JPS6160188A - Character recognizer - Google Patents

Abstract

PURPOSE:To shorten the recognizing time of each character and also to improve the recognizing accuracy by sorting the types of characters in response to plural different recognizing means and using a recognizing means accordant with the corresponding sorted type of character. CONSTITUTION:The characters read by a reading part 2 are sent to a primary feature extracting circuit 15 in case the printing types are sorted by a character type reference memory 12. When the types of characters are obtained from an input pattern, the character data is normalized by a normalizing circuit 13 and sent to a character type deciding circuit 14. Then a decision parameter is calculated and sent to the circuit 15. For a printing type pattern and the numerical items, the range is limited for a sorting dictionary memory of a primary collation circuit 16 and the candidates are selected and sent to a primary deciding circuit 17. For other types of characters, the features of the corresponding characters are read out of the dictionary memory of the circuit 16 according to the decision parameter. Then the types of characters are decided and the candidates are selected and sent to the circuit 17.

Description

[Detailed Description of the Invention] [Field of Industrial Application] ■ The present invention relates to a character recognition device (hereinafter referred to as OCR) that reads and recognizes handwritten characters including kanji written on forms, etc. The present invention relates to a character recognition device that is capable of processing and reducing erroneous recognition.

In recent years, the progress of OCR has been remarkable, and OCR that can read printed letters and handwritten characters that target alphanumeric characters and kana characters is widely used in form processing operations, etc. The development of Japanese character recognition technology including the following is active, and various methods are being tried.

In the recognition of such handwritten characters including kanji, hiragana, etc., a method is desired that has a faster processing speed and fewer recognition errors.

[Conventional technology]

FIG. 3 shows a block diagram of OCR for Japanese characters, which targets handwritten characters including Chinese characters.

In the figure, form 1 is a form in which the customer's address, name, product name, etc. are written in each field.

The reading unit 2 has a function of scanning the reflected light of the light irradiated onto the form 1 through a lens system 2a with an image sensor 2b, reading one frame of characters, and transmitting the characters as image data to the binarization circuit 3. have

The main control section 4 has a function of controlling each section and executing a character reading and recognition processing program.

The image memory 5 contains binarized image data.

That is, it stores image data of read characters.

The character cutting circuit 6 has a function of cutting out one character from one frame of characters stored in the image memory 5 and sending it to the recognition circuit 10 based on the format information sent from the format information memory 9.

The feature extraction circuit 7 has a function of extracting character features sent from the recognition circuit 10, ie, character stroke number 1 curve coefficients, etc., and sending them to the recognition circuit 10.

The dictionary memory 8 stores characteristics of characters that serve as standards for recognition.

That is, the characteristics of kanji, hiragana, hiragana, alphabetic characters, numbers, symbols, etc. are stored and have a function of sequentially sending them to the recognition circuit 10 upon request from the recognition circuit 10.

The format information memory 9 stores information indicating the position where characters are written on the form 1, and the position where the read characters are written is stored in the image memory 5, character cutting circuit 6. It also has the function of sending information to the weaving circuit 10.

The recognition circuit 10 receives from the feature extraction circuit 7 the features of the character sent from the single character extraction circuit 6, compares them with the features of the characters sequentially sent from the dictionary memory 8, and calculates the degree of matching. It has a function to send out character codes as candidate strings in descending order of degree.

The post-processing unit 11 has a function of outputting the input candidate sequence in a sieve or a number according to a predetermined degree of matching.

Since it has such a configuration and function, the character recognition method will be explained. First, the characters on the form 1 are read and binarized image data is stored in the image memory 5.

Next, the image data is sent to the single character cutting circuit 6, and based on the character position information sent from the formant information memory 9,
One character is cut out and sent to the recognition circuit 10.

The recognition circuit 10 sends the input character data to the feature extraction circuit 7, which extracts the features of the character data. Therefore, the characteristics of the characters are sequentially read out from the dictionary memory 8 and compared with the characteristics of the character data, and the characters with a high degree of − quality are selected as candidate characters as answers for recognition.

If there are multiple candidate characters, character codes are sequentially output as a candidate string.

The output candidate string of character codes is sieved by the post-processing section 11, output, and stored in a storage means such as a floppy disk.

Character recognition is sequentially performed on the image data stored in the image memory 5 in this manner.

In addition to the method using feature extraction in the above example, there is a method devised for printed characters that is similar to a pattern matching method that uses correlation.

There are also methods that combine these two methods and use the recognition results of both methods, or there are methods that select the two methods using a simple method.

[Problem that the invention seeks to solve]

The above conventional method has the following problems.

■Feature extraction methods can express even small differences as features, and are suitable for distinguishing between categories with only slight differences, but for kanji with a large number of strokes, too many features are extracted, Identification is greatly affected by fluctuations in both sides, dirt on the paper surface, and fluctuations caused by writing instruments.

■The method using correlation is strong against minute fluctuations in characters, but
It is unsuitable for distinguishing between categories with only slight differences.

(2) In addition, both methods that use the methods (2) and (2) above together or selectively are not effective because the amount of processing increases and the number of parts that involve both methods increases.

Therefore, even in the above-described method, many similar characters are likely to be erroneously recognized.

The characters used in Japanese include kanji, hiragana, katakana, alpha, numerals, Arabic numerals, symbols, etc., so reading them with a single recognition algorithm has different reading accuracy depending on the type of character. Otherwise, the processing itself becomes wasteful.

[Failure to solve the problem]

The present invention provides a plurality of recognition means for recognizing character image data using different methods, a classification means for classifying the character types of the image data into correspondence with the recognition means according to the complexity and curve coefficient of the character image data, and This is a character recognition device that selects and recognizes image data based on the classification of the means, by selecting the most suitable recognition means for the classified character type, and thereby the above-mentioned problems can be solved.

The character types in the format definition mentioned here include kanji, kana characters, numbers, alphabets, and symbols, as well as objects that can be encoded, such as circuit symbols. There are roughly four types of character types for selecting recognition methods.
Although this overlaps with the character types in the format definition in some places, things like ``Kanji'' and ``Japanese'' include multiple character types for classification. The four main types are rounded groups of letters centered around ``hiragana''.

These include character groups such as kanji and symbols with little rain, such as ``1 katakana'' and ``dai-ko-gomon'', groups of kanji with each stroke (the above are mainly large-sized handwritten characters), and printed characters. Hirakana and Katakana, which are character types based on formant definitions, can correspond to classification character types, but when dealing with Japanese texts, it is absolutely necessary to determine this classification character type. .

Furthermore, it is possible to cope with a case where a text part and a part such as a drawing, table, etc. exist in a document, and it is necessary to recognize both of them.

[Effect]

According to the present invention, the classification means uses, for example, the complexity of character image data, that is, the number of strokes, the curve coefficient, etc., as parameters for character type determination, and classifies the character types so that they correspond to a plurality of different recognition means. By performing recognition using recognition means suitable for the classified character type, the processing time for recognizing individual characters can be shortened, high-speed processing can be performed, and recognition accuracy can be increased and misrecognition can be reduced.

〔Example〕

Hereinafter, one embodiment of the present invention will be described with reference to FIGS. 1 to 3. FIG. 1 is a block diagram showing an embodiment according to the present invention, and FIG. 2 is a flowchart of FIG. 1. The same reference numerals indicate the same objects throughout the figures.

In FIG. 1, the main control section 4a has the function of controlling each section to read characters, classifying character types according to the recognition means, and executing a character recognition processing program corresponding to the recognition means.

The single character cutting circuit 6a cuts out one character from the image memory 5 based on the field and line information when format information is stored in the format information memory 9, and when format information is not stored, or If only a portion of the characters are stored, when the reading unit 2 reads the characters, the character area extracting circuit 9a+line extracting circuit 9b refers to the character area, line, and field read earlier.

and a field extraction circuit 9c has a function of processing and cutting out one character.

The character type reference memory 12 is stored in advance in the format information memory 9.
Holds the character type parameter when specified by .

The normalization circuit 13 has a function of normalizing the character pattern sent from the single character cutout path 6a to a size of nXm (for example, 48 x 4B dots).

The character type determination circuit 14 has a function of counting the nxm black number or curve coefficient of the character pattern normalized by the normalization circuit 13 to calculate determination parameters and sending them to the primary feature extraction circuit 15.

-The next feature extraction circuit 15 selects a character type based on the determination parameter sent from the character type determination circuit 14, -The next matching circuit 16 reads and matches the features of the corresponding character, and selects a candidate character. It has a function of sending candidate sequences to the primary determination circuit 17. That is, candidates are selected by extracting features according to the classified character type, such as rounded characters, angular characters, stroke characters, or printed characters. For typefaces (printed characters), correction is performed to match input cover patterns (or normalized patterns), and candidates are selected by comparison with standard patterns for classification.

The primary matching circuit 16 has a built-in classification dictionary memory (not shown) which is a storage means for storing classification characteristics of characters.

The primary judgment circuit 17 examines the candidate contents of the candidate sequence sent from the feature extraction circuit 15, the significant difference in similarity between the candidates, etc.
Something that can be clearly distinguished, something that has the same shape.

It has a function to classify and output those that require secondary judgment.

That is, the range shown by the one-dot chain line in the figure has a primary recognition function.

The secondary feature extraction circuit 18 has a function of reading character features corresponding to the candidate string sent from the -order determination circuit 17 and sending them to the secondary matching circuit 19.

The secondary matching circuit 19 includes a dictionary memory and has a function of reading out and matching the features of the candidate string sent from the secondary feature extracting circuit 18 and the features of the corresponding characters in the dictionary memory.

The secondary determination circuit 20 examines the candidate contents of the candidate string sent from the secondary matching circuit 19, significant differences in similarity between candidates, etc.
It has the function of outputting clearly distinguishable items as answers, similar items as candidate sequences, and classifying items that are difficult to identify and outputting them as reject candidates.

That is, the range indicated by the two-dot chain line in the figure has a secondary recognition function.

Since it has such a configuration and function, the recognition processing method will be explained with reference to the flowchart in FIG. 2.

(1) First, when a character read by the reading section 2 is stored in the image memory 5, one character is cut out by the one character cutting circuit 6a.

■If there is a type classification in the character type reference circuit 12.

In other words, if the character type can be specified in advance and 1, for example, a printed document is to be read, the character type information is stored in the format information memory 79, and the extracted character data is normalized and character type determined. First, it is sent to the -order feature extraction circuit 15. Furthermore, if the information is a numerical item such as a telephone number or date of birth, it is similarly immediately sent to the primary feature extraction circuit 15.

■Also, if there is no character type information in the format information memory 9 (if the character type is obtained from the input cover pattern, that is, for documents containing a mixture of general handwritten characters and printed characters, the extracted character data is processed in the normalization circuit 13). It is normalized and sent to the character type determination circuit 14.

(2) For character data, determination parameters are calculated in the character type determination circuit 14 and sent to the primary feature extraction circuit 15.

■In the case of a print pattern or a numeric item, the range of the classification dictionary memory of the -next matching circuit 16 is limited, and matching with the read print pattern or handwritten math pattern is performed to select a candidate and create a candidate string. It is sent to the primary judgment circuit 17. In the case of other character types, the primary matching circuit 16 reads out the characteristics of the corresponding character from the classification dictionary memory based on the determination parameter, performs character type determination, selects a candidate, and sends the candidate string to the primary determination circuit 17. Sent. ′ □ −
In the C-order determination circuit 17, candidates are determined based on the content of the candidates and the degree of similarity between the candidates, and those that can be clearly determined are output as answers, and those that are isomorphic 3, such as ``i'' between kanji and kana, are output as answers.
The work, force 1 mouth, etc. are sent to the post-processing section with a □ lag attached and the drawing method omitted as a complement to the bowl.
゛In addition, those requiring secondary judgment are classified and classified.■Thus, the secondary feature extraction circuit 18 extracts the features, and the secondary matching circuit 19 compares them with the features in the dictionary memory to perform the secondary judgment on the candidate string. to circuit 20.

(2) The secondary determination circuit 20 outputs those that can be clearly determined as answers, and when some candidates are selected, sends the candidate string to a post-processing section (not shown). In addition, if the degree of matching is low and the selection of candidates cannot be narrowed down, they are output as a list of rejected candidates.

In this way, it is possible to specify the middle type in advance.

In both cases, the selection of candidates can be simplified, and especially when prior information is available, this can be greatly simplified and high-speed processing can be achieved. Furthermore, by performing secondary determination according to the input characters, highly accurate recognition processing with fewer misrecognitions can be performed. Furthermore, it is possible to reliably identify characters that can be distinguished based on their shape, and to output similar character pairs that cannot be identified with a flag attached, thereby allowing effective post-processing.

The above recognition method □ can be applied to objects that can be coded. For example, if there is a circuit diagram or the like in a document, recognition becomes possible by coding the circuit symbol or the like.

〔Effect of the invention〕

As explained above, according to the present invention, (1) recognition processing can be significantly speeded up even when multiple character types are mixed.

■It is possible to improve recognition accuracy and reduce misrecognition.

■Can be applied to objects that can be encoded.

There is an effect.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment according to the present invention, FIG. 2 is a flowchart of FIG. 1, and FIG. 3 is a block diagram showing a conventional method. In the figure, 4.4a is the main control unit, 5 is the image memory, 6.6a is the single character extraction circuit, 7.15 is the feature extraction circuit, 8 is the dictionary memory, 9 is the format information memory, and 10 is the recognition circuit. , 11 is a post-processing unit, 12 is a character type reference memory, 13 is a normalization circuit, 14 is a character type determination circuit, 16 is a primary collation circuit, 17 is a primary determination circuit,
18 is a secondary feature extraction circuit, 19 is a secondary matching circuit, 2
0 indicates a secondary judgment circuit. Tea 2 Diagram P: I Candy axis t1 Fighting letter t J + Rhino 7su H,, tllri 1 sentence ya η pe ro rn Kariyamashi ■ (yrs, T oh 78°O ■(E,,. ■ )■ H. ■( ″″１compliant3′′′ ・Yes)■ ■(Dining] Fig. 3

Claims (3)

[Claims] (1) A character recognition device that irradiates light onto characters written on a medium and recognizes the image data of the characters obtained from the reflected light of the irradiated light, which recognizes the image data of the characters by different methods. A plurality of recognition means for recognizing, a classification means for classifying character types of image data according to the recognition means, and a recognition means for recognizing the image data by the recognition means corresponding to the character types based on the classification of the classification means. Characteristic character recognition device. (2) The character recognition device according to claim 1, wherein the classification means classifies the character types of the image data based on preset character type information. (3) The character recognition device according to claim 1, wherein the classification means extracts characteristics of characters in the image data to classify character types.