JP2906591B2 - Character classification method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a character classification system in optical character recognition.

[Conventional technology]

With the diversification of information processing systems, various data input methods are required, and character recognition technology is being put to practical use as a powerful data input method. However, the current character recognition technology is far inferior to humans in character reading ability, and a character reading device having higher recognition ability is desired.

In order to enhance the character recognition ability, improvements are being made in each of a pre-processing method, a normalization method, a feature extraction method, a classification / identification method, and a post-processing method that constitute the character recognition method.

Among the character recognition technologies, Kanji recognition has many types of characters to be recognized, complicated shapes, many sets of characters with similar shapes,
In multi-font printing kanji recognition and handwritten kanji recognition, the variety of shape deformation has been a problem. In this,
Regarding the problem of the large number of character types to be recognized, a hierarchical classification process that narrows down the character type candidates by relatively simple processing and then produces the final recognition result by careful processing is generally used. I have. The processing at the first stage of the classification hierarchy is generally called a large classification or a rough classification, and the classification processing at a later stage is often called a detailed classification or identification. There are many methods known as major classifications in printed kanji recognition, but as a pioneering and proven method, for example, the document "2,000
Development of printed Kanji OCR that reads character types at 100 characters / second, Kenichi Mori, Kunio Sakamoto, Nikkei Electronics, 1977.10.31, pp.102-12
8 (October 1977) ". There are also many proposals for large classification in handwritten kanji recognition. For example, in the literature, "large classification by stroke density function of handwritten kanji, Naito and Yodogawa, IETF Technical Report PRL79-
3 (1979) is known. Kanji characters have simple shapes such as "day", slightly complicated shapes such as "den", and very complicated shapes such as "taki". The character type candidates are narrowed down on the basis of the standard to reduce the burden of detailed classification or identification processing.

[Problems to be solved by the invention]

A problem in the classification process featuring the complexity of the character shape is the setting of the complexity scale. Although the above-mentioned “complexity index” has a good track record in printed kanji recognition, there is a problem in that stability is a feature lacking when the shape of a character is largely fluctuated like handwritten kanji. Similarly, the "stroke density function" described above was a pioneering attempt in the study of handwritten kanji recognition, but also left a problem with stability as a feature,
Often used with other features.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character classification system that takes up the complexity of a character shape as a feature of large classification in kanji recognition, stably extracts the complexity of the character shape, and effectively utilizes the feature.

[Means for solving the problem]

According to a first aspect of the present invention, there is provided a character classification method, comprising: a character pattern storage unit for storing a character pattern represented as a binary digital figure on a two-dimensional lattice; and a character pattern as storage content from the character pattern storage unit. Read and store two sufficiently large values in the storage locations corresponding to the pixels in the character portion, and store in the storage locations corresponding to the pixels in the background portion of the character pattern in the horizontal direction from the pixels in the background portion of the pattern. The number of pixels in the continuous background,
A characteristic pattern creating unit that stores a set of the number of pixels of the background portion continuing in the vertical direction from the pixels of the background portion of the pattern, and reading a value stored in the storage location from the characteristic pattern creating unit, A process of vertically adding the reciprocal of the value of the background pixel continuous in the horizontal direction of the pixel of the background portion or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion to obtain the sum is performed in each vertical direction on the character pattern. A histogram in the horizontal direction by performing on the column of pixels in the same manner, read the value stored in the storage location from the characteristic pattern creation unit, and calculate the background pixel continuous in the vertical direction of the pixel in the background portion. A process of adding the reciprocal of the value or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion in the horizontal direction to obtain the sum is performed for each horizontal pixel column on the character pattern. To create a vertical histogram,
Characteristic histogram extraction means for storing the values of the horizontal histogram and the vertical histogram, reading the values of the horizontal histogram and the values of the vertical histogram, and referencing each of the recognition dictionaries with the value of each histogram as a feature value And a classification means for performing a classification process by collating with a pattern and outputting a classification result.

According to a second aspect of the present invention, there is provided a character classification method, comprising: a character pattern storage means for storing a character pattern represented as a binarized digital figure on a two-dimensional grid; and a character pattern as storage content from the character pattern storage means. Read and store two sufficiently large values in the storage locations corresponding to the pixels in the character portion, and store in the storage locations corresponding to the pixels in the background portion of the character pattern in the horizontal direction from the pixels in the background portion of the pattern. The number of pixels in the continuous background,
A characteristic pattern creating unit that stores a set of the number of pixels of the background portion continuing in the vertical direction from the pixels of the background portion of the pattern, and reading a value stored in the storage location from the characteristic pattern creating unit, A process of vertically adding the reciprocal of the value of the background pixel continuous in the horizontal direction of the pixel of the background portion or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion to obtain the sum is performed in each vertical direction on the character pattern. A histogram in the horizontal direction by performing on the column of pixels in the same manner, read the value stored in the storage location from the characteristic pattern creation unit, and calculate the background pixel continuous in the vertical direction of the pixel in the background portion. A process of adding the reciprocal of the value or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion in the horizontal direction to obtain the sum is performed for each horizontal pixel column on the character pattern. To create a vertical histogram,
Characteristic histogram extraction means for storing the values of the horizontal histogram and the vertical histogram; reading the value of the horizontal histogram and the value of the vertical histogram; calculating the average value of the horizontal histogram; A density feature extraction means for storing in a storage location, calculating an average value of the vertical histogram and storing the average value in the storage location as a vertical density feature; reading values of the horizontal density feature and the vertical density feature; And a classification means for performing a classification process as an amount and outputting a classification result.

According to a third aspect of the present invention, there is provided a character classification method, comprising: a character pattern storage unit for storing a character pattern represented as a binarized digital figure on a two-dimensional grid; and a character pattern as storage content from the character pattern storage unit. Read and store two sufficiently large values in the storage locations corresponding to the pixels in the character portion, and store in the storage locations corresponding to the pixels in the background portion of the character pattern in the horizontal direction from the pixels in the background portion of the pattern. The number of pixels in the continuous background,
A characteristic pattern creating unit that stores a set of the number of pixels of the background portion continuing in the vertical direction from the pixels of the background portion of the pattern, and reading a value stored in the storage location from the characteristic pattern creating unit, A process of vertically adding the reciprocal of the value of the background pixel continuous in the horizontal direction of the pixel of the background portion or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion to obtain the sum is performed in each vertical direction on the character pattern. A histogram in the horizontal direction by performing on the column of pixels in the same manner, read the value stored in the storage location from the characteristic pattern creation unit, and calculate the background pixel continuous in the vertical direction of the pixel in the background portion. A process of adding the reciprocal of the value or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion in the horizontal direction to obtain the sum is performed for each horizontal pixel column on the character pattern. To create a vertical histogram,
Characteristic histogram extraction means for storing the values of the horizontal histogram and the vertical histogram; reading the value of the horizontal histogram and the value of the vertical histogram; calculating the average value of the horizontal histogram; A density feature extraction means for storing in a storage location, calculating an average value of the vertical histogram and storing the average value in the storage location as a vertical density feature; reading values of the horizontal density feature and the vertical density feature; A first classification unit that performs a classification process as an amount and stores a classification result; reads the value of the horizontal histogram and the value of the vertical histogram from the characteristic histogram extraction unit; The classification result is read and set as a candidate character type, and the value of each histogram is used as a feature amount in the recognition dictionary. Wherein performs classification processing by collating the candidate character types of the reference pattern, characterized in that it consists of a second classifying means for outputting the detailed classification result.

[Action]

A feature extraction method of a glyph complexity measure which plays a central role of the present invention will be described.

FIG. 4 is a diagram for explaining graphic processing for feature extraction. FIG. 4 (a) is a diagram showing the lengths of horizontally continuous pixels in the background portion (black pixel portion) and the background portion (white pixel portion) of the figure, and FIG. () Is a diagram showing the length of vertically consecutive pixels in the background portion of the figure. That is, on the ij coordinate system as shown in FIG. 4, if the pixel (i, j) is a pixel of the background portion, the length of the column of the pixel of the background portion that continues horizontally and vertically through the pixel is Is determined. This will be referred to as a horizontal and vertical interval between the background pixels (i, j). In FIG. 4, the horizontal interval between pixels (i, j) is h _ij , and the vertical interval is v _ij
And the horizontal interval between the pixels (i ′, j ′) is
h _i ′ _j ′ and the vertical interval is v _i ′ _j ′.

If the pixel (i, j) and the pixel (i + 1, j) is a background pixel, the horizontal spacing h _ij and h _{i + 1, j,} it holds the _{h ij = h i + 1,} j, also the pixel (i, j) and If the pixel (i, j + 1) is a background pixel, v _ij = vi _{, j + 1} holds for the vertical intervals v _ij and v _{i, j + 1} . That is, the horizontal interval between the background pixels adjacent in the horizontal direction is continuous and the vertical interval between the background pixels adjacent in the vertical direction is equal.

FIG. 5 is a diagram for explaining a method of determining a scale of complexity in the horizontal direction and the vertical direction. The measure of horizontal complexity is determined by the following procedure.

The horizontal interval _hij of all the background pixels (i, j) is obtained.

H _kL is set to a sufficiently large value for the pixel (k, l) in the figure portion (ground portion). For example, the length of the horizontal width of the figure, or a value several times, several tens times, or the like may be used.

Accumulating the reciprocal of sufficiently large value corresponding to the fractional pixel of the inverse or the value of the horizontal gap h _ij vertically. That is, the following equation is calculated.

A histogram of Figure 5 to calculate the H _i for the columns of the pixels (a). This is called a horizontal characteristic histogram.

Calculating an average value of the horizontal direction characteristic histogram H _i. This will be referred to as a horizontal density feature.

The horizontal characteristic histogram shows a measure of horizontal complexity as a function or as a vector quantity, and the horizontal density feature shows a measure of horizontal complexity as a scalar quantity.

The method of determining the vertical complexity scale is the same as the following procedure.

The vertical interval v _ij of all the background pixels (i, j) is obtained.

Vh _kL for pixel (k, l) in figure part (other part)
Is set to a sufficiently large value. For example, the length in the vertical direction of the figure, or a value several times, several tens times, or the like may be used.

Accumulating the reciprocal of sufficiently large value corresponding to the fractional pixel of the inverse or the value of the horizontal gap h _ij in the horizontal direction. That is, the following equation is calculated.

The histogram of FIG. 5B is obtained by calculating V _j for each pixel column. This is called a vertical characteristic histogram.

Calculating a mean value in the vertical direction characteristic histogram V _j. This will be referred to as a vertical density feature.

The vertical characteristic histogram shows a measure of vertical complexity as a function or as a vector quantity, and the vertical density feature shows a measure of vertical complexity as a scalar quantity.

FIG. 6 is a diagram showing an example of the complexity of a graphic and a horizontal characteristic histogram corresponding thereto. FIG. 6 (a) shows an example of a complicated figure, and FIG. 6 (b) shows an example of a simple figure. Comparing the horizontal characteristic histogram H _i for each figure, FIG. 6 high density in the horizontal direction (a)
In FIG. 6B, the histogram also has a large value, and in FIG. 6B, the histogram generally has a small value. The same is true for the vertical direction, where the horizontal and vertical characteristic histograms measure the complexity of the figure.

〔Example〕

FIG. 1 is a diagram showing the configuration of an embodiment of the character classification system of the first invention.

This character classification method includes a character pattern storage unit 1, a characteristic pattern creation unit 2, a characteristic histogram extraction unit 3,
And a large classification means 5.

The character pattern storage unit 1 stores a binarized digital figure, and may be a normal storage unit.

The characteristic pattern creation unit 2 reads the character pattern signal from the character pattern storage unit 1, obtains the horizontal and vertical intervals of the background pixel described in the section of the operation, stores the value in each storage location, and A sufficiently large value is stored in a storage location corresponding to a pixel in the ground portion of the figure, and can be realized by a conventional technique.

The characteristic histogram extracting means 3 reads the values of the horizontal and vertical intervals or sufficiently large values from each storage location of the characteristic pattern creating means 2 and reads the horizontal characteristic histogram and the vertical characteristic histogram described in the section of the operation. Is calculated, and the values of these two characteristic histograms are stored, and can be realized by a conventional technique.

The large classifying means 5 reads the horizontal characteristic histogram and the vertical characteristic histogram from the characteristic histogram extracting means 3 and performs, for example, a collation calculation with reference patterns prepared for each character type using these as characteristic amounts. A similarity or dissimilarity is determined by the above-described method, and classification processing in ordinary character recognition is performed, in which ordering is performed in order of reliability, and can be realized by a conventional technique. The matching calculation in this case includes various methods, such as a method of performing a linear vector calculation by regarding a histogram as a multidimensional vector, and a method of performing a nonlinear vector calculation by regarding a histogram as a series of scalar quantities. Recognition, supervised by Kenichi Mori,
(No. 1) of IEICE (November 1988)
Chapter Pattern Recognition Theory.

FIG. 2 is a diagram showing the configuration of an embodiment of the character classification system according to the second invention.

This character classification method includes a character pattern storage unit 1, a characteristic pattern creation unit 2, a characteristic histogram extraction unit 3,
It comprises a density feature extraction means 4 and a coarse classification means 6.

The character pattern storing means 1, the characteristic pattern creating means 2, and the characteristic histogram extracting means 3 are the same as the respective components of the character classification method shown in FIG.

The density feature extraction means 4 is a characteristic histogram extraction means 3
From the horizontal characteristic histogram and the vertical characteristic histogram, and as described in the operation section, calculate the average value of the horizontal characteristic histogram to obtain the horizontal density characteristic, and calculate the average of the vertical characteristic histogram values. A value is calculated and two values of each density feature are stored as a vertical density feature, and can be realized by a conventional technique.

The rough classification means 6 reads the horizontal density characteristic and the vertical density characteristic from the density characteristic extraction means 4 and sets them as characteristic amounts, and performs classification processing in ordinary character recognition such that the character types are ordered in order of reliability. And can be realized by conventional techniques. In this case, for example, two density features
A method of ordering in ascending order by calculating a distance from a reference vector prepared for each character type assuming that it is a dimensional vector, or a table using two feature quantities (H, V) as shown in FIG. , And selecting a candidate character type determined at each position of the table.

FIG. 3 is a diagram showing the configuration of an embodiment of the character classification system of the third invention.

This character classification method includes a character pattern storage unit 1, a characteristic pattern creation unit 2, a characteristic histogram extraction unit 3,
It is composed of a density feature extracting means 4, a first classifying means 7, and a second classifying means 8.

The character pattern storing means 1, the characteristic pattern creating means 2, the characteristic histogram extracting means 3, and the density feature extracting means 4 are the same as the respective constituent elements of the character classification method shown in FIGS.

The first classifying unit 7 reads the horizontal density feature and the vertical density feature from the density feature extracting unit 4 and sets them as feature amounts, and performs classification in normal character recognition such that the character types are ordered in the order of reliability. The processing is performed to store the candidate character types, and can be realized by the conventional technology in the same manner as the rough classification means 6 described with reference to FIG.

The second classifying unit 8 reads candidate character types from the first classifying unit 7, reads a horizontal direction characteristic histogram and a vertical direction characteristic histogram from the characteristic histogram extracting unit 3, and uses these as characteristic amounts. Similarity or dissimilarity is determined and ordering is performed efficiently in order of reliability by performing collation calculation only with the reference pattern of the kind. As in the large classification process 5 in FIG. Can be realized.

〔The invention's effect〕

As described above, according to the present invention, stable character classification based on the complexity of the shape of a character pattern can be realized, which is very useful for improving the performance of kanji recognition and expanding applications.

[Brief description of the drawings]

FIG. 1 is a diagram showing the configuration of an embodiment of the character classification system of the first invention, FIG. 2 is a diagram showing the configuration of an embodiment of the character classification system of the second invention, and FIG. FIG. 4 is a diagram showing the configuration of an embodiment of the character classification system of the present invention, FIG. 4 is a diagram for explaining graphic processing for feature extraction, and FIG. 5 is a scale for determining the complexity in the horizontal and vertical directions. FIG. 6 is a diagram for explaining the method, FIG. 6 is a diagram showing an example of the complexity of the graphic and the corresponding horizontal characteristic histogram, and FIG. 7 is a diagram for determining the position of the table by two feature amounts (H, V). FIG. 9 is a diagram for explaining a method of selecting a candidate character type determined at each position of the table. 1 ... Character pattern storage means 2 ... Character pattern creation means 3 ... Characteristic histogram extraction means 4 ... Density feature extraction means 5 ... Large classification means 6 ... Rough classification means 7 ... First classification means 8 ... ... Second classification means

Claims (3)

(57) [Claims] 1. A character pattern storage means for storing a character pattern represented as a digitized digital figure on a two-dimensional lattice, and a character pattern as storage content is read from the character pattern storage means, and a character part is read. Two sufficiently large values are stored in the storage locations corresponding to the pixels, and the storage locations corresponding to the pixels in the background portion of the character pattern are stored in the storage locations corresponding to the pixels in the background portion of the pattern portion in the left-right direction. A characteristic pattern creating unit that stores a set of the number of pixels and the number of pixels in the background portion that is vertically continuous from the pixels in the background portion of the pattern; and stored in the storage location from the characteristic pattern creating unit. The value is read and the reciprocal of the value of the background pixel continuous in the left-right direction of the pixel of the background portion or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion is added in the vertical direction to obtain the sum. Is performed on each vertical pixel column on the character pattern to create a horizontal histogram, and similarly, the value stored in the storage location is read from the characteristic pattern creating means, and the background portion is read. The process of adding the reciprocal of the value of the background pixel continuous in the vertical direction of the pixel in the vertical direction or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion in the horizontal direction to obtain the sum is performed in each horizontal pixel on the character pattern. Characteristic histogram extracting means for generating a vertical histogram by performing the above process on the columns, storing the values of the horizontal histogram and the vertical histogram, and reading the values of the horizontal histogram and the values of the vertical histogram. A classifier that performs a classification process by comparing each histogram value with each reference pattern of a recognition dictionary as a feature amount and outputs a classification result. Character classification system consisting of. 2. A character pattern storage means for storing a character pattern represented as a binarized digital figure on a two-dimensional lattice, and a character pattern which is a storage content is read from the character pattern storage means, and Two sufficiently large values are stored in the storage locations corresponding to the pixels, and the storage locations corresponding to the pixels in the background portion of the character pattern are stored in the storage locations corresponding to the pixels in the background portion of the pattern in the left-right direction. A characteristic pattern creation unit that stores a set of the number of pixels and the number of pixels in the background portion that is vertically continuous from the pixels in the background portion of the pattern; and stored in the storage location from the characteristic pattern creation unit. The sum is obtained by reading the values and adding in the vertical direction the reciprocal of the value of the background pixel continuous in the left-right direction of the pixel of the background portion or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion. Is performed on each vertical pixel column on the character pattern to create a horizontal histogram, and similarly, the value stored in the storage location is read from the characteristic pattern creating means, and the background portion is read. The process of adding the reciprocal of the value of the background pixel continuous in the vertical direction of the pixel in the vertical direction or the reciprocal of a sufficiently large value corresponding to the pixel of the character part in the horizontal direction to obtain the sum is performed on each horizontal pixel on the character pattern. Characteristic histogram extracting means for generating a vertical histogram by performing the above process on the columns, storing the values of the horizontal histogram and the vertical histogram, and reading the horizontal histogram value and the vertical histogram value Calculating the average value of the horizontal histogram, storing it in a storage location as a horizontal density feature, and calculating the average value of the vertical histogram Density feature extraction means for storing the horizontal density feature and the vertical density feature in a storage location as vertical density features, performing classification processing using both density features as feature amounts, and outputting a classification result. Character classification system. 3. A character pattern storage means for storing a character pattern represented as a binarized digital figure on a two-dimensional lattice, and a character pattern as a storage content is read from the character pattern storage means, Two sufficiently large values are stored in the storage locations corresponding to the pixels, and the storage locations corresponding to the pixels in the background portion of the character pattern are stored in the storage locations corresponding to the pixels in the background portion of the pattern in the left-right direction. A characteristic pattern creation unit that stores a set of the number of pixels and the number of pixels in the background portion that is vertically continuous from the pixels in the background portion of the pattern; and stored in the storage location from the characteristic pattern creation unit. The sum is obtained by reading the values and adding in the vertical direction the reciprocal of the value of the background pixel continuous in the left-right direction of the pixel of the background portion or the reciprocal of a sufficiently large value corresponding to the pixel of the character portion. Is performed on each vertical pixel column on the character pattern to create a horizontal histogram, and similarly, the value stored in the storage location is read from the characteristic pattern creating means, and the background portion is read. The process of adding the reciprocal of the value of the background pixel continuous in the vertical direction of the pixel in the vertical direction or the reciprocal of a sufficiently large value corresponding to the pixel of the character part in the horizontal direction to obtain the sum is performed on each horizontal pixel on the character pattern. Characteristic histogram extracting means for generating a vertical histogram by performing the above process on the columns, storing the values of the horizontal histogram and the vertical histogram, and reading the horizontal histogram value and the vertical histogram value Calculating the average value of the horizontal histogram, storing it in a storage location as a horizontal density feature, and calculating the average value of the vertical histogram Density feature extraction means for storing in a storage location as a vertical density feature by reading values of the horizontal density feature and the vertical density feature, performing a classification process using both density features as feature amounts, and storing a classification result Means, reading the value of the horizontal direction histogram and the value of the vertical direction histogram from the characteristic histogram extracting means, and reading the classification result of the first classifying means as a candidate character type; A second classification unit that performs a classification process by comparing the candidate character type in the recognition dictionary with a reference pattern as a feature amount and outputs a detailed classification result.