JPS5855551B2 - Discrimination feature extraction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention is an identification feature extraction device, particularly for extracting features of character lines constituting recognition target characters in a recognition system that targets characters that are complex and have many categories, such as handwritten kanji figures. The present invention relates to a discriminative feature extraction device.

Conventionally, the following devices have been developed as this type of device.

In other words, in a handwritten kanji recognition system, there is a device that microscopically traces character lines and extracts the geometric shapes of character lines, such as end intersections and bending points, as features, or Alternatively, a device has been developed that approximates an appropriate curve and extracts the approximation parameters as features.

In addition, in the printed kanji identification system, ■ The outline of the character line is observed with an appropriate mask, the direction of the character line outline within the mask is determined, and this is scanned and counted over the entire screen. and a device that extracts numerical values representing complexity in the vertical direction, or ■A device that projects the density distribution of characters onto the vertical and horizontal axes and extracts a density distribution function that is degenerate in the direction of the axes as a feature, or ■Characters A device that divides the screen into appropriate sizes and extracts the average density of the divided areas as a feature, or ■ Sets four rectangular inspection areas around the top, bottom, left and right of the characters, and examines the character lines in the areas. A device that extracts four types of codes as features from each area by counting the amount, or ■ Counting the proportion of black part in the peripheral area at each stage while gradually thickening the character line, and calculating the value. Devices have been developed that extract characteristics that change with thickening processing, or (1) devices that extract the outline stroke shape of character figures.

However, among the above-mentioned devices, the device that focuses on the end intersections of character lines for handwritten kanji has the characteristic and difficulty of performing microscopic tracing of the character lines.
Due to handwriting deformations such as local contact or separation of character lines that normally occur, extracted features vary greatly and become unstable.

Therefore, in a recognition system that uses these features, in order to absorb these instability in the identification part, a huge amount of data or a long identification logic is prepared as a dictionary, and a large amount of counting and complicated judgment processing are required. must be done.

By the way, a recognition system for 881 educational kanji is said to require several minutes of processing time on a high-speed computer.

In addition, a system that recognizes kanji by approximating kanji using line segments only achieves a recognition rate of about 85φ for 200 character types, and a practical system that targets kanji of ordinary use as it is. It seems difficult to extend this to

Furthermore, a common drawback of both is that the structure of character lines in kanji figures ranges from very simple to complex, and although it would seem that a broad classification that utilizes this property is effective, classification based on this information is difficult. Instead, both simple character identification and complex character identification are determined using the same complicated procedure, which is not necessarily efficient.

On the other hand, among the above-mentioned devices, in the recognition system for printed kanji, practical devices are being constructed for kanji used in a single font; At present, we believe that it is impossible to use the conventional technology as it is in a system that targets handwritten kanji, which are much more diverse and include large variations.

Currently, various feature extraction devices have been developed for handwritten alphanumeric characters, kana, etc., and have achieved great results.

From this, it is possible to consider extending these devices to include handwritten kanji.

However, not only quantitative expansion but also qualitative development that can deal with the complexity of kanji is necessary.

In fact, only a few attempts have been made to evaluate the capabilities of conventional devices for handwritten Chinese characters.

In addition, as a recognition method for printed alphabets, the white part of the character background is divided into regions according to the number of character lines viewed from there in the vertical and horizontal directions, and a vector is calculated by counting the total area for each type of region. Attempts have been made to form a vector and recognize the vector as a feature of a character.

However, this feature extraction method had the following three drawbacks.

The first point is the type of region, that is, the number of features that can appear.

The type of feature is determined by the number of character lines seen in the vertical, horizontal, and horizontal directions.

In the conventional configuration described above, since the alphabet was targeted, the characters were simple.

Therefore, the number of character lines is limited to three types: 0, 1, and 2 or more.

Therefore, 3 types each in the up, down, left and right directions, total 3X3
We considered the characteristics of 81 types of X3X3-81.

Now, when the recognition target is a complex character like a kanji, the number of character lines naturally increases, and sufficient recognition cannot be achieved if two or more lines are treated as one.

In the case of kanji, there are cases where at least 10 or more character lines are lined up, so if you treat them separately from 0 to 10, it will be 11'.
= 14641 types of features may appear.

That is, as the characters to be recognized become more complex, the configuration of the recognition device becomes extremely difficult.

The second drawback is the problem of the appearance position of features.

In this conventional method, the areas on the screen are totaled to form a vector regardless of the position of the area.

This is not a problem when the character has a simple shape because it can express to some extent which part of the character the feature itself is extracted from, but it is an effective method for objects that are complex and have many similar categories like kanji. isn't it.

In other words, a problem arises in which characters of the same type are extracted with only different appearance positions and cannot be distinguished from each other.

A third drawback is that differences between individual features are not calculated.

In other words, even if the actual character shapes are very similar, if different types of features are assigned to them,
The two cases were treated no differently than two completely different shapes.

For example, if there is an area with 5 character lines in the vertical and horizontal directions, and 5 character lines in the vertical and left directions, and 6 character lines in the right direction, even though the two do not actually differ much, Its area is counted as a completely different feature.

Such differences easily occur due to deformation;
It is unreasonable to treat these things in the same way as two completely different things.

The conventional identification method using the characteristics must be complicated in order to absorb this variation.

The present invention was made to solve the above problems, and
For example, if you look at characters from the horizontal or vertical coordinate axis,
The number of character lines constituting a character is counted along the coordinate axis and extracted as a vector on the coordinate axis of the character figure.

This is called a stroke density function that expresses the complexity of a character graphic with respect to the coordinate axis.

Characters are identified by matching by comparing each element as a vector of this function.

The aim is to utilize the properties of kanji shapes, which have a variety of character line density distributions, from simple to complex configurations, to efficiently classify and recognize them.

The drawings will be explained in detail below. FIG. 1 shows an embodiment conceptually representing an identification feature extraction device according to the present invention.

Reference numeral 1 denotes a register for storing recognized characters, with the horizontal direction being the X axis and the vertical direction being the Y axis, each having a resolution of 128 bits.

2 and 3 are character line number counting devices that scan the characters stored in the register 1 in parallel with the horizontal or vertical axis and count the number of character lines using the position on the vertical or horizontal coordinate axis as a parameter.

4 and 5 are 128-word registers that store the number of character lines in the axis direction, that is, the density of character lines counted by the counting device 2 and the counting device 3, as a function regarding the coordinate axes.

FIG. 2 shows in detail the register 2 or 3 shown in FIG. This is one column of 128-bit data.

6 reads out 2 bits in parallel while moving 1 bit at a time from one end of the 128-bit data 1', and sequentially reads 1 bit at a time.
This is a bit extraction circuit that outputs to the device 7 27 times.

Reference numeral 7 denotes a pattern extraction device which outputs the value “1” to the device 8 described later only when the right bit of the 2-bit data read out by the bit extraction circuit 6 is “0” and the left bit is “1”. be.

8 is a counting device which counts the number of [1's] outputted by the pattern extraction device 7 and outputs the result to the counting device 2 or the counting device 3 in FIG.

FIG. 3 is a block diagram of a matching section that compares features.

The stroke density functions accumulated in storage devices 4 and 5 of FIG. 1 are connected and stored in a 256 word register shown at 9 in FIG. 3.

Reference numeral 10 denotes a storage device that stores separately stored standard stroke density functions for each category to be identified.

Reference numeral 11 denotes an arithmetic unit which calculates the square of the difference between corresponding elements on the coordinate axes for the stroke density function held in the register 9 and the standard stroke density function stored in the storage device 10.

Reference numeral 12 denotes an arithmetic unit that adds the square of the difference calculated by the arithmetic unit 11 and the value of the temporary register 13, and stores the result in the register 13 again.

The arithmetic unit 12 repeats the operation by the number of dimensions of the stroke density function.

14 is a storage device prepared for each target category;
This storage device stores the matching results between the stroke density function stored in the register 9 and the characteristics for each category stored in the storage device 10.

To operate this, first store the character data in the form of binary values (0, 1) in register 1 shown in Figure 1.7 Next, start from the left side of the character, that is, the smaller X-axis value, and start from the same - X-axis coordinate. For example, for the data of a character intercept of a vertical single-row valve with
The bit extraction circuit 6 shown in FIG. 2 observes two bits at a time while shifting one bit at a time starting from the smaller Y-axis value.

Then, the pattern extracting device 7 and counting device 8 shown in FIG.

Next, the number of character strokes is counted in the same manner for a vertical line of character segment data located one bit to the right on the X-axis.

This operation is repeated to count the number of strokes for all character screens.3 As a result, character features are extracted in the register 4, with the X axis as a coordinate variable and the number of strokes as a function value.

That is, this feature indicates how many strokes the recognized character had within the character frame using the X-axis coordinate value as a parameter.

From now on, this feature will be referred to as a stroke density function in the X-axis direction.

A similar process is performed using the Y-axis direction as a coordinate variable, and a stroke density function in the Y-axis direction is obtained in the register 5.

Identification clears device 13 to 0 and 25 for each category.
Differences are calculated for the six-dimensional stroke density functions and stored in the storage device 14.

The category with the smallest value among the results may be used as the identification result.

The feature that is extracted and used by this device, that is, the stroke density function, reflects the number of strokes of a character, and it can be used for many types of characters, such as kanji figures, which have many types, from simple characters with a few strokes to complex characters with over 20 strokes. Recognition objects can be effectively classified according to their complexity.

That is, unlike similar conventional methods, in the present feature extraction device, the type of feature does not matter, and the number of dimensions of a feature is determined by the resolution for representing characters and the resolution for finding features.

When the target character is complex, the number of feature dimensions increases in the conventional method, but the present invention has the advantage that such a problem does not occur.

Furthermore, since features are extracted and compared along the set coordinate axes, there is an advantage that the appearance position of the features is reflected in character identification.

Furthermore, since the features are compared by comparing the differences between elements, there is no drawback that comparisons cannot be made if the types of features differ as in the conventional method.

For example, an example of a character with a relatively simple shape among commonly used kanji is shown in FIG.

Stroke density functions in the X-axis direction and the Y-axis direction are determined for these characters.

Next, strokes are arranged at equal intervals in the X-axis direction and the Y-axis direction within the character frame according to the number of strokes indicated by the stroke density function.

In other words, if the above-mentioned "number of strokes indicated by the stroke density function" is "1" corresponding to the X coordinate position a, then the center of the straight line given by the function x = a within the quadrilateral circumscribing the character. Draw a dot at the point (that is, the point divided into 1/2).

If we assume that the value is ``2'' corresponding to the x-coordinate position, then dots will be placed at two points that divide the straight line given by the function x=b into 1/3 each. I will continue to do so.

If the above-mentioned "number of strokes indicated by the stroke density function" is "1" corresponding to the Y coordinate position m, which is the same, then the line given by the function y = m within the quadrilateral circumscribing the character. Draw a dot at the point where it was divided into 1/2.

Also, if the value is "2" corresponding to the Y coordinate position n, then 1/3 on the straight line given by the function y = n
Draw a dot at each of the two divided points.

The figures shown in Figure 5 are obtained by performing such processing, and each character shown in Figure 4 has its own unique figure (
character) is obtained.

This can be considered to be the result of feature description for the Kanji figure using the stroke density function.

Next, Figure 6 shows an example of a character with the most complex shape among the commonly used kanji.

Figures (characters) similarly synthesized from the stroke density functions for these characters are shown in FIG.

It can be seen from both the examples in FIGS. 5 and 7 that the stroke density function accurately expresses the degree of complexity of the stroke structure of the recognized character.

Using the stroke density function obtained by this device, it is possible to construct a major classification device for commonly used kanji.

For example, regarding the character data of 2074 commonly used kanji characters written by 10 people, 5 of them are used as a standard pattern, and the average of the stroke density functions is calculated.

Stroke density functions are determined for the other five characters as unknown input kanji, and classification is performed based on the Euclidean distance from the average stroke density function determined above.

As a result, the classification rate assumes that if the correct answer appears within a certain rank, it is recognized correctly, and 95φ is 1 for unknown input kanji.
The correct answer was within 0th place, and it was effectively classified.

This major classification device has a processing time of 3 seconds/character and an average pattern capacity of 133 KB, which was determined by software simulation using a minicomputer.As a device of this kind,
Sufficiently faster speeds than expected from conventional equipment,
It can be said that we were able to achieve miniaturization.

[Brief explanation of the drawing]

FIG. 1 is an embodiment conceptually showing the identification feature extraction device according to the present invention, FIG. 2 is an embodiment showing the configuration of register 2 or register 3 shown in FIG. Block diagrams showing the matching method, Figures 4 to 7
The figure shows a detailed illustration of the invention. 191. ...128xl 28-bit recognized character storage register, 2...Character line number counting device, 3...
... Character line number counting device, 4 ... Stroke density function storage device, 5 ... Stroke density function storage device, 6 ... Bit extraction device, 7 ...・・・
...Pattern extraction device, 8...Counting device, 9.
...Stroke density function storage device, 10...
...Standard stroke density function storage device, 11...
・Difference, square calculation device, 12... Addition device, 13
. . . Temporary storage device, 14 . . . Matching result storage device.

Claims (1)

[Claims] 1. In a discriminating feature extraction device in a figure recognition system that extracts features by scanning figures including kanji, a two-dimensional figure storage device stores given figure information, and at least two a counting device that counts the number of intersections between a cutting line that cuts the given figure according to coordinate values on two coordinate axes and a figure stroke of the given figure; A storage device corresponding to the coordinate axes is provided, and features of the given figure are extracted based on the contents of the storage device, and the characteristics are identified by comparison at corresponding positions on the coordinate axes. , a patent characterized in that the difference is determined based on the difference in the number of stroke intersections, and identification is performed by summing up the differences at each point on the coordinate axis.