JP2906758B2 - Character reader - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character reading apparatus for automatically reading handwritten characters and printed characters, and more particularly to a character for reading handwritten deformations, similar fonts, multi-font printed characters, and omni-font printed characters. It relates to a reading device.

[0002]

2. Description of the Related Art Conventionally, in a character reading apparatus for reading a character pattern such as a handwritten character or a printed character, individual character data is extracted in a character cutout section from character string data handwritten or printed on a form or a document. After the clipping process is performed, the individual character recognition process is performed.
The general individual character recognition process performs a feature extraction process on each unknown input character pattern cut out by the character cutout unit, and obtains the obtained feature vector and a character recognition dictionary designed in advance from a set of learning character patterns. The distance value or similarity is calculated, and the determination unit determines the character category having the minimum distance value or the maximum similarity value as the read result.

As a means of improving the reading performance in character reading, there is a method of performing character reading by providing a character recognition dictionary including a plurality of subcategories corresponding to various deformations for one category. (References:
Miyamoto. N. , Nakajima. N. an
d Kawatani. T: "High performer
mance optical character r
eaderforhand printed number
rals. alphabets. and Katakan
a ", NTT Review, Vol. 1, No. 2,
pp. 73-81 (July. 1989).

However, in the conventional character reading apparatus, since recognition processing is performed independently for each character, there is no information of other character data written on a form or a document. , When a character pattern similar to multiple character categories is input due to the writer's habit, etc.
There is a drawback that identification is difficult, determination is impossible, or erroneous recognition occurs.

FIG. 2 shows an example of characters written on a form. Although five characters are entered in the character frame, the second character is difficult to discriminate whether it is a numeral 1 or 7, and the reading is rejected because it cannot be determined by the conventional character reading device. However, when a human reads a character, he or she does not always pay attention to one character and recognizes it individually, but also uses information of characters written before and after.

In the example of the characters shown in FIG. 2, if only one character is focused on, it is difficult to determine whether the second character is 1 or 7, but the fourth character can be reliably determined to be 7. If it is assumed that the data is the data written by the same writer, the fourth character and the second character have a difference in the character shape using the fourth character determination information. Since the eye can be determined to be 7, the second character is determined to be 1. Therefore, it is difficult for a conventional character reading device that performs the determination process independently for each character to obtain a reading performance close to that of a human.

An object of the present invention is to fill in a form or a document written by the same writer by taking advantage of the fact that the variation in the character shape of the same category written by the same writer is small. It is an object of the present invention to provide a character reading device that can read a character shape that is difficult to read by the conventional individual character recognition method that performs processing for each character by reading from the entire character character shape information.

[0007]

According to a first aspect of the present invention, there is provided a character reading apparatus having a recognition dictionary for storing feature vectors of a plurality of sub-categories for one character category, and recognizing a feature vector of an input character pattern. In a character reading device that reads characters on a form or a document using individual character recognition means that performs category determination processing based on the distance value between feature vectors of each subcategory of a dictionary, the distance value between the subcategory feature vectors is small. A determination result that stores a proximity flag memory that sets and stores a proximity flag for a subcategory pair, a category name of a first candidate and a second candidate of a distance value, a subcategory feature vector number, and a determination flag indicating whether to read or reject. If the i-th character indicates rejection after the judgment process of the memory and one sheet of form is completed, the i-th sentence A comparison is made as to whether the same category as the second candidate category of the eye exists as the first candidate category of another character, and the same category exists at the j-th character on the form or document, and the first category of the i-th character Candidate, second
The proximity flag is set for the candidate subcategory pair, and
When the proximity flag is not set for the first candidate of the character and the first candidate subcategory pair of the j-th character, the first candidate of the i-th character
It is characterized by having forced character determination means for determining a candidate category as a determination result and forcibly determining a rejected character again.

The character reading apparatus of the second invention has a recognition dictionary for storing a plurality of sub-category feature vectors for one character category, and includes a feature vector of an input character pattern and a sub-category of each sub-category of the recognition dictionary. In a character reading device that reads characters on a form or a document using an individual character recognition unit that performs a category determination process based on a distance value between feature vectors, a proximity flag is set to a subcategory pair having a small distance value between the subcategory feature vectors. The same writer as a proximity flag memory that stores and stores a first candidate and a second candidate of a distance value, a category name, a subcategory feature vector number, and a determination flag indicating whether to read or reject. To pre-store the judgment frequency of each subcategory feature vector in the recognition dictionary obtained from After the determination processing of the degree distribution memory and one sheet of form is completed, when the i-th character indicates rejection, the same category as the second candidate category of the i-th character is used for other determined characters. A comparison is made as to whether the same category exists as the first candidate category. If the same category exists at the j-th character on the form or document, the sub-category feature vector number of the second candidate at the i-th character, The judgment frequency distribution memory is referred to based on the subcategory feature vector number of one candidate, and when the judgment frequency is equal to or less than the threshold value, the first candidate category of the i-th character is used as the judgment result, and the rejected character is forcibly judged again. A forced character determination unit. A character reading device according to a third invention is characterized in that, in the determination frequency distribution memory of the character reading device according to the second invention, the content of the determination frequency distribution memory is updated during the operation of reading a form or a document.

[0009]

After the judgment processing for characters on a form or document is completed, the data in the judgment result memory is referred to, and the proximity flag is set for a character whose distance value between the first candidate category and the second candidate category is close and indicates rejection. According to the content of the memory or the content of the determination frequency distribution memory, the same category as the second candidate category of the character indicating rejection exists as the first candidate category of another character on the form or document, and a plurality of different categories If the determination is made with high reliability without proximity, the character indicating rejection is forcibly determined as the first candidate category.

[0010]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The constructions of the first, second and third inventions will be described below with reference to the drawings. FIG. 1 is a configuration diagram showing one embodiment of the first, second, and third inventions. The scanner unit 1 performs a binarization process on a form or document image data optically scanned to generate a black and white binary level character string pattern. The character cutout unit 2 cuts out a character string pattern based on the size and pitch information of the character string pattern, and performs individual character cutout processing. The feature extracting unit 3 extracts character features such as the shading feature and the outline feature of the character, and obtains an N-dimensional feature vector f = (f ₁ ,.
.., F _N ). In the recognition dictionary unit 4,
M types of character categories C _{1 to be} recognized (I = 1,.
., M), the sub-category C _IJ (I = 1,...) Divided into L sub-categories
, M, J = 1,..., L) sub-category feature vector g _IJ obtained by calculating a set of feature vectors
= (G _{IJ 1} ,..., G _IJN ).

Next, a distance calculator 5 calculates the distance between the feature vector of the input character pattern and each sub-category feature vector based on equation (1).

D ² (f, g _IJ ) = (f−g _IJ ) ^t (f−g _IJ ) (1) The determination unit 6 rearranges the distance values obtained by the distance calculation unit 5 in ascending order. And writes the obtained determination result in the determination result memory 8. As shown in FIG. 3, the determination result data stored in the determination result memory 8 is obtained by the determination unit 6 for each character one by one in order for n characters written on a form or a document. Write the result. The result of the k-th character is a judgment flag h ^(k) indicating judgment or rejection, a first candidate category name C _{I 1} ^(k) having the first (minimum) distance value, and a sub-category feature vector number S _{I. 1 J 1} ^(k) , distance value is second
Second candidate category name of position (different from the first position ^category) C _{I 2} ^(k), consisting of the sub-category feature vector number ^{_{S I 2 J 2 (k)}} .

As shown in equations (2) and (3), the determination flag is different when the difference between the distance value of the first candidate and the distance value of the second candidate of the k-th character is smaller than the threshold value ε. Since the category and the distance value are close to each other, the judgment flag is set (judgment flag h _k = 1). When the value is larger than the threshold value ε, the judgment flag is not set (judgment flag h _k = 0). In a conventional character reading apparatus, a character for which a determination flag is set is rejected (impossible to read).

When D ² (f, _{gI 2 J 2} ^(k) ) − D ² (f, g _{I 1 J 1} ^(k) ) ≦ ε, h _k = 1 (2) D ² (f, _{g I 2 J)} It performed based on the _{^{2 (k)) -D 2 (}} f, g I 1 J 1 (k))> ε h k = 0 (3) the equation (4 calculation of distance between the sub-category feature vector time), the formula As shown in (5), when the distance between dictionaries is smaller than a preset threshold value δ, a proximity flag is set (R _{IJI 'J'} = 1), and as shown in equation (6), When the value is equal to or larger than the threshold value δ, 0 is set in the proximity flag.
(R _{IJI 'J'} = 0) is written to the proximity flag memory 7.

D ² (g _IJ , g _{I 'J'} ) = (g _IJ -g _{I 'J'} ) ^t (g _IJ -g _{I 'J'} ) (4) D ² (g _IJ , g _{I 'J) '} ) <Δ (I ≠ I') R _{IJI 'J'} = 1 (5) When (I = I ') R _{IJI' J '} = 0 (6) D ² (g _IJ , g _{I' J '} ) ≧ δ R _{IJI' J '} = 0 (7) The contents of the proximity flag memory 7 are shown in FIG. In FIG. 4, the sub-category feature vector numbers 21 and 22 are used as addresses, and the proximity flag value R _{IJI 'J'} shown as 23 is used as data. In the case of the same category such as _SIJ and _{SIJ '} , the value of the proximity flag is 0. For different categories such as _SIJ and _{SI'J '} (I ≠ I '), the expression (5) , (6) and (7), the contents of the proximity flag R _{IJI 'J'} shown in 23 are determined and stored in the proximity flag memory 7.

Next, the contents of the judgment frequency distribution memory 10 are shown in FIG. Each subcategory feature vector number 24, 2
5 is the address, and the determination frequency P shown in 26
_{IJI 'J'} is a matrix for data. The judgment frequency distribution memory 10 and the threshold value register 11 are controlled by the judgment frequency distribution control unit 9, and control the contents of the memory based on the sub-category feature vector numbers obtained by the judgment unit 6.

In the character reading device of the second invention, an algorithm for obtaining a determination frequency distribution from a learning pattern previously written by the same writer is as follows. When the subcategory feature vector number in which the learning character pattern written by the same writer is determined is S _IJ, and the counter of the number of characters determined by S _IJ is q _IJ , step. 1
Each component of the judgment frequency P _{IJI 'J'} is initialized.

P _{IJI 'J'} = 0 (I = 1,..., M, J = 1,..., L, I ′ = 1,..., M, J ′ = 1 _,. , L) (8) step. 2 Initialize each component of the counter.

Q _IJ = 0, r _IJ = 0 (I = 1,..., M, J = 1,..., L) (9) step. 3. The distance between the feature vector and each subcategory feature vector is calculated based on equation (1) for all the characters in the set of learning character patterns written by the same writer, and the subcategory feature vector number S that minimizes the distance value is calculated. _{When IJ} is set, add the value of the counter.

Q _IJ = q _IJ +1 (10) step. 4 For all learning character patterns, step.
After executing step 3, the ratio between the frequency of occurrence of the category and the frequency of occurrence of the sub-category is subjected to threshold processing to obtain r
Update the value of _IJ .

[0021]

(Equation 1)

The learning characters entered by one scribe are written in step. 2 to step. Execute step 4. 5 When r _IJ = 1 and r _{I 'J'} = 1 (I = 1,..., M, J = 1,..., L, I ′ = 1,... M, J ′ = 1 ,..., L) P _{IJI 'J'} = 1 (12) step. 6 In the learning character database in which the scribe is changed, step. 2 to step. Step 5 is executed.
When the number of writers is α, the threshold θ set in the threshold register 11 is obtained by Expression (13).

Θ = f (α) f: monotonically increasing function (13) In the character reading device of the third invention, an algorithm for obtaining a determination frequency distribution from an unknown input character pattern during an operation of reading a form or document is as follows. Become like When the subcategory feature vector number for which an unknown input character pattern is determined is S _IJ and the counter of the number of characters determined by S _IJ is q _IJ , step. 1 Judgment frequency P
_Initialize each component of _{IJI 'J'} .

P _{IJI 'J'} = 0 (I = 1,..., M, J = 1,..., L, I ′ = 1,..., M, J ′ = 1 _,. , L) (14) step. 2 Initialize each component of the counter.

Q _IJ = 0, r _IJ = 0 (I = 1,..., M, J = 1,..., L) (15) step. 3. The distance between the feature vector and each subcategory feature vector is calculated based on equation (1) for the characters entered in the input form or document, and the subcategory feature vector number that minimizes the distance value is S _IJ
Then, add the value of the counter.

Q _IJ = q _IJ +1 (16) step. 4 For all learning character patterns, step.
After executing step 3, the ratio between the frequency of occurrence of the category and the frequency of occurrence of the sub-category is subjected to threshold processing to obtain r
Update the value of _IJ .

[0027]

(Equation 2)

For one input form or document, step. 2 to step. Execute Step 4. step. 5 When r _IJ = 1 and r _{I 'J'} = 1 (I = 1,..., M, J = 1,..., L, I ′ = 1,... M, J ′ = 1 ,..., L) P _{IJI 'J'} = 1 (18) step. 6 The threshold value θ to be set in the threshold value register 11 is obtained by Expression (19).

Θ = θ ₁ θ ₁ : constant (19) The character reading apparatus of the present invention has the comprehensive judgment unit 12, and stores characters which have been subjected to rejection processing in the judgment result memory 8 in the past. It is possible to read by performing a rescue process using the determined result information of the completed form or the entire document.

The processing of the overall judgment section in the character reading apparatus of the first invention will be described with reference to the flowcharts of FIGS.
Performs recognition processing for all characters on a form or document,
At the stage where the determination result is stored in the determination result memory 8, the determination flag h ^(k) of the k-th character is checked in process 28, and the determination flag is set (h ^(k) =
Since 1) indicates rejection, it is subject to this processing.
In process 32, the first candidate category C _{I 1} ^(k) ,
The second candidate category C _{I 2} ^(k) a check is _{R I 1 J 1 I 2 J} 2 is a proximity flag. If the proximity flag is not set (R _{I 1 J 1 I 2 J 2} = 0), the processing shifts to the next character, and the proximity flag is set (R
_{If I 1 J 1 I 2 J 2} = 1), the search processing data of another character is searched, and in a process 36, the category equal to the second candidate category C _{I 2} ^(k) is the first candidate category of another character. Check if it exists as. In the process 38, when the same category exists and the proximity flag of the corresponding character is not set (R
_{I 1 'J 1' I 2 'J 2'} = 0), the process 41 determines the category C _{I 1} ^(k) , and the process 40 clears the determination flag h ^(k) . In cases other than the above, when the determination flag is not set, the category is determined to be C _{I 1} ^(k) in the processing 41, and when it is set, the rejection processing is performed in the processing 46.

Next, the processing of the overall judgment section in the character reading apparatus according to the second and third aspects of the present invention will be described with reference to the flowcharts of FIGS. Recognition processing is performed for all characters on the form or document, and at the stage where the determination result is stored in the determination result memory 8, the determination flag h ^(k) for the k-th character is checked in processing 49, and the determination flag is determined. When standing (h ^(k) = 1), it indicates rejection, and this processing is symmetric. The judgment processing data of another character in the judgment result memory 8 is searched, and in a process 55, it is checked whether a category equal to the second candidate category C _{I 2} ^(k) exists as the first candidate category of another character. If the same category exists, the judgment frequency distribution memory 10 is referred to based on the subcategory feature vector number of the first candidate being determined and the subcategory feature vector number of the first candidate of the character indicating rejection, and processing is performed. At 59, the judgment frequency P _{I 2 J 2 I 1 'J 1'} is compared with the threshold value θ which is the content of the threshold value register 11, and PI _{2 J 2 I 1 'J 1'} ≦ θ (20) At this time, in step 61, the category is determined to be C _{I 1} ^(k) , and in step 60, the determination flag h ^(k) is cleared. In cases other than the above, when the judgment flag is not set, judgment is made in the category C _{I 1} ^(k) in step 61, and when the judgment flag is set, rejection processing is performed in step 66.

In FIG. 12, the reading result of the character reading apparatus of the first invention will be described while comparing it with the reading result of the prior art. The character shape of the second character 68 in FIG. 12 is ambiguous whether it is category 1 or category 7, and the rejection process is performed in the conventional character reading device, and the rejection result in the related art is rejection 72. However, with the character reading device of the present invention, it is possible to read based on the character information of a form or the entire document. Since the fourth character 70 has a character shape that can be determined as Category 7, the character of the second character 68 is determined again using the character information of the fourth character 70. As a determination method, when the determination results of all the characters are stored in the determination result memory 8, the fourth
The determination result of the character is rejection 72, and the first candidate category in the process 29 is 1 and the second candidate category is 7. The proximity flag in process 31 is category 1
Since the sub-category feature vector vectors of and are close to each other in the character shape 82 of category 1 and the character shape 85 of category 7, the first candidate proximity flag and the second candidate proximity flag become 1. In the processing 36, a comparison is made in the process 36 as to whether or not the same category as the second candidate category name 7 shown in 73 in the determination result of the other characters on the form or document exists as the first candidate category. In character 70, the same category exists as the first candidate category. In the processing 39, regarding the subcategory feature vector of the character shape 84, since there is no other category that is particularly close, the first candidate proximity flag and the second candidate proximity flag become 0. Therefore, it is possible to forcibly determine that the category is 1 as the reading result in the process 41.

Referring to FIG. 14, the reading result of the character reading apparatus according to the second and third aspects of the present invention will be described while comparing with the reading result of the prior art. The character shape of the second character 76 in FIG. 14 is ambiguous whether it is category 1 or category 7, and rejection processing is performed in the conventional character reading device, and the reading result is
It will be rejected 80. However, the character reading apparatus according to the second aspect of the present invention can read the form or the entire document based on the character information. In the fourth character 78, there is a character shape that can be clearly determined to be category 7,
The character of the second character 76 is determined again using the character information of the fourth character 78. As a determination method, at the stage when the determination results of all characters are stored in the determination result memory 8, the determination result of the second character 76 is rejection 80, but the first candidate category in the process 50 is 1 The second candidate category is 7, and in the process 55, a comparison is made as to whether or not the same category as the second candidate category indicated by 81 in the determination result of the other characters on the form or document exists as the first category. , 82, which is the same category in the determination result of the fourth character, exists as the first candidate. In this case the first candidate sub category feature vector number S _{1 2} and address subcategories feature vector number S _{7 2} of the fourth character 78 of the second character 76, the number of determination frequency indicating the determination frequency distribution memory to the processing 57 P
_{1 2 7 1} is loaded and compared with the set threshold value θ.

FIG. 15 shows the contents P _{IJ 71} (I = 1, J = 1) of the judgment frequency distribution memory based on the sub-category feature vector number _S71 having a character shape like 84.
1,..., L, L = 4). FIG.
The scribe who fills in the character shape of category 7 as character shape 84 in FIG.
5. The frequency of writing in different character shapes such as character shape 86 is extremely low. This is a property that when a scribe writes a character, the variance of the character shape for the same category is small and a similar character shape is entered.
Take advantage of this property. In process 59, the determination frequency P
_Since P ₇ 27 ₁ ≤θ (21) holds for 7271, the determination flag is cleared in processing 60 and the determination result is forcibly determined to be C ₁ ⁽²⁾ and category 1 in processing 61.

[0035]

As described above, according to the present invention, distortion caused by a writer's habit is absorbed based on the entire information of character data written before and after on a form. By doing so, it is possible to read characters similar to a plurality of categories. In addition, even when a single character, such as a multi-font character of a print character, has the same character shape as a different category of another font, and it is necessary to judge from the entire character character shape, such as when the character cannot be read, the character reading of the present invention is performed. The device can read. Although the present invention has been described with reference to characters, the present invention can be easily realized with respect to images, sounds, and graphics. Also, the explanation was made using the Euclidean distance as a measure indicating the closeness between the feature vector and the recognition dictionary, but other distances (Maharanobis distance, city block distance, etc.) and similarities (simple similarity, compound similarity, etc.) ) Is also applicable.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining an embodiment of a character reading device according to first, second and third aspects of the present invention;

FIG. 2 is a diagram for explaining a reading result of a conventional character reading device.

FIG. 3 is a diagram for explaining the contents of a determination result memory of the character reading device according to the first, second, and third inventions.

FIG. 4 is a diagram for explaining the contents of a proximity flag memory of the character reading device according to the first, second, and third inventions.

FIG. 5 is a diagram for explaining the contents of a judgment frequency distribution memory of the character reading device according to the second and third inventions.

FIG. 6 is a part of a flow for explaining a process of a comprehensive judgment unit of the character reading device of the first invention.

FIG. 7 is a part of a flow for describing a process of a comprehensive judgment unit of the character reading device of the first invention.

FIG. 8 is a part of a flow for describing a process of a comprehensive judgment unit of the character reading device of the first invention.

FIG. 9 is a part of a flow for explaining a process of a comprehensive judgment unit of the character reading device of the second invention.

FIG. 10 is a part of a flow for explaining a process of a comprehensive judgment unit of the character reading device of the second invention.

FIG. 11 is a part of a flow for explaining a process of a comprehensive judgment unit of the character reading device of the second invention.

FIG. 12 is a diagram for comparing and explaining a character reading result by the character reading device of the first invention and a character reading result by the conventional technique.

FIG. 13 is a view for explaining examples of character shapes of category 1 and category 7 which are read by the character reading apparatuses of the first, second and third inventions.

FIG. 14 is a diagram for comparing and explaining a character reading result by the character reading device according to the second and third inventions and a character reading result by the conventional technique.

FIG. 15 is a diagram for explaining states of distributions stored in a judgment frequency distribution memory according to the second and third inventions.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Scanner part 2 Character extraction part 3 Feature extraction part 4 Recognition dictionary part 5 Distance calculation part 6 Judgment part 7 Proximity flag memory 8 Judgment result memory 9 Judgment frequency distribution control part 10 Judgment frequency distribution memory 11 Threshold register 12 Comprehensive judgment part 13 Judgment result

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G06K 9/03 G06K 9/62 G06K 9/68 G06K 9/72

Claims (2)

(57) [Claims] 1. A recognition dictionary for storing feature vectors of a plurality of subcategories for one character category, based on a distance value between a feature vector of an input character pattern and a feature vector of each subcategory of the recognition dictionary. In a character reading device that reads characters on a form or a document by using an individual character recognizing unit that performs category determination processing, a feature vector of an input character pattern and each sub-card of a recognition dictionary are used.
Distance value between categorical feature vectors is 1st and 2nd
Sai first candidate, category name of the second candidate, the first candidate distance values, category name of the second candidate, and the determination result memory for storing a determination flag indicating whether to reject or subcategory feature vector number, and read, If the determination flag is i-th
When the character indicates rejection, the first and
2 If the proximity flag is not set for the candidate category,
The i-th character is determined to be unreadable and the process proceeds to the next character.
When the proximity flag is on, the i-th character
The same category as the second candidate category has already been judged
Compare whether the character exists as the first candidate category
Determining the frequency for storing in advance determined frequency of each sub-category feature vectors in the sub-category, wherein a proximity flag memory for storing sets a proximity flag on smaller sub-category pair of distance values between the vectors, the recognition dictionary determined from the same writer to After the determination processing of the distribution memory and one sheet has been completed, if the i-th character indicates rejection, the i-th character
A comparison is made as to whether the same category as the second candidate category of the character exists as the first candidate category of another determined character. When the same category exists as the jth character on the form or document, the i-th character The judgment frequency distribution memory is referred to based on the subcategory feature vector number of the second candidate of the eye and the subcategory feature vector number of the first candidate of the jth character. A character reading device comprising: forced character determination means for forcibly determining a rejected character again using a first candidate category of an eye as a determination result. 2. A determination frequency distribution memory, a character reader according to claim 1, wherein updating the contents of the determination frequency distribution memory during a read operation of the form or character.