JPS63100586A - Pattern recognizing method - Google Patents

Abstract

PURPOSE:To eliminate an erroneous recognition caused by an erroneous designation, by selecting automatically one of the standard patterns which are used thereafter, based on a selection reference containing the shortest distance appeared in each standard pattern group while recognizing a pattern by using all of plural kinds of standard pattern groups during registration. CONSTITUTION:A triangular mark is a feature vector generated from the input pattern of an alphanumeric character, and circular marks attached with alphanumeric characters of 'P', 'F', 'B', and '?' are feature vectors of the standard patterns of the corresponding alphanumeric characters. In the feature vector space, distances between the input pattern and the standard patterns 'P', 'F', 'B', and '?' are d1, d2, d3 and d4 in the same order, and distances to the remaining all standard patterns become larger values than d4. Accordingly, the input pattern is recognized to be the most adjacent standard pattern 'P' having the shortest distance to the input pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a pattern recognition method used in an optical character reader (OCR) in an information processing system.

BACKGROUND OF THE INVENTION Optical character reading devices used in conventional technical information processing systems generally perform character recognition using a pattern matching method.

This character recognition method using the pattern matching method creates a feature vector from the input pattern to be recognized, compares it with all feature vectors of the standard pattern being registered, and selects the nearest neighbor with the shortest distance in the feature vector space. The standard pattern is recognized as an input pattern.

Input patterns have various fonts (fonts), so in order to increase the recognition rate, it is recommended to include various fonts in each standard pattern or prepare a group of standard patterns for each font. When inputting a character pattern to be recognized, the user is asked to specify a standard pattern group corresponding to the font.

Problems to be Solved by the Invention Among the conventional character recognition methods mentioned above, the method of including various fonts in each standard pattern has the problem that the number of standard patterns to be compared increases and the recognition speed decreases. There is.

In addition, using a method that prepares a standard pattern group for each font and then specifies the use of the standard pattern group according to the font at the start of recognition not only requires the user to have correct knowledge of the font, but also prevents misunderstandings. There is a problem that the recognition rate decreases due to incorrect specification based on the method.

Structure of the Invention Means for Solving the Problems According to the character recognition method of the present invention which solves the problems of the prior art described above, a plurality of standard pattern groups are registered for each font or the like.

In addition, for the appropriate number of input patterns to be recognized initially, pattern recognition is performed using all of the multiple types of standard pattern groups being registered, and during this recognition, the shortest distance that appears in each standard pattern group is included. One of the standard patterns is selected based on selection criteria.

Then, for subsequent input patterns, pattern recognition is continued using one of the standard patterns selected above, thereby effectively preventing a decrease in recognition speed or a decrease in recognition rate due to incorrect specification. has been done.

Hereinafter, the operation of the present invention will be explained in detail together with examples.

Embodiment FIG. 1 is a conceptual diagram for explaining distance in a feature vector space in an embodiment of the present invention.

In Figure 1, the triangle marks are feature vectors created from the alphanumeric input pattern to be recognized, rPJ, rFJ, r
BJ, r? A circle marked with an alphanumeric character J is a feature vector of a standard pattern of corresponding alphanumeric characters.

In the above feature vector space, input pattern and standard pattern rPJ, rFJ, rBJ, r? The distance to J is dIn dt, "In a4" around the same time, and the distance to all remaining standard patterns (not shown) is larger than d4. Therefore, the input pattern has the shortest distance between it and this. It is recognized that the standard pattern rPJ is the nearest neighbor.

In this example, the alphanumeric standard pattern groups are BOLDFACE, LETTERGOTHIC,
It is registered for each of eight types of fonts, including PRESTIGEELITE. Immediately after the start of recognition, for an appropriate number of input patterns, for example 7, the standard pattern group is compared with the input pattern for all 8 types of fonts being registered, and the shortest distance d1 is divided by the next shortest distance d8. value d
A calculation of +/dz is performed and the minimum d.

The recognition result obtained for the font in which the value /a appears is taken as the overall recognition result.

Similarly, for the second input pattern, for all eight types of fonts being registered, the input pattern is compared with the standard pattern group, and the shortest path N a + is determined by the next shortest distance d! The value d+/dz divided by is calculated. and,
a obtained for the first input pattern for each font
By adding the d+/dx value obtained for the current input pattern to t/dg M, d for each font is
The total sum of +/dt values is calculated, and the recognition result obtained for the font in which the minimum sum of ds/dz values appears is taken as the overall recognition result.

Similarly, for the third input pattern, for all eight types of fonts being registered, the standard pattern group is compared with the input pattern, and the at/ax value is calculated by dividing the shortest path jiI d+ by the next shortest distance d8. will be calculated. Then, by adding the d+/dz value obtained for the current input pattern to the sum of the dt/ax values for each font obtained for the second input pattern, a new sum of d+/dx values for each font is obtained. is calculated, and the minimum d+
The recognition result obtained for the font in which the sum of /di values appears is taken as the overall recognition result.

Thereafter, the same recognition and calculation of the sum of d+/di values as described above are repeated until the seventh input pattern. At this point, the font of the standard pattern group in which the minimum sum of d+/dt values appears is determined to be the font of the input pattern to be recognized. Then, for the eighth and subsequent input patterns, matching is performed only with the standard pattern group of the font determined and selected as described above.

Even after character L2 recognition using the standard pattern selected as described above has started, the sum of d+/di values for the latest eight character patterns, or the average value of d+/dg values for each character pattern. continues to be calculated and updated, and when this exceeds a predetermined value, it is assumed that a font change has occurred during the process, and the state returns to the state immediately after the start of recognition. In other words, the subsequent eight input patterns are compared with the standard pattern group of all fonts, and the minimum d1/
The standard pattern group of fonts in which the sum of the values d appears is selected as a new matching target.

FIG. 2 to FIG. 4 show experimental results of the pattern recognition method of the present invention.

The experimental results shown in FIG.
Enter six alphabetic characters in sequence, and d displays the alphanumeric characters recognized from the shortest distance for each input pattern and font, and a decimal number with two decimal places.

/a The Z value is calculated.

For example, if the input pattern is the alphabet rAJ, font 1
Then, the recognition result based on the shortest distance in the same space with the feature vector registered as the standard pattern of this font 1 is rAJ, d1/dx vector is 0.77, and the recognition result of font 2 is registered as the standard pattern of this font 2. The recognition result based on the shortest distance in the same space with the feature vector is rAJ, the d+/dz value is 0.96, and the similar recognition result for font 3 is "8J, the d+/dz value is 0.28, etc." In addition, "Minimum value" in the rightmost column indicates the number of the font in which the minimum d+/di value appears for each input pattern. For example, if the input pattern is the alphabet rAJ, the minimum d+/dt value (0,
The number of font 3 in which 28) appears is written.

The "additional value" in the bottom column of Figure 2 is for each font.
d+/d obtained for 26 individual letters from to Z
This is the added value of the i value. This value is the minimum value (10, 69) for font 3, the next smallest value (17, 48) for font 6, and is a value of 20 or more for all the other six fonts.

Figure 3 shows the accuracy rate of whether the determined font matches the actual font of the input pattern for all fonts in relation to the number of additions of the sum of d+/dz values (number of input patterns). These are experimental results.

If the number of additions is 1, that is, the font is determined only from the minimum d+/dz value for one input pattern, the accuracy rate is 71.8%. Furthermore, as the number of times the total is added before making a decision, that is, the number of input patterns increases to 3.5.7, the accuracy rate of the decision increases to 92.9%, 97.
It gradually approaches 100% at 9%, 99.6%, and so on.

FIG. 4 shows the results of an experiment in which the accuracy rate of font determination for various selection criteria was compared when the number of additions was one.

In the figure, t'at □　　　　　　　　　　　　　　 It is.

Also, the selection criteria d, /d2&d, are d+/dt and dl
This is a case where a font is selected under the condition that both of the fonts are minimum.If this condition is not met, such as when only one of the fonts is minimum, it is rejected as it cannot be determined whether it is correct or incorrect.

As shown in the figure, if the font is determined based on the minimum d+/di, an accuracy rate of 73.6% is obtained, and if the font is determined based on the minimum dt/d- or d+/dis, both It can be seen that an accuracy rate of 72.5% is obtained. Also, (ax −ti, )/d1. When the selection criterion was the minimum, the accuracy rate was 71.4%, and when the selection criterion was the minimum d, the accuracy rate was 69.1%.

The above is a case where the number of additions is one, but the accuracy rate can be increased to a practical value by increasing the number of additions for any selection criteria including the shortest distance d mentioned above.

The above example illustrates the case where the recognition target is a character pattern, but it may also be a voice pattern, etc. In this case, instead of the font for the character pattern, a group of standard patterns such as by gender, age group, region of use, etc. be registered.

Effects of the Invention As explained in detail above, the pattern recognition method of the present invention performs pattern recognition using all of the registered standard pattern groups for an appropriate number of input patterns that are initially recognized. During recognition, one of the standard patterns to be used thereafter is automatically selected based on selection criteria including the shortest distance that appears in each standard pattern group, so all standard pattern groups can be matched from beginning to end. Unlike when the pattern is used as a target, the recognition speed is improved, and unlike when the user is required to specify a group of standard patterns to be used, there is no risk of misrecognition due to incorrect specification.

[Brief explanation of the drawing]

FIG. 1 is a conceptual diagram for explaining distances in a feature vector space according to an embodiment of the present invention, and FIGS. 2 to 4 are conceptual diagrams showing experimental results according to the present invention. Patent applicant: NEC Home Electronics Co., Ltd.

Claims (4)

[Claims] (1) In a pattern matching pattern recognition method that performs pattern recognition based on the distance in the feature vector space between the input pattern to be recognized and the standard pattern being registered, multiple types of standard pattern groups are registered. Then, for the appropriate number of input patterns to be recognized first, pattern recognition is performed using all of the multiple types of standard pattern groups being registered, and during this recognition, the selection including the shortest distance that appears in each standard pattern group is performed. A pattern recognition method characterized in that one of the standard patterns is selected based on a criterion, and pattern recognition is continued using one of the selected standard patterns for subsequent input patterns. (2) The selection criterion including the shortest distance is a value obtained by dividing the shortest distance by the next shortest distance, and the standard pattern with the smallest value is selected. Pattern recognition method. (3) A patent claim characterized in that the selection criterion including the shortest distance is a value obtained by dividing the shortest distance by the average value of a plurality of distances below this, and the standard pattern with the smallest value is selected. The pattern recognition method according to item 1. (4) A patent characterized in that the selection criterion including the shortest distance is the shortest distance divided by the average value of the distances from the second to the second, and the standard pattern with the smallest value is selected. A pattern recognition method according to claim 1.