JPH0991385A - Character recognition dictionary adding method and terminal ocr device using same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily add recognition data on a character which has individual features to a character recognition dictionary by adding and learning the recognition data. SOLUTION: A segmentation processing part 4 cuts a character image to be recognized out of a character image read by a document read part 1, a recognition processing part 5 recognizes a character according to recognition algorithm, and a candidate character selection part 6 selects the most likelihood candidate character and displays the character at a display part 7. When a writer decides that the character is a misrecognized character and corrects it into the correct character, a display/correction conversion processing part 9 generates a management table containing the misrecognized character, corrected character, and the address of the cut image of the character. A recognition data image storage memory 10 inputs and stores the management data and the cut image having the address shown from the segmentation processing part 4. A recognition processing part 5 adds and learns recognition algorithm according to the management table and cut image.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for adding a character recognition dictionary used when reading and recognizing characters and a terminal OCR device using the method.

[0002]

2. Description of the Related Art Conventionally, as shown in FIG. 5, an apparatus of this type has a signal processing section 2 for reading an original read by an original reading section 1.
And binarize it and store it in the image storage memory 3. The cut-out processing unit 4 takes in the image data from the binarized image data according to the image cut-out algorithm, separates the recognition area, and the recognition processing unit 5 performs the recognition process. The recognition processing unit 5 is equipped with several recognition processing algorithms for recognition processing, and a method for automatically selecting an optimized recognition algorithm by detecting features of data in which recognition regions are separated, or attribute information Methods and the like selected according to the above are known. In the recognition processing, generally, a method using neurons, a method for recognizing by vector distance calculation, etc. are known. Since the character recognized in this way usually consists of a plurality of candidates, the candidate character selection unit 6 selects the optimum character and outputs it as the recognition result.

The recognition processing technology is also advanced, and the recognition rate (correct answer rate) and the like can be recognized with high accuracy (probability). In particular, printed characters such as printed characters and symbols show values exceeding 99%. However, the recognized characters include not only printed characters but also handwritten characters, and it is required to recognize these handwritten characters with high accuracy. In Japanese, there are kanji, katakana, and hiragana, which make recognition processing more difficult than in Western characters. In addition, the large difference due to individual writing quality such as habits also makes recognition processing difficult.

When recognizing a handwritten character, a cutout process for separating an appropriate image cutout area from image data is performed. For the clipping method, mark information is added to the edge of the original, the specified area is printed with a drop-out color frame in the entry area, and the distribution of dots between characters is automatically histogram-less (X-axis and Y-axis). It is known that the image cutout area is separated from the image data by detecting based on (for both axes). In recent years, recognition of handwritten characters written in a free format without a frame, printed characters (ligatures, trademarks), etc. has also been performed.

[0005]

When recognizing characters, especially handwritten characters, the cause of the incorrect recognition is examined. When the characters are dirty, the written characters are thin and difficult to read.
There are common causes for each person who said the manuscript is skewed, individual causes such as habits, and typographical errors. Even if the cause is common to each person or is specific to an individual, if the degree is small, the possibility of correct recognition is high. Therefore, a conventional recognition algorithm generally recognizes a large number of handwritten characters and extracts features required for recognition based on the learning data thereof or has feature parameters multidimensionally. In order to recognize even individual habits, an extremely large number of people have to learn habits, which improves the recognition accuracy of a particular individual, but cannot expect the recognition effect for another person. For this reason, character recognition must be performed assuming the average recognition accuracy for many people, and recognition software for such average character feature extraction data or multidimensional data of feature parameters does not recognize character recognition. The current situation is that the manufacturer of the device fixes it (in other words, forms a ROM) and incorporates it into the device. On the other hand, it is a specific individual (s) who purchase and use these devices. For the user, whether it is a habit or not, if the device does not correctly recognize the character that he / she wrote, the device has no meaning. As described above, the conventional device is a device that is difficult for a user who has a habit of writing.

The present invention has been made in view of the above problems, and an object of the present invention is to make it possible to easily add character recognition data enabling recognition of characters having personal characteristics to a character recognition dictionary. To do. Another object of the present invention is to facilitate additional learning by using personal learning data that enables character recognition having personal characteristics.

[0007]

In order to achieve the above-mentioned object, in the invention of claim 1, character recognition is performed based on a recognition dictionary formed by reading characters and recognizing the characters of the read character image according to their characteristics. Then, display this recognition result, and when the displayed recognition character is corrected,
The corrected character and the character image of the character are stored in association with each other, and the characteristics of the stored character image are extracted and added to the character recognition dictionary.

According to the second aspect of the present invention, the personal learning paper on which personal handwritten characters are written is read, and the characters of the read character image are recognized based on a recognition dictionary composed of information for recognizing the characters. Display the recognition result,
When the displayed recognition character is corrected, the corrected character and the character image of the character are stored in association with each other,
The characteristics of the stored character image are extracted and added to the character recognition dictionary.

According to another aspect of the invention, there is provided a reading means for reading a character, a recognition dictionary storing means for storing a character recognition dictionary composed of information for recognizing a character of the read character image according to its characteristics, and the reading means for reading. A recognition means for recognizing a character image based on the character recognition dictionary, a display means for displaying the recognized character, a correction means for correcting when the displayed recognized character is erroneous, and the character information. Management storing means for managing and storing character images correspondingly,
The management apparatus includes a feature extraction unit that extracts features of the character image stored in the management storage unit, and a control unit that adds the features extracted by the feature extraction unit to the character recognition dictionary as character recognition information.

According to another aspect of the present invention, external storage means can be detachably connected to the management storage means, and the corrected character information and its character image are managed and stored in correspondence with each other in the external storage means. .

According to the invention of claim 5, the corrected character information for each individual and the character image thereof are stored in the external storage means together with the identification number of the individual. When the identification number is input, the corrected character information corresponding to this identification number is stored. And the character image will be referenced.

[0012]

According to the invention of claim 1, the characters are read and the characters recognized based on the character recognition dictionary are displayed. When the displayed character is not correctly recognized and is displayed as a different character, the operator corrects the character. Since the corrected character and the character image obtained by reading this character are stored in association with each other, the characteristic of this character image is extracted to create character recognition data, and this is added to the character recognition dictionary. As a result, the character can be correctly recognized thereafter, and the recognition accuracy is improved especially in the case of an individual habit character.

According to the second aspect of the present invention, the personal learning paper is written in katakana, hiragana, numbers, kanji that are frequently used, etc. by the operator of the apparatus, reads the paper, displays the recognition result, and is erroneous. In this case, the operator corrects the characters. The corrected character and the character image obtained by reading this character are stored in association with each other, the characteristics of this character image are extracted to create character recognition information, and this is added to the recognition dictionary. Thereby, the personal data can be added to the recognition dictionary in a short time.

According to the third aspect of the invention, the recognition means recognizes the read character image based on the recognition dictionary and displays it on the display means. If the displayed character is incorrect, the operator corrects the character using the correction means to correct the character, which is erroneously recognized in the management storage means, and the corrected character is associated with the character image of the erroneously recognized character. Remember. The feature extraction means extracts the features of this character image to create character recognition data,
The control means adds this data to the character recognition dictionary. This improves the character recognition accuracy.

According to the fourth aspect of the present invention, the external storage means stores personal information such as a character or a character image which has been erroneously recognized due to a habit character and is corrected, and the user can use the terminal OCR.
When the device is used, by connecting the external storage means to the device, the device can be a dedicated device that can recognize even habit information of the user. Further, the data of the external storage means enables additional learning to the connected terminal OCR device.

According to the invention of claim 5, the personal information of a plurality of persons can be stored in the external storage means. Therefore, the personal information is distinguished by the identification number. Information is available for reference.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a terminal OCR device of the present invention. The document reading unit 1 reads a document to form character image data, and the signal processing unit 2 binarizes the character image data. The image storage memory 3 stores the binarized character image data. The cutout processing unit 4 cuts out the image data from the character image data stored in the image storage memory 3 according to a cutout algorithm, and separates the recognition region of the character. The recognition processing unit 5 has a recognition processing algorithm necessary for recognizing the characters of the data in the separated recognition area, and performs the recognition processing according to the recognition processing algorithm.

Since the recognition processing unit 5 normally outputs a plurality of characters as a recognition result for one recognition target, the candidate character selection unit 6 selects an optimum character from these recognition candidate characters. The display / correction conversion processing unit 9 has a display unit 7 and a keyboard 8, displays the optimum character selected by the candidate character selection unit 6 on the display unit 7, and when this character is different from the character of the manuscript, the operator A correction character is entered from the keyboard 8 and a correction process is performed. Recognition data / image storage memory 10
Stores the corrected data, the correction history, the data obtained by separating the recognition area by the cutout processing unit 4, and the like.
The memory card 11 can be detachably connected.
Personal character recognition data is stored in the IC memory card 11, and writing quality information represented by handwritten characters of each individual,
Contains information such as habit information, inclination information, edition, and characteristics of making,
The stored data can be validated with a personal password or the like. IC memory card 11 is FLASH memory, EEP
It is composed of a ROM, a magnetic disk, or the like. The communication unit 12 transmits the recognition-processed data, character image data, and the like through a line 13 to a host computer (not shown) or the like.

The operation of the terminal OCR device configured as described above will be described below with reference to FIG. In the document reading unit 1, characters (especially handwritten characters) written on the document sheet are imaged on a reading element by a lens and the image signal is output to the signal processing unit 2. Although the CCD element is used as the reading element in this embodiment, the present invention is not limited to this, and an image pickup tube or the like may be used. After the output of the CCD element is binarized by the signal processor 2, the binarized data is temporarily stored in the image storage memory 3. Although the capacity of the image storage memory 3 is limited to the control of the mechanical system, a buffer memory of about several tens of lines may be used as long as the mechanical system can stop the document. In this embodiment, the capacity is such that one page of A4 size original can be stored. As a result, the data for one page of the image read by the document reading unit is stored in the image storage memory 3.

Next, the cut-out processing unit 4 operates the image storage memory 3
The region to be recognized is cut out from the stored data in accordance with the cutout algorithm. As this cutting method, a general method is to print a black mark on both ends in advance on a filling sheet, write a character or the like to be recognized in the black mark, and cut out from the beginning of the mark in the cutting processing unit 4. The distance is determined in advance, and the stored data is cut out according to the cutout algorithm and output to the recognition processing unit 5. In this case, the position to be recognized is predetermined, and characters and the like cannot be written in free positions. As another method, there is a method in which the information such as the above-mentioned positioning of the black mark is not printed and the character recognition area is automatically detected and cut out. With this method, the characters to be entered can be entered freely, but there is also a technical problem regarding the cutting method. For example, recognition becomes difficult when a ligature or an adjacent character comes into contact. In the latter method, generally, a histogram of a pixel distribution in the axial direction, that is, an X-axis direction histogram, a Y-axis direction histogram is used to cut out a recognition region, detect the inclination of a document, and correct the inclination.
The line spacing is identified.

The data cut out by the cutout processing unit 4 is input to the recognition processing unit 5 and the recognition processing is performed. The recognition processing unit 5 performs recognition processing on the data in the recognition area based on a predetermined algorithm, and outputs the recognition result as a candidate character to the candidate character selection unit 6. The recognition processing unit 5 performs the recognition process using, for example, a known neuron or an algorithm using inter-vector distance calculation. In this embodiment, the processing algorithm of the neuron system is used, but the present invention is not limited to this and other algorithms can be used. As the preprocessing of the recognition processing, known processing such as thinning processing, line density equalization, normalization processing (including non-linearity) is performed. In the recognition processing, each recognition area is processed by feature extraction for extracting the feature of the contour direction density + background density of the recognition area, and the corresponding normal plural candidate characters are output as the recognition result. This recognition processing method is different from a general processing algorithm such as pattern matching, and is characterized in that recognition can be performed with high accuracy on a small scale.

The recognition processing section 5 arranges candidate characters having multidimensional characteristics close to the multidimensional characteristics (for example, the characteristics are within 100) extracted as the characteristics in the order of candidate high rank, and outputs them to the candidate character selection section 6. The algorithm by the neuron used in this embodiment is a processing form that imitates a nerve, and the processing is basically executed in parallel. A learning process is performed to generate the multidimensional feature, and the feature is generated based on the learning result. Therefore, if a large number of learning processes are performed, a better recognition accuracy can be achieved. The recognition dictionary storage means, the recognition means, the feature extraction means, and the control means described in the claims represent the functions of the recognition processing unit 5.

The candidate character selection unit 6 displays / corrects / converts the characters in the candidate high order of the output candidate characters of the recognition processing unit 5.
Output to The display / correction conversion unit 9 visualizes (displays) the recognition result to the operator, and when the recognition result is different from the entered content and is erroneous recognition, the keyboard 8 integrated with the display unit 7 is used to It is possible to correct characters that were erroneously recognized.

In the recognition data / image storage memory 10, the corrected data having been corrected, the history of the correction processing and the image input to the recognition processing unit 5 of the corrected character are stored in association with each other. It corresponds to the management storage means in the claims. The image data (image) cut out by the cutout processing unit 4 is cut out for each recognition area, and is input to the recognition processing unit 5 to be recognized and the recognition data / image storage memory 1 by the line A.
In case of erroneous recognition, it is stored together with the above-mentioned correction processing history and the like. The erroneously recognized image data is sent to the recognition processing unit 5 on the line B, and the recognition processing unit 5 uses the data as data having an important personal weight for performing new learning on neurons or the like. Become. Based on this data, the neurons and the like can perform additional learning, and personal habits and the like can be recognized with high recognition accuracy (probability).

Data having such an important weight is I
When storing and holding in a medium such as a C card, selecting valid / invalid with a personal code or password, and performing recognition learning, this data is absorbed by the recognition processing unit 5 and used as learning data. It can be so. Therefore, the IC card 11 can be attached to and detached from the recognition data / image storage memory 10.

When the recognition result displayed on the display unit 7 is correct, a communication unit 12 for transmitting the recognition result to a host computer (not shown) is provided, and the communication unit 12 communicates using the communication line 13. To do. Which communication procedure is used in the present invention is arbitrary.

Next, the character recognition process will be described in detail with reference to FIG. In FIG. 2, processing after the cutout data cut out by the cutout processing unit 4 described in FIG. 1 will be described. The method described above is used for the clipping algorithm. The cutout data 14 is cut out with a certain area as shown in FIG. The square frame indicated by the broken line is added for the purpose of explanation and is not in the actual area. The cut-out data 14 thus cut out becomes a recognition-processed image 15 having X and Y regions, and is input to the recognition processing unit 5 to be subjected to recognition processing. In the present embodiment, the recognition algorithm uses neurons, but the recognition algorithm is not limited to this. The recognition processing is performed, and as a result, the candidate character 16 closest to the recognition processed image 15 is output from the recognition processing unit 5. Candidate character 1
As for 6, the recognition processing unit 5 usually derives a plurality of characters from the highest rank closest to the recognition processed image 15 to the lower rank as a recognition result, but the number of candidate characters varies depending on the recognized characters. It should be noted that the character recognition will be described by limiting it to one character for easy understanding in the present description, but in the actual processing, a plurality of characters are recognized, and the candidate character 16 also occurs for each character.

The candidate character 16 is output to the display / correction conversion unit 9 as code information. For example, the letter A is 41HEX
(Hexadecimal number display). The present invention is a terminal in which the recognition result can be confirmed / corrected by the terminal OCR device. Therefore, the terminal OCR device is equipped with the display 7 and the keyboard 8 so that the character writer can confirm the recognition result on the terminal and make a mistaken recognition. If there is, correction work can be performed on the spot and correct data can be obtained. If there is no erroneous recognition, the correction work is unnecessary and the communication unit 12 is used to transmit the data stored in the recognition data / image storage memory 10 to the host computer or the like.

Next, the display operation will be described. The result of the recognition processing is displayed / corrected by the conversion processing 9 to generate a display screen in a predetermined format. This display format depends on the application. The recognition result is also stored in the recognition data / image storage memory 10 at the same time. The storage area may use any location in the memory. If there is no error in the recognition result, the stored recognition result data is transferred using the communication line 13 or the like.

Next, a case will be described in which the recognition result includes erroneous recognition and the character writer (operator) corrects it. At present, the recognition accuracy cannot reach 100%, and erroneous recognition may occur. Therefore, the writer can correct the erroneously recognized character (the character of the recognition result displayed on the display 7) using the keyboard 8. The correction is automatically performed when the cursor is placed on the erroneously recognized character displayed on the display 7 and the correction character to be changed is input. To correct Kanji characters, it is necessary to select and correct the specified characters after using Kana-Kanji conversion, but numbers, hiragana, etc. can be specified as they are.

The data corrected in this way is used for additional learning in the recognition processing section 5. According to the present invention, the recognition processing unit 5 can additionally learn a habit character depending on an individual and can supply a personal database at the time of additional learning. In the additional learning process, features corresponding to the recognition processing algorithm such as multidimensional feature data or vector features are extracted from the recognition region data cut out similarly to the recognition process by a feature extraction algorithm. In the additional learning, it is known that in the case of a vector feature, it is easy to add a filter or the like for comparing the features, and it is easier to perform additional learning than a neuron using a multidimensional feature.

Next, the additional learning will be described. When additional learning is originally required, general additional learning such as ligatures,
Learn features that should be new features such as symbols, trademarks, and foreign languages, and learn features that should be features for each individual when adding features and for the purpose of improving current recognition accuracy. May be added. The present invention is directed to the latter additional learning that learns the characteristics of each individual and adds the characteristics. Therefore, it is necessary to efficiently collect individual learning data. In the present invention, the erroneously recognized character is determined from the information corrected by the correction process, and the cut-out data 14 of the erroneously recognized data is output from the cut-out processing unit 4 and stored in the recognition data / image storage memory 10 together with the management information. Collect by doing.

FIG. 3 shows a recognition data / image storage memory 1
The learning data management table of 0 is shown. Since the learning data can store a plurality of pieces of personal information, a personal identification code is first provided and used to identify the personal information. This personal information can be used as external reference data by the recognition processing unit 5. Note that these pieces of information are code conversion tables, and are different from the multidimensional features and the like used in the recognition processing by the recognition processing unit 5.
The recognition result code is a character representing the result of recognition by the recognition processing unit 5, but in FIG. 3 it is provisionally written as a number or katakana instead of a code for easy understanding. The correction code is obtained by correcting the character recognized by the recognition processing unit 5 when the writer judges that the character is incorrect. In FIG. 3, the writer is provided in the display / correction conversion processing unit 9 with the numeral 7 as the numeral 1. The information is corrected by the keyboard 8. The number of corrections represents the number of correction frequencies. In relation to the number of corrections, the same number of image storage address values as the number of corrections are recorded. This image storage address value is cut out by the cutout processing unit 4 in FIG.
This is the address of the cutout data stored in.

The recognition processing section 5 inputs learning data based on this management table and performs additional learning. For example, when learning is performed using neurons, the recognition processing unit 5 absorbs the image data used for recognition from the image storage address value together with the correction code (correct character code) and performs additional learning using a known learning algorithm as additional learning data. To do. The reason why the additional learning can be easily carried out in this way is that the present invention is a terminal in which the character recognition person can immediately confirm and correct the recognition result of the apparatus by using the terminal OCR apparatus. In other methods that do not use such a terminal, for example, when a neuron is adopted, learning of the neuron must be performed again for many characters in order to collect personal information, and the misrecognized character There are not many efficient ways to learn just. Note that some center OCR devices that place OCRs in each terminal, not in the center, transfer handwritten characters from the terminal's facsimile to the center personal computer and perform recognition processing on the center personal computer. There is no apparatus that can be used by the operator (in this case, not the person who entered the facsimile message) to perform additional learning using this correction information. In the case of such a center OCR device, since the operator who corrects the error is not the writer, even if the device misrecognizes it, it may not be recognized as an error. Also, if the person who entered the text is correct, it is possible to correct it correctly, but the operator of the center cannot necessarily correct the text. Therefore, the recognition accuracy is lower than that of the terminal OCR device of the present invention.

The present invention uses a method of storing a management table and personal information such as a cutout image having an address described in the management table in a portable recording medium such as an IC card. A FLASH memory is used as an element for storing data in consideration of capacity, price, etc.
It is also possible to use a large-capacity memory such as a hard disk built into a card. However, a capacity of at least about 1 MB is sufficient to store the cutout image. When the capacity is exceeded, additional learning may be executed, stored in the recognition processing unit 5 to process the data stored in the IC card, and then cleared. By additionally learning the personal information in this manner, the present invention can significantly improve the accuracy of erroneous recognition due to the habit of the individual.

In the above description, after the additional learning of the personal information, the recognition accuracy of the data of the individual is improved. However, correct recognition can be performed using the management table and the cut-out image of the address described in the management table without performing additional learning. That is, when a cut-out image having the same characteristics as the cut-out image that has been erroneously recognized is input, this is erroneously recognized as "7" in the management table of FIG. 3, for example, the recognition result code, and then the correction code "1" is recognized. If it is output as a result, it means that the correct recognition result is output. As a result, if the characters of the writer are correctly recognized, the recognition accuracy is improved. This method is effective when recognition processing is performed again before additional learning is performed.

Next, another embodiment will be described. In the above-described embodiment, personal information is collected and stored in the management table or the like every time a unit of text written by an individual (for example, one order form) is erroneously recognized, and additional learning is performed. .
However, in this case, a considerable amount of sentences must be recognized and processed in order to collect a lot of habits of the individual. In the present embodiment, such personal information is collected intensively and additional learning is performed using this.

A personal learning form is used to collect personal information. This sheet for personal learning is to be filled with numbers and letters that cover a number of individual habits such as numbers 0-9, all hiragana and katakana, and standard kanji. Each person fills in the designated character (including the number. The letters mentioned so far and the letters in the future shall include the number) in the designated position. The entered data is subjected to recognition processing, and the recognition result is displayed on the display 7. If there is an erroneously recognized character, the writer immediately performs a correction process, thereby sequentially creating management information such as a management table. If this work is carried out several times, the display / correction conversion processing unit 9 automatically corrects and displays misrecognized characters for personal handwritten information, and the writer corrects the misrecognized characters. The number of times comes to half. If there are almost no correction characters, it means that "collection of personal correction information" to be added to the "current recognition processing algorithm" has been completed. This is a stage at which correct recognition can be performed by the management table described above. Although the recognition accuracy is improved if this is left as it is, finally, by performing additional learning, it is possible to recognize a habit character or the like from the recognition processing unit 5 and output a correct candidate character. To perform this additional learning, the recognition data / image storage memory 1
The cut-out image information stored in 0 is the line B in FIG.
Is used to return to the recognition processing unit 5 again, the feature of the personal information is extracted based on the recognition region data that is erroneously recognized, and this is learned. With such additional learning, the subsequent recognition processing is performed by combining the previously learned multidimensional feature (learning by neuron) and the vector feature (learning by the inter-vector distance calculation) with the additional feature learned this time.

FIG. 4 is an operation flow chart of another embodiment.
On the personal learning sheet, the writer writes the characters to be written enclosed by the broken line at a predetermined position (S1). The recognition processing of the entered data is executed by the terminal OCR device (S2), and the display / correction conversion processing unit 9 displays the recognition result on the display 7
(S3). If this display is erroneously recognized (S4), the writer corrects the correct character using the keyboard 8. When this correction is performed, the correction conversion table (management table) shown in FIG. 3 is created, and the correction contents are stored and the correction conversion table is displayed (S7). The cut-out image (corrected image) of the corrected character is input from the cut-out unit 4, is given a storage address, and is stored in the recognition data / image storage memory 10 (S8). In this way, the creation of the management table is completed (S9) and stored in the recognition data / image storage memory 10. It should be noted that the recognition data / image storage memory 10 may be stored in an IC card.

When the management table is created in this way,
The recognition processing unit 5 performs the recognition process again with the erroneously recognized corrected image, and performs correct recognition by referring to the management table (S10). The processing of steps S4 and S7 to S10 is performed for each erroneously recognized character, and is performed until all erroneous recognitions are completed (S11). In addition, when the misrecognized character disappears in step S4, but all of the learning as described above is not completed (S5), it is filled in on another personal learning sheet (S6), and steps S2 to S4 and S7 are performed. ~ Repeat S10.

The above-mentioned state is the state in which the habit character of an individual can be correctly recognized by the management table. Although the recognition accuracy is improved if this is left as is, the correct candidate character is output directly (without using a management table or the like) from the recognition processing unit 5 by additional learning. Since correct recognition can be performed using the management table, this additional learning need not be performed immediately after the management table is created, but may be performed at any time. In addition,
The additional learning itself is a known technique. The recognition processing unit 5 sucks up the learning result such as the management table and the cut-out image addressed in the management table from the recognition data / image storage memory 10 or the IC card (S12).
Features are extracted from the extracted clipped image (S
13) reflect this feature in the recognition algorithm (S
14) and the additional learning ends.

[0042]

As is apparent from the above description, according to the present invention, when a read character is recognized and displayed, if the character is erroneously recognized, the character writer or the like corrects it. By managing and storing the correction information including the correction character and the cut-out image of the character, the recognition accuracy can be improved without using this and additional learning.
This correction information is stored in an external storage means such as an IC card,
By enabling detachable connection with the terminal OCR device, when the connection is established, the device can be specialized by the owner of the IC card. In addition, it is possible to perform additional learning by using this correction information, and in particular, it is possible to improve the recognition accuracy of an individual. By using an external storage means such as an IC card, additional learning can be performed for other terminal OCR devices. The correction information of an individual can be intensively collected by using a sheet for individual learning, and additional learning can be performed using this correction information. Also,
Even if the recognition program is implemented as software or ROM, additional learning can be easily performed. The recognition algorithm of the present invention can be applied to any recognition algorithm without being limited to the algorithm for neuron or the vector distance calculation. Further, according to the present invention, since only the cut-out image corresponding to the corrected character needs to be stored, the memory amount can be small. Although the recognized character can be transferred to the host computer, the correction information of the erroneously recognized character can also be transferred and the host computer can perform the recognition processing based on this correction information. As described above, according to the present invention, the correct recognition can be performed based on the erroneous recognition data of the individual, so that the recognition accuracy of the individual is significantly improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention.

FIG. 2 is a block diagram showing detailed processing of character recognition.

FIG. 3 is a diagram showing a management table created in a recognition data / image storage memory.

[Fig. 4] Flow chart for intensively collecting a large amount of personal information and performing additional learning.

FIG. 5: Conventional character recognition processing flow chart

[Explanation of symbols]

 1 Document Reading Section 2 Signal Processing Section 3 Image Storage Memory 4 Cutout Processing Section 5 Recognition Processing Section 6 Candidate Character Selection Section 7 Display Section, Display 8 Keyboard 9 Display / Correction Conversion Processing Section 10 Recognition Data / Image Storage Memory 11 IC Memory Card 12 communication units 13 lines

Claims (5)

[Claims] 1. A character is read based on a recognition dictionary composed of information for reading a character and a character of the read character image according to its characteristics, displaying the recognition result, and correcting the displayed recognized character. In this case, the corrected character and the character image of the character are stored in association with each other, and the characteristic of the stored character image is extracted and added to the character recognition dictionary. . 2. A personal learning paper on which personal handwritten characters are written is read, recognition is performed based on a recognition dictionary composed of information for recognizing the characters of the read character image according to their characteristics, and the recognition result is displayed. When the displayed recognition character is corrected, the corrected character and the character image of the character are stored in association with each other, and the characteristic of the stored character image is extracted and added to the character recognition dictionary. A method for adding a character recognition dictionary characterized by the above. 3. A reading means for reading a character, a recognition dictionary storing means for storing a character recognition dictionary composed of information for recognizing a character of a read character image according to its characteristics, and a character image read by the reading means. The recognition means for recognizing a character based on the character recognition dictionary, the display means for displaying the recognized character, the correction means for correcting when the displayed recognized character is incorrect, and the character information and the character image are associated with each other. Management storing means for managing and storing by means of the above, feature extracting means for extracting features of the character image stored in this managing storing means,
A terminal OCR device comprising: a control unit that adds the features extracted by the feature extraction unit as character recognition information to the character recognition dictionary. 4. An external storage means can be detachably connected to the management storage means, and the corrected character information and its character image are managed and stored in correspondence with each other in the external storage means. The terminal OCR device according to claim 3. 5. The corrected character information for each individual and the character image thereof are stored together with an identification number of the individual in the external storage means, and when the identification number is input, the corrected character information corresponding to the identification number and the character image are referred to. The terminal OCR device according to claim 4, wherein