JPH06318270A - Method and device for reading out chracter - Google Patents

Abstract

PURPOSE:To execute character recognizing processing with high identifying capacity for similar characters by utilizing preceding and succeeding character shape information. CONSTITUTION:A projection vector obtained at the time of executing distance calculation between a feature vector obtained by a feacture extracting part 3 and a recognition dictionary vector stored in a recognition dictionary memory 4 by a distance calculating part 5 is stored in a projection vector register 7. An evaluation function selecting table memory 8 finds out an evaluation function corresponding to the area of the projection vector, and when a judging part 11 judges that a judgement condition is satisfied, stores an evaluation function number 14 in an evaluation function selecting register 9. An evalution function table memory 10 outputs an evaluation value 16 based upon a specified evaluation function and the judging part 11 determines a category based upon the value 16 and a distance value found out by the calculating part 5. Thereby an evaluation function can be changed in accordance with the areas of projection vectors for the preceding and succeeding character shapes and even a character which can not be judged can be judged by changing its evaluation function and recognizing it again.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character reading method and a device for automatically reading handwritten characters and printed characters, and more particularly to a character reading method and a device for reading characters of a similar typeface having handwriting variations.

[0002]

2. Description of the Related Art Conventionally, in a character reading method and device for reading a character pattern such as a handwritten character or a printed character, character string data written on a form is cut out into individual characters by a character cutting section. That is, the feature extraction unit performs a feature extraction process on the character pattern. The feature extraction process absorbs the deformation of the character pattern and converts it into a feature vector in which the number of dimensions of the character pattern is reduced. In general individual character recognition processing, the feature extraction processing is performed on each unknown input character pattern cut out by the character cutout unit, and the distance between the obtained feature vector and the character recognition dictionary vector designed in advance from the learned character pattern set. The determination is performed by calculating the value and using the character category having the minimum distance value as the reading result.

In the character recognition dictionary, it is desirable to represent a set of feature vectors of learning character patterns by a representative vector of a low dimension as much as possible. Design of the character recognition dictionary using the principal component analysis method as a statistical data compression method. There is a technique.
The principal component analysis method is effective as a method for compressing a data set having a correlation, and is obtained by solving the eigenvalue problem of the covariance matrix of the data set and selecting the corresponding eigenvectors in descending order of eigenvalues. When applied to the character recognition dictionary design, paying attention to the fact that the transformation of character patterns has correlation for each category, using the correlation between the feature vectors of the learning character pattern set, and using the principal component analysis method. Compress. The recognition dictionary vector is composed of an average vector composed of average values of respective components of the feature vector of the learning character pattern set, and a principal component axis vector which is an eigenvector obtained by the principal component analysis of the feature vector. (References:
"Projection distance method in handwriting recognition", Ikeda, Tanaka,
Motooka, Journal of Information Processing Society of Japan, Vol. 24, NO. 1,
Jan. (1983)

However, in this conventional character reading method, after cutting out the individual characters,
Since each character is individually and individually recognized, there is no information on the character data written before or after the form or document, and there are multiple character categories due to the habit of the writer. When a similar character pattern is input, it is difficult to identify it, and there is a drawback in that it cannot be determined or misrecognition occurs.

FIG. 2 is a diagram showing comparison between the determination results according to the prior art and the present invention, (a) shows an example of characters written on a form, and (b) and (c) respectively show the prior art. , Shows the determination results according to the present invention. In FIG. 2 (a), five characters are entered in the character frame, but it is difficult to distinguish the second character 12 from the numeral "1" or "7", and the conventional character reading method shown in FIG. As indicated by reference numeral 51 in (b) of FIG. However, when a person reads a character, he or she does not necessarily pay attention to only one character and individually recognizes it, but also uses information of characters written before and after. In the example of the characters in FIG. 2A, when only one character is focused, it is difficult to identify whether the second character 12 is '1' or '7', but the fourth character 13 is surely Since it is possible to determine “7”, it is determined that the second character is likely to be “1”. It is difficult to obtain reading performance close to that of a human with the conventional character reading method in which the determination processing is performed for each character independently.

[0005]

A character reading method according to the present invention is a recognition dictionary comprising an average vector of feature vectors extracted from a set of learned character patterns and a principal component axis vector obtained by principal component analysis of the feature vectors. A character reading method for reading characters on a form or a document using an individual character recognition method that determines the category of a character pattern using the distance value between the vector and the extracted feature vector of the unknown input character pattern as a scale. An evaluation function defined for each category with a domain of a projection vector obtained by an inner product operation between a difference vector between the feature vector of the unknown input pattern and the average vector and the principal component axis vector, The evaluation function is selected and updated according to the area of the projection vector of the preceding and following characters on the form or document. And means for performing the determination process again on the basis of the evaluation value obtained from the evaluation function for the character pattern that cannot be determined in the determination process using the distance value and the evaluation value obtained from the evaluation function as a scale. It is characterized by

Further, the character reading apparatus of the present invention includes a feature extraction unit for performing a process of extracting character features, an average vector of feature vectors extracted by the feature extraction unit from a set of learned character patterns, and a principal component analysis of the feature vectors. It has a recognition dictionary memory for storing a recognition dictionary vector consisting of the principal component axis vector obtained by, and the distance value between the recognition vector and the feature vector extracted by the feature extraction unit from the unknown input character pattern In a character reading device that reads characters on a form or a document using an individual character recognition method that performs determination processing as a scale, a difference vector between the feature vector of the unknown input character pattern and the average vector, and the principal component axis vector And the projection vector register that stores the projection vector value obtained by the inner product operation between An evaluation function selection table memory having a vector register value as an address, an evaluation function selection register for storing data for selecting an evaluation function adapted to a projection vector area of characters before and after the form or document, and a selection An evaluation function table memory that outputs an evaluation value using the contents of the projection vector register as an address according to the evaluated function, and a determination unit that performs determination processing based on the distance value and the evaluation value are provided.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the character reading apparatus of the present invention. In the scanner unit 1, the optically scanned form image data is binarized to generate a monochrome binary level character string pattern. The character cutout unit 2 cuts out individual characters from the character string pattern based on the pitch information of the character string pattern and the like. The feature extraction unit 3 extracts the character features such as the grayscale feature and the outline feature of the character, and the feature vector f = (f ₁ ,
.., f _N ) is generated. In the recognition dictionary memory 4, an average vector g _i = (g _i1 , ..., g _iN ) obtained from a set of feature vectors of learning patterns of the character category C _i (i = 1, ..., M) to be recognized. And the principal component axis vector φ _ik ＝
A recognition dictionary vector consisting of (φ _ik1 , ..., φ _ikN ), (k = 1, ..., L) is stored, and the distance calculation unit 5 stores the feature vector f and the recognition dictionary vector of the category C _i. The distance value D _{i of} is calculated by the equation (1).

[0009]

The minimum distance value D _{j1 is stored in the} category register 6.
The category C _j1 that takes is stored.

[0011]

In the projection vector register 7, a projection vector R _i = (r _i1 , ..., R _ik , ..., _Which is obtained from the inner product of the feature vector f and the principal component axis vector φ _ik shown in the equation (3).
r _iL ) is stored.

R _ik = φ _ik · (f−g _i ) ... (3) The evaluation function selection table memory 8 is a memory that gives data for selecting an evaluation function adapted to the area of the projection vector for each category. The value of the category register 6 and the value of the projection vector register 7 are used as addresses, and the output data is the number of the function to be selected for each area of the projection vector of each category.

The evaluation function selection register 9 is a register prepared as many as the number of categories, and stores the latest value of the number of the function to be selected corresponding to each category.

The evaluation function table memory 10 stores data for giving an evaluation value according to the area of the projection vector for each category. The category name as the contents of the category register 6 and the evaluation function selection register 9 are stored. The evaluation value 16 is output as data, with the number of the function to be selected, which is the content of, and the value of the projection vector register 7 as an address.

As the first determination processing, the determination unit 11 compares the distance value D _i for each category obtained by the distance calculation unit 5 with a preset threshold value and the evaluation value 16 for each category. Is performed to determine whether to perform the category determination process or the undeterminable process. Further, only the character that cannot be determined in the first processing is subjected to the determination processing again, and based on the value of the evaluation value 16 which is the output of the evaluation function table memory 10, the category determination processing is performed or the determination is not possible. A second decision is made whether to process.

Next, the operation of this embodiment will be described. The scanner unit 1 optically scans and binarizes the character data written in the form shown in FIG. 2A, and the character cutout unit 2 cuts out individual characters. . The feature extraction unit 3 performs a conversion process on the feature vector f, and the distance calculation unit 5 calculates the distance between the feature vector and the recognition dictionary vector based on the equation (1). In the determination unit 11, the first candidate category C _j1 obtained by the equation (2) and (4)
The distance value of the second candidate category C _j2 obtained by the formula is checked.

[0018]

As thresholds for the distance value, ε (> 0), δ
(> 0) is defined, and D _j1 <δ (5) D _j2 -D _j1 > ε (6) Further, P _j1 (R _j1 ) is set for the output value Pi (Ri) of the evaluation function table memory. > 0 (7) P _i (R _i ) = 0 (for all i where i ≠ j1) (8)

In the first judgment, (5), (6),
When both the expressions (7) and (8) are satisfied, the determination process is performed on the category C _j1 , and otherwise the determination determination process is performed.
In the second determination, only the first undeterminable character is processed, and the determination category is set to C _j1 when both the expressions (7) and (8) are satisfied.

The first candidate category C _j1 obtained by the distance calculation unit 5 is stored in the category register 6,
The projection vector obtained by the equation (3) is stored in the projection vector register 7. Evaluation function selection table memory 8
For each category, an evaluation function number 14 to be used according to the area of the projection vector is output. When the determination result of the determination unit 11 performs the determination process for the category C _j1 , the evaluation function selection register 9 for each category is determined based on the determination signal 15.
The evaluation function number 14 is stored in. When the determination unit 11 does not satisfy the equations (5) and (6), the determination signal 15 is not output and the evaluation function number 14 is not stored in the evaluation function selection register 9.

In the example of the character pattern shown in FIG. 2A, the result obtained by the determination unit 11 is the second character 12 because the characters other than the second character 12 have standard character shapes.
Only the character cannot be determined, and the other characters are subjected to category determination processing. Therefore, the second character 12 and the fourth character 13 will be described in detail for the present embodiment.

FIG. 3 is a diagram for explaining the recognition processing of the unknown input pattern in the individual character recognition method using the principal component analysis method, and the distribution of the feature vector obtained from the set of learning character patterns of the number "1". 18 shows a distribution 19 of learning character patterns of the numeral "7". In the principal component analysis method obtained as the eigenvalue expansion of the covariance matrix, the first principal component axis vector φ ₁₁ and the orthogonal second principal component axis vector φ ₁₂ are obtained through the mean vector g ₁ of the feature vectors of '1'. To be Similarly for '7', the average vector g ₇
The first principal component axis vector φ ₇₁ and the second principal component axis vector φ
₇₂ is obtained. The distance between the feature vector f of the unknown input pattern and the recognition dictionary vector is the length of a perpendicular line to the plane generated from the feature vector f by the average vector and the principal component axis vector, and the distance D _{1 from the} category '1' and the category ' It is obtained as the distance D _{7 from} 7 '. Further, the projection vectors R ₁ and R ₇ obtained by the equation (3) are obtained as the position vector of the intersection of the perpendicular of the distances D ₁ and D ₇ with the respective average vectors as the origin and the plane. In the case of the distribution of the feature vectors of FIG. 3, there is a region 29 in which the feature vectors of the two categories “1” and “7” are close to each other, and FIG.
In the similar character shapes of “1” and “7” as shown in the second character 12 of, the feature vector f generally exists in the proximity area 29.

FIG. 4 is a diagram for explaining the evaluation function in the present invention. With reference to FIG. 4, the second function shown in FIG.
A determination method when the character data indicated by the 12th character and the 4th character 13 is input will be described. In FIG. 4, f ⁽²⁾ is the feature vector of the second character 12, and f ⁽⁴⁾ is the fourth
A feature vector of the thirteenth character is shown, and a vector obtained by projecting the feature vector f ⁽²⁾ is a projection vector R ₇ ⁽²⁾ , based on the average vector of the recognition dictionary of '7' and the principal component axis vector.
The vector obtained by projecting the feature vector 32 is the projection vector R
₇ ⁽⁴⁾ . The component of the projection vector R ₇ ⁽²⁾ with respect to the first principal component axis is r ₇ ⁽²⁾ , and the component of the projection vector R ₇ ⁽⁴⁾ with respect to the first principal component axis is r ₇ ⁽⁴⁾ . The lower half of FIG. 4 shows an example of an evaluation function showing the relationship between the evaluation value and the component of the projection vector in the recognition dictionary of '7' with respect to the first principal component axis.

In this example, the evaluation function represents the reliability. The evaluation value is larger as it is closer to the average vector of the set of learning character patterns, and the evaluation value is smaller as the distance from the average vector is increased. . Further, as a feature of the present invention, a plurality of evaluation functions can be stored, and the evaluation function can be changed based on the information of the projection vectors of the preceding and succeeding characters. In the example of FIG. 4, two types of evaluation functions having different slopes, that is, an evaluation function P ₇₁ (r) and an evaluation function P ₇₂ (r) are provided as evaluation functions for the category “7”,
Similarly, the evaluation function P ₁₁ (r) and the evaluation function P ₁₂ (r) are provided as the evaluation function for the category “1”.

The evaluation function is stored in the evaluation function table memory 10, and a plurality of evaluation functions are stored for each category. The kind of evaluation function to be used is designated in the evaluation function selection register 9.

FIG. 5 is a diagram for explaining a method of selecting an evaluation function according to the present invention. The contents of the evaluation function selection table memory 8 will be described with reference to FIG. Region 4 where the projection vector of the feature vector is the average vector
When it is 0, the evaluation function P ₇₁ (r) is selected, and when it is in the area 41 away from the average vector, the evaluation function P ₇₁ (r) is selected.
₁₁ Select (r). As the initial value of the evaluation function selection register 9, a value is set so as to select the evaluation function P ₁₁ (r). That is, the evaluation function P ₁₁ shown in FIG.
(R) and the evaluation function P ₇₁ (r) are used.

In the first determination, when the second character 12 in FIG. 2 is input, (5), (6), (7),
Since the judgment result is rejected because the expression (8) is not satisfied, the judgment signal 15 is not generated, and the contents of the evaluation function selection table memory 8 are not stored in the evaluation function selection register 9. Therefore, the contents of the evaluation function selection register 9 are initially set. The evaluation function P ₇₁ (r) is selected with the value as it is, but when the fourth character 13 is input, the projection vector R ₇ ⁽⁴⁾ corresponds to the area 40, so that the number of the evaluation function P ₇₂ (r) is The evaluation function selection table memory 8 outputs the evaluation function number 14, and the judgment unit 11 also satisfies the expressions (5), (6), (7), and (8). Function P
_{The 72} (r) number is stored in the evaluation function selection register 9 of the category of “7”. That is, after the fourth character 13 is recognized, P ₇₂ (r) is used as the evaluation function.

In the second judgment, the information of the second character 12 is set again as an undecidable character, and since the area of the projection vector is in the area 40, the evaluation function P is the output of the evaluation function selection table memory 8. ₇₁ (r) is selected, but since the expressions (5) and (6) are not satisfied, the determination signal 15 is not output and the contents of the evaluation function selection register 9 are the same as those of the evaluation function P ₇₂ (r). Up to now, P ₇₂ (r) is used as the evaluation function. As an output of the evaluation function table memory 10, P ₁₁ (R ₇ ⁽²⁾ )> 0, P ₇₂ (R ₇
^{Since (2)} ) = 0, the conditional expressions (7) and (8) are satisfied, and the category "1" is obtained as the determination result 56 as shown in (c) of FIG.

Further, since the character information necessary for the second determination is only the category name and the projection vector,
If the values of the category register 6 and the projection vector register 7 are saved when the determination unit 11 cannot determine, the scanner unit 1, the character cutting unit 2, the feature extraction unit 3,
The process in the distance calculation unit 5 does not need to be performed again.

In this embodiment, characters are used as targets, but the same operation is applied to images, voices, and graphics.

[0032]

As described above, according to the present invention, based on the information of the character data written before and after the form,
It is possible to read characters similar to a plurality of categories due to the habit of a writer. Further, since the evaluation function is defined in the space compressed by using the principal component analysis, it is efficient in saving the character feature information before and after. Furthermore, the evaluation function in the present invention can be set freely according to the system, and the evaluation function to be used can also be freely selected, so that flexible reading processing is possible. Further, the present invention can be easily realized not only for characters but also for images, sounds and figures.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of a character reading device of the present invention.

FIG. 2 is a diagram showing comparison of determination results according to a conventional technique and the present invention.

FIG. 3 is a diagram for explaining a recognition process of an unknown input pattern in the individual character recognition method using the principal component analysis method.

FIG. 4 is a diagram for explaining an evaluation function in the present invention.

FIG. 5 is a diagram for explaining a method of selecting an evaluation function according to the present invention.

[Explanation of symbols]

 1 Scanner Section 2 Character Extraction Section 3 Feature Extraction Section 4 Recognition Dictionary Memory 5 Distance Calculation Section 6 Category Register 7 Projection Vector Register 8 Evaluation Function Selection Table Memory 9 Evaluation Function Selection Register 10 Evaluation Function Table Memory 11 Evaluation Section 14 Evaluation Function Number 15 Judgment signal 16 Evaluation value

Claims (2)

[Claims] 1. A recognition dictionary vector composed of an average vector of feature vectors extracted from a set of learning character patterns and a principal component axis vector obtained by principal component analysis of the feature vectors and an extracted feature vector of unknown input character patterns. In a character reading method for reading characters on a form or a document by using an individual character recognition method for determining a category to which a character pattern belongs, using a distance value between and as a scale, the feature vector of the unknown input pattern and the average vector And the evaluation function defined for each category with the area of the projection vector obtained by the inner product operation between the difference vector and the principal component axis vector, and the projection vector of the preceding and following characters on the form or document. Of the evaluation function by selecting and updating the evaluation function adapted to the region of The character reading method is characterized in that the judgment process is performed again on the character pattern that cannot be judged in the judgment process using the evaluation value obtained from the above as a scale based on the evaluation value obtained from the evaluation function. 2. A feature extraction unit for extracting character features, and an average vector of feature vectors extracted by the feature extraction unit from a set of learned character patterns and a principal component axis vector obtained by principal component analysis of the feature vectors. A recognition dictionary memory that stores a recognition dictionary vector consisting of and, and performs a determination process using the distance value between the feature vector extracted from the feature extraction unit from the unknown input character pattern and the recognition dictionary vector as a scale. In a character reading device for reading characters on a form or a document by using a character recognition method, it is obtained by an inner product operation between a difference vector between the feature vector of the unknown input character pattern and the average vector and the principal component axis vector. Add the projection vector register that stores the projection vector value and the value of this projection vector register. Memory, an evaluation function selection table that stores data for selecting an evaluation function adapted to the area of the projection vector of the preceding and following characters on the form or document, and the evaluation function selection table according to the selected evaluation function. An evaluation function table memory for outputting an evaluation value with the contents of the projection vector register as an address,
A character reading device comprising: a determination unit that performs determination processing based on the distance value and the evaluation value.