to the character recognition mode.
A recognition mode designation signal for requesting the setting to the character recognition mode is output from 06, and the dictionary learning mode is ended (step S106).

[0025]

However, the conventional dictionary data learning method using the OCR 100 has the following problems.

That is, first, in the case of creating new dictionary data with a plurality of sample characters, in the above conventional example, the influence of the reading order is reduced by weighting the previously created dictionary data. However, the finally created dictionary data differs depending on the reading order of the sample characters, and is not an average of the characteristics of n kinds of sample characters.

Secondly, since the dictionary data is updated for each sample character, when a new dictionary data is created with a plurality of samples, the learning end signal from the OCR 100 must be waited each time the work is interrupted and the processing is performed. It takes time.

Thirdly, every time a new sample character is read, f ′ = α f + (1-α) x must be calculated to create new dictionary data, which is a practical obstacle. In order to carry out this calculation and store new dictionary data in the dictionary storage unit 104 within a short time, the microcomputer 105 that operates at high speed must be installed in the OCR 100.

The above-mentioned second and third problems can be solved by storing read data of all sample characters in the OCR 100 and calculating new dictionary data by the microcomputer 105 capable of high speed processing. OCR
It is necessary to mount a memory 103 having a capacity for storing the read data of all sample characters and a microcomputer 105 capable of high-speed processing in 100.

However, the handy type optical character recognizing device 100 is required to be downsized and low in price in order to differentiate it from a large stationary character recognizing device, and the memory 103 having a large capacity and the like. It is not possible to mount the expensive microcomputer 105 that operates at high speed.

Fourth, since character recognition is compared with a constant,
Setting the constant increases the read error.

That is, the dictionary data is represented only by an m-dimensional vector, and the distance between vectors (the degree of disagreement k) with the feature data of the sample character is compared with a constant C ₁ .
Since it is a condition for character recognition, the influence of the degree of variation of sample characters to be learned is ignored, and if the constant C ₁ is too small for the variation of the characteristic data of the sample characters, many "unrecognizable" errors occur, If the constant C ₁ is too large with respect to the variation in the characteristic data of the sample characters, many errors "recognize erroneous characters" occur.

On the other hand, it is difficult to visually confirm the degree of variation from the sample character group, and there is a problem that the constant C ₁ must be set while repeating character recognition.

Further, if the constant C ₁ is large, the dictionary data of a plurality of characters and the character recognition condition are often satisfied. Even if the character recognition condition is satisfied by the dictionary data of a certain character, the character data will not be matched with other characters. The distance and the constant C ₂ must be used for comparison, and it took time for character recognition.

The present invention has been made to solve the above-mentioned problems, and new dictionary data can be created from a large number of sample characters in a short time, and the created dictionary data can be read from the sample characters. It is an object of the present invention to provide a dictionary data learning method for a handy type optical character recognition device that is not affected by the order and does not cause a character recognition error even if there are variations between sample characters.

[0036]

In order to achieve the above object, a dictionary data learning method of a handy type optical character recognition apparatus according to the present invention is an image sensor for optically reading a character to be read and an output from the image sensor. The storage unit for storing the read data of the read character, the dictionary data storage unit for storing the dictionary data, the character recognition mode for performing the character recognition of the read character by comparing the read data with the dictionary data, and the read data A handy type optical character recognition device provided with a microcomputer that operates in two modes, a dictionary learning mode, for storing dictionary data created based on the dictionary data storage section, and a handy type optical character recognition device in the dictionary learning mode. A host that outputs the character code corresponding to the read data of the read character as educational data to the microcomputer of the character recognition device Computer, which reads sample characters as read characters for dictionary creation with an image sensor, uses a microcomputer that operates in dictionary learning mode to extract characteristic data from the sample character read data, and creates it based on the characteristic data. In the dictionary data learning method of the hand-held optical character recognition device that rewrites the dictionary data by storing the stored dictionary data in association with the education data in the dictionary data storage unit, the hand-held optical character recognition device is an image sensor. Read the sample characters printed by at least two different printing means for one reference character, and output the characteristic data extracted from the read data of each sample character to the host computer, and the host computer outputs two or more different samples. Of the reference character from the character feature data Create a write data, along with education data corresponding to the reference characters and outputs to the microcomputer.

According to a first aspect of the present invention, an image sensor is provided.
The sample character printed by at least two different printing means is read for the reference character of the character, and the characteristic data extracted from the read data of each sample character is output to the host computer. Therefore, it is not necessary to provide a memory capacity for storing the read data of each sample character.

The host computer creates the dictionary data of the reference character from the characteristic data of two or more different sample characters and outputs it to the microcomputer. This host computer is a host computer that outputs educational data, and can also be used as a host processing device that inputs the code of the character recognized by the handheld optical character recognition device and processes the data. Data can be created.

Further, since the new dictionary data is created by collectively processing the characteristic data of two or more sample data, it is not affected by the reading order of the sample characters.

Further, in the dictionary data learning method of the handy type optical character recognition device according to the second aspect, the dictionary data of the reference character is defined as the average value of the feature data of two or more different sample characters and the variance amount of the feature data. It is characterized by being created with a square root σ.

According to the second aspect of the present invention, the dictionary data of the reference character is represented by the average value of the feature data of two or more different sample characters and the square root σ of the variance of the feature data. By comparing the distance (the degree of disagreement k) between the vectors and and the constant C ₁ based on the square root σ of the variance, a character recognition error does not occur even if the variations in sample characters are different.

Further, in the dictionary data learning method of the handy type optical character recognition device of the third aspect, the characteristic data of the sample character divides the read data of the sample character into m masks, and the sample character for each mask. Is a m-dimensional vector x that is assigned to each of m components, and the number of sample characters is n, the dictionary data is

[0043]

(Equation 1)

And f , which represents the average of the characteristic data of the sample characters,

[0045]

(Equation 2)

And is represented by σ representing the square root of the variance of the feature data.

According to a third aspect of the present invention, the average value is calculated from the characteristic data of two or more sample characters represented by the m-dimensional vector x.

[0048]

(Equation 1)

The square root of the amount of variance is calculated from

[0050]

(Equation 2)

Since the respective dictionary data are obtained from the above, the average and the degree of variation of the read data as the image data can be obtained by a calculation formula suitable for the processing by the host computer.

[0052]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configurations of the OCR 10 and the host computer 6 according to the present invention. In FIG. 1, 1 is an image sensor including a CCD,
The read character is optically read and converted into an electric signal.

Reference numeral 2 is an amplification / binarization circuit for amplifying a minute analog signal input from the image sensor 1 and then converting it into a binary signal representing a gradation. 3 is SRAM
In the storage unit that also serves as the image data storage unit such as (static RAM), the binary signal representing the light and dark output from the amplification / binarization circuit 2 is subjected to pattern processing as shown in FIG. 6A. It is stored as read data.

That is, at the address of the storage unit 3 corresponding to each position of the character to be read, the light / dark information of the binary signal at that position (for example, 0 for a light dot in the background color portion and 1 for a dark dot in the character portion). Is stored, and read data composed of image data similar in shape to the character to be read is formed on the memory of the storage unit 3.

A microcomputer 4 which operates in two modes, a character recognition mode and a dictionary learning mode, is connected to the latter stage of the storage unit 3. In any mode, the microcomputer 4 extracts a character string, cuts out a character from the read data stored in the storage unit 3,
Each process of normalization and feature extraction is executed.

The character string extraction is to extract read data including a plurality of characters from the entire read data when a plurality of characters are read at the same time.

The character cut-out is to cut out read data corresponding to a read character for each character from the read data. For example, in the read data, when the periphery of the portion where the dark dots “1” are continuous is surrounded by the bright dots “0”, the portion around the dark dots “1” is cut out as one character.

Since the cut-out one-character portion is compared with the dictionary data, it is adjusted to the same size as when the dictionary data was created by the normalization process, and the normalized read data is output to the host computer 6. At the same time, the characteristics of each part are extracted from the normalized read data.

In the feature extraction, for example, in the storage unit 3, the normalized read data is divided into 6 vertically and horizontally and is divided into 36 masks in total, and the mask is calculated from the number and position of dark dots "1" in each mask. The feature number, which is the value of the read data representing the read character, is calculated for each. When calculating the number of features, the central position of the mask is weighted and the number of dark dots “1” in the mask is totaled.

For example, regarding the number of features of the lower right mask in FIG. 6A, the number of dark dots "1" around the mask in FIG. 6B is "6", and the number of bright dots "0" in the center of the mask is "6". Is “2”, the sum of the number “1” around the mask “6” * the weighting amount “1” and the number “1” in the mask center “2” * the weighting amount “2” is “10”. It is represented by.

Similarly, the characteristic number of all 36 masks is obtained, and the characteristic data of the character to be read is represented by a 36-dimensional vector in which this is assigned to 36 components.

That is, assuming that the feature data is a 36-dimensional vector x , this vector x has the feature numbers x ₁ and x ₂ of each mask when the masks of FIG. 6A are ordered in the horizontal direction from the top. , X ₃ , ... X ₃₆ , it can be expressed as x = ( x ₁ , x ₂ , x ₃ , ..., x ₃₆ ).

In the character recognition mode, the microcomputer 4 compares the characteristic data x with the dictionary data and outputs the recognition data of the read character to the host computer 6 with a predetermined character code. The data x is output to the host computer 6, and new dictionary data input from the host computer 6 is stored in the dictionary data storage unit 5.

A dictionary data storage unit 5 is composed of a non-volatile memory such as an electrically rewritable ROM (EEPROM) and stores the dictionary data to be compared with the read character.

This dictionary data storage unit is connected to the microcomputer 4 and can read dictionary data into the microcomputer 4 or write new dictionary data from the microcomputer 4.

As shown in FIG. 1, the OCR 10 having these configurations further includes a host computer 6 capable of bidirectional data communication with the microcomputer 4.
Is connected.

The host computer 6 is provided with a data input section 7, a display section 8 including a display, and a characteristic data storage section 9 for storing characteristic data. The data input unit 7 is a keyboard, and an operator operates the keyboard to select an operation mode of the microcomputer 4 or perform predetermined control described later.

The display which is the display unit 8 displays normalized read data, a temporary recognition result described later, and the like.

Next, a method for creating new dictionary data by the OCR 10 and the host computer 6 having such a configuration will be described with reference to the flowchart of FIG.

First, the operator operates the data input section 7 to put the microcomputer 4 of the OCR 10 into the dictionary learning mode, and the host computer 6 sends a learning mode designating signal for requesting the setting of the dictionary learning mode to the OCR 10.
To (step S1).

After receiving this signal, the microcomputer 4 sets itself in the dictionary learning mode, and then, in response to the response signal, the host computer 106 sets this mode.
To (step S2).

The operator confirms this notification on the display unit 8,
Next, regarding the characters of the font to be stored as dictionary data, a plurality of types of sample characters (for example, 10 types) printed by various printers are read from the image sensor 1. At this time, it is not always necessary to continuously read the 10 types of sample characters at once.

In this embodiment, a plurality of sample characters (for example, "A", "B", "C") from one type of printer are used for the characters of the font to be stored as dictionary data.
"D", "E", "3", and "4" (7 characters) are printed, and these sample characters are read at the same time.

Each time a plurality of sample characters are read from the image sensor 1 as the characters to be read, the microcomputer 4 executes the above-mentioned character string extraction and character cutout processing, and cuts out read data for each sample character.

Then, the cut-out read data is subjected to normalization and feature extraction processing to extract feature data x from the read data of the sample characters.

At the same time, the microcomputer 4
Although it is a dictionary learning mode, like the character recognition mode,
By calling the temporary dictionary data stored in the dictionary data storage unit 5 and comparing it with the characteristic data x ,
Performs temporary character recognition of sample characters. A method for character recognition by comparing the characteristic data x with the dictionary data will be described later.

When the temporary character recognition of one sample character "A" is completed, the normalized read data of the sample character, the feature data x , and the temporary recognition data recognized by the temporary character recognition are temporarily stored in the storage unit 3. Remembered.

The microcomputer 4 further cuts out the read data of the next sample character "B" from the extracted character string and executes the same processing as described above to read all the sample characters contained in the extracted character string. Repeat the process for the data.

For the read data of all the sample characters included in the extracted character string, read data and characteristic data
x , when the temporary recognition data is stored in the storage unit 3, these data are sequentially read from the storage unit 3 and output to the host computer 6 (steps S3 to S5).

That is, in step S3, the temporary recognition data recognized by the temporary character recognition from the first sample character "A" to "4" are sequentially output to the host computer 6 by a predetermined character code.

Then, in step S4, the code indicating the order of the sample characters and the read data of the sample characters are continuously output in order from the first sample character.
Further, in step S5, a code indicating the order of the sample characters and the characteristic data x of the sample characters are continuously output in order from the first sample character.

The host computer 6 displays the temporary recognition result and the read image of the sample character on the display unit 8 from the temporary recognition data and the read data among the input data.

FIG. 3 shows a state in which the temporary recognition result and the read image are displayed on the display section 8. The read “A”, “B”, “C”, “D”, “E”, “3” and “4” are displayed. The temporary recognition result (11 ₁ , 11 _2, ... 11 ₇ ) of the sample characters and the read image (12 ₁ , 12 _2, ... 12 ₇ ) are displayed in the frame 20 below the display, divided for each sample character. ing.

The operator confirms the temporary recognition result by comparing this display with the read image, and if there is a sample character that is erroneously recognized, the operator inputs the correct character from the data input unit 7 and tentatively recognizes the character. Correct the result. For example, when the read image 12 ₂ shows the character “B”, but the temporary recognition result 11 ₂ shows “D” and there is an erroneous recognition, the cursor on the display is moved to the temporary recognition result 11 _2.
In accordance with the above, the correct character "B" is input from the data input unit 7.

When the operator confirms that the displayed temporary recognition results of all the sample characters correspond to the read image, the operator operates the data input section 7 to confirm this. By this confirming operation, the character code corresponding to the temporary recognition result displayed at that time becomes the educational data of the sample character and is input to the host computer 6.

At this time, the characteristic data input from the OCR 10 is stored in the characteristic data storage section 9 in association with the above-mentioned educational data for each sample character.

The display frame 21 displayed above the display of FIG. 3 shows for each character how many kinds of sample character characteristic data are stored in the characteristic data by different types of printers. Indicates a character to be stored as dictionary data. In the present embodiment, the characters to be stored as dictionary data are the alphabets from "A" to "Z" and the numbers from "0" to "9", so 36 characters 13 ₁ , 1
3 ₂ ... 13 ₃₆ is displayed for each frame.

The number of read sample types in the lower part of the display frame 21 is 14
Is the character data displayed in the upper part of the character data of several different sample characters by the method described above.
It shows whether or not it is stored in the characteristic data storage unit 9.

Therefore, as described above, "A""B""C"
When the characteristic data is newly stored for the sample characters “D”, “E”, “3”, and “4”, the read sample type numbers 14 ₁ , 14 ₂ displayed below the characters in the display frame 21.
・ ・ Will be displayed with 1 added.

If the number of samples n required to create new dictionary data is "10", the number of read sample types 14 in the lower part of the display frame 21 is set to "10" or more, and Sample characters printed by a different printer are prepared and read again from the image sensor 1 of the OCR 10.

In this way, when the processes from step S3 to step S5 and the above-described process in the host computer 6 are repeated and all the read sample types 14 in the lower part of the display frame 21 become "10" or more, The operator operates the data input unit 7 to instruct the host computer 6 to create new dictionary data, assuming that necessary feature data is prepared.

In response to this instruction, the host computer 6 creates dictionary data for each character in order from "A". That is, if one character of a character for which new dictionary data is to be created is used as a reference character, 10 or more characteristic data having the reference character as educational data are read from the characteristic data storage unit 9, and the characteristic data is used as a characteristic. New dictionary data composed of the average value of the data and the square root σ of the variance of the feature data is created.

If the characteristic data x is represented by an m-dimensional vector, and the number of characteristic data for the reference character, that is, the number of sample characters is n, and the characteristic data x of the i-th sample character is x _i , The average value f of the data is

[0095]

(Equation 1)

It can be obtained by, and is represented by an m-dimensional vector.

Further, if the j-th component of the i-th feature data x is x _ij and the j-th component of the average value f represented by the m-dimensional vector obtained above is f _j , The square root σ of the variance is

[0098]

(Equation 2)

It can be obtained by

With respect to one reference character, when dictionary data consisting of the average value f and the square root σ of the variance is created, the reference character is moved to another character and the same calculation is performed to obtain the dictionary data of all 36 characters. create.

When all the new dictionary data is created, the host computer 6 creates a character code, which is educational data, for each character in order from the first character, and then an average value f of the dictionary data corresponding to the educational data. And the square root σ of the variance amount are output to the microcomputer 4, and the microcomputer 4 is instructed to associate these dictionary data with the educational data and write them in the dictionary data storage unit 5 (step S6).

The microcomputer 4 stores new dictionary data in association with the education data in the dictionary data storage unit 5, and then notifies the host computer 6 that the dictionary data has been rewritten (step S7), and the operation is performed. The person confirms this by the display on the display unit 8.

Next, a method of recognizing a character to be read by using the dictionary data thus created will be described with reference to the flowchart of FIG. 4 showing the operation in the character recognition mode.

When the operator operates the data input section 7 to set the microcomputer 4 of the OCR 10 to the character recognition mode, the host computer 6 outputs a recognition mode designating signal requesting the setting of the character recognition mode to the OCR 10 ( Step S10).

Upon receiving this signal, the microcomputer 4 sets itself to the character recognition mode, and then notifies the host computer 6 of the setting of this mode by a response signal (step S11).

The operator confirms this notification on the display unit 8,
Then, the read character to be recognized is read from the image sensor 1.

Every time when the character to be read is read from the image sensor 1, the microcomputer 4 performs the character string extraction and character cutout processing as in the dictionary learning mode, and cuts out the read data for each character to be read.

Subsequently, the cut-out read data is subjected to normalization and feature extraction processing to extract feature data y from the read data of the read character.

Character recognition of the read character is based on this feature data.
This is performed by comparing y with dictionary data.

First, the dictionary data is stored in the dictionary data storage unit 5
The characteristic data y and the average value f are compared. This comparison is performed by finding the distance k between the two vectors, since the feature data y and the average value f are both represented by an m-dimensional (36-dimensional in this embodiment) vector.

That is, assuming that the distance k is the degree of disagreement with the dictionary data of a certain specific character, the degree of disagreement k is represented by the feature data y and the j-th component of the average value f as y _j and f _j , respectively.

[0112]

(Equation 3)

It can be obtained by

When the dissimilarity k is less than or equal to the square root σ * 3 of the variance which is dictionary data of the same specific character, that is, when k <3σ is satisfied, the read character is a character related to the dictionary data. To determine.

Here, the variance amount σ ² is a measure of the variation of the n kinds of feature data described in the dictionary learning mode. The variance amount σ ² increases when the variance is large, and the variance amount σ when the variance is small. ² becomes smaller. Statisticalally, in the variation of the characteristic data according to the normal distribution, 99.8% of unknown characteristic data is included in the average value f + 3σ of the characteristic data. That is, if the characteristic data y of the character to be read is the average value f ± 3σ of the characteristic data, it can be determined that the characteristic data is within the same character variation range and is otherwise an abnormal value, that is, characteristic data different from the character. .

[0116] Therefore, with respect to the mean value f in m-dimensional space as shown in FIG. 5, the mean value feature data y ₁ distance k ₁ is to be read characters is within 3σ with f are to be read characters mean that It can be determined that the character has the dictionary data of f .

On the other hand, the characteristic data y ₂ of the read character whose distance k _{2 from the} average value f is 3σ or more can be determined to be not the character having the dictionary data of the average value f , and the other The same comparison as the dictionary data for characters is performed.

By comparing the distance k with the average value f with 3σ in this way, it is possible to incorporate the degree of variation of the sample characters into the judgment condition of character recognition, as compared with the case of comparing with a constant. The error "Unrecognized" and the error "Error recognition" do not occur frequently.

Further, it is not necessary to determine the character recognition condition twice for the dictionary data of one character, and the character recognition can be performed in a short time.

Upon recognizing the read character as a specific character, the microcomputer 4 outputs the character code of the specific character to the host computer 6 as recognition data (step S12). Further, even when the character recognition condition is not satisfied for any of the dictionary data relating to any character, the character code of “*” indicating “unrecognizable” is output to the host computer 6 as recognition data.

When the character recognition is completed for one character to be read, the microcomputer 4 further cuts the read data of the next character to be read from the extracted character string and executes the same process as above to extract the extracted character. The process is repeated for the read data of all the read characters included in the column.

When the recognition data for all the characters to be read is output to the host computer 6, the microcomputer 4 notifies the host computer 6 that the character recognition is completed (step S13), and the character recognition mode is completed.

In the above embodiment, an example in which the host computer 6 and the OCR 10 are connected one-to-one has been described.
However, the present invention is not limited to this. For example, two or more OCRs 10 may be connected to one host computer 6, and the common host computer 6 may be used to create dictionary data for each 0CR 10.

With this configuration, the calculation for creating the dictionary data can be performed by the host computer 6 capable of high-speed processing, and it is necessary to provide each OCR 10 with a large-capacity memory and a high-speed operating microcomputer 4. Therefore, the OCR 10 can be reduced in size and price.

Further, in any of the above-mentioned embodiments, the host computer 6 and the OCR 10 do not necessarily have to be always connected, and may be connected only when new dictionary data is created in the OCR 10.

Further, the host computer 6 uses the OCR1
It is preferable to use a host computer that inputs recognition data obtained by character recognition by 0 and processes the recognition data, but not limited to this, a host computer that processes recognition data for character recognition and a host that creates new dictionary data. The computers may be different.

[0127]

As described above, according to the present invention, O
The CR 10 only outputs the characteristic data extracted from the read data of each sample character to the host computer 6, and the new dictionary data is created by the host computer 6 only.
Therefore, it is not necessary to provide the OCR 10 with a large memory capacity and an expensive microcomputer that operates at high speed. Therefore, it is possible to reduce the size and cost of the OCR 10.

Since the new dictionary data created by the host computer 6 is created by collectively processing the characteristic data of two or more sample data, it is not affected by the reading order of the sample characters. Also, because it is not necessary to update the dictionary data for each sample character, there is no need to wait during the dictionary data creation process each time, and multiple sample characters can be read continuously and dictionary data created in a series of operations. it can.

According to the second aspect of the present invention, the dictionary data of the reference character is represented by the average value of the feature data of two or more different sample characters and the square root σ of the variance of the feature data. By comparing the distance between vectors with the feature data (degree of disagreement k) with the boundary value based on the square root σ of this variance, even if the variations in the feature data of the sample characters are different, character recognition error Does not occur.

Further, it is not necessary to determine the character recognition condition twice for the dictionary data of one character, and the character recognition can be performed in a short time.

According to the third aspect of the present invention, the average value and the degree of variation are obtained from the characteristic data of two or more sample characters represented by the m-dimensional vector x by a calculation formula suitable for processing by the host computer. be able to.

[0132]

[Brief description of drawings]

FIG. 1 is a handy type optical character recognition device 1 according to the present invention.
2 is a block diagram showing the configurations of 0 and a host computer 6. FIG.

FIG. 2 is a flowchart showing operations of the host computer 6 and the OCR 10 in the dictionary learning mode.

FIG. 3 is a diagram showing an operation screen displayed on a display unit 8 in a dictionary learning mode.

FIG. 4 is a flowchart showing operations of the host computer 6 and the OCR 10 in the character recognition mode.

FIG. 5 is an explanatory diagram in which a mean value f represented by an m-dimensional vector and characteristic data y ₁ and y ₂ of a read character are represented in a two-dimensional plane for explanation.

FIG. 6 shows a storage state of read data stored in an image data storage unit. FIG. 6A shows an image data storage unit in which 36 masks are formed by vertically and horizontally dividing into 6 parts, and FIG. It is explanatory drawing which expanded the lower right mask of the image data storage part.

FIG. 7 is a block diagram showing configurations of a conventional handy type optical character recognition device 100 and a host computer 106.

FIG. 8 is a conventional OCR 100 and host computer 10.
6 is a flow chart showing the operation of FIG.

[Explanation of symbols]

 1 Image Sensor 3 Storage Unit 4 Microcomputer 5 Dictionary Data Storage Unit 6 Host Computer 10 Handheld Optical Character Recognition Device (OCR)

Claims (3)

[Claims] 1. An image sensor (1) for optically reading a read character, a storage section (3) for storing read data of the read character output from the image sensor (1), and dictionary data. A dictionary data storage unit (5), a character recognition mode in which the read data and the dictionary data are compared to recognize the character to be read, and dictionary data created based on the read data are stored in the dictionary data storage unit (5). A handy type optical character recognition device (1) having a microcomputer (4) which operates in two modes of a dictionary learning mode which is stored in
0) and a host computer (6) that outputs a character code corresponding to the read data of the read character as educational data to the microcomputer (4) of the handy type optical character recognition device (10) in the dictionary learning mode. The sample character is read by the image sensor (1) as a read character for dictionary creation, and the feature data is extracted from the read data of the sample character by the microcomputer (4) operating in the dictionary learning mode. In the dictionary data learning method of the handy-type optical character recognition device, in which dictionary data created based on the dictionary data is rewritten by storing the dictionary data in the dictionary data storage unit (5) in association with the educational data. The character recognition device (10) has less than one reference character from the image sensor (1). Also 2
The sample characters printed by the different printing means described above are read, and the feature data extracted from the read data of each sample character is output to the host computer (6), and the host computer (6) features the two or more different sample characters. Create standard character dictionary data from the data,
A dictionary data learning method for a handy type optical character recognition device, which outputs to a microcomputer (4) together with educational data corresponding to a reference character. 2. The reference character dictionary data is created by an average value of the feature data of two or more different sample characters and a square root σ of the variance amount of the feature data.
Dictionary data learning method for the handy type optical character recognition device described. 3. The characteristic data of a sample character is an m-dimensional pattern in which the read data of the sample character is divided into m masks and the value of the read data indicating the sample character for each mask is assigned to each of the m components. is a vector x, the sample number of characters when it is n, dictionary data, [number 1] And f , which represents the average of the feature data of the sample characters, and 3. The dictionary data learning method for a handy type optical character recognition device according to claim 1 or 2, characterized in that it comprises σ representing the square root of the variance of the feature data.