JPS61246885A - Plural dictionary producing system - Google Patents

Abstract

PURPOSE:To avoid a learning pattern from being erroneously classified by producing plural dictionaries by taking into account the relationship between areas. CONSTITUTION:Areas U1 and U2 are ones for producing the dictionaries of categories alpha and beta, respectively. An area U4 is the one where no dictionary is produced. An area U5 other than the area U1, U2 and U4 is the one where the dictionary of any category can be produced. In this case, where the learning pattern beta* belonging to the category beta arises, the dictionary of the category betahas to be produced within a circle U3 with a radius alpha. However it can be produced in a blank area U5 or in the area U2 other than the overlapping area U4.

Description

[Detailed Description of the Invention] [Summary] The present invention relates to a multiple dictionary creation method, and when creating a new dictionary, learning patterns that would have been correctly classified if only the dictionary was used or new patterns of other existing categories are used. This eliminates the negative effects of misclassification by dictionaries and enables correct classification of all learning patterns.

[Industrial application field]

The present invention relates to a multiple dictionary creation method. In general, each character written by humans has its own unique characteristics. This is no exception, even when it comes to simple numbers. For example, if you write the number ``8'', there is no problem if it is written as ``8'' correctly, but it may look like ``3'' or cannot be recognized at all, and even if the machine can recognize ``8'', it may not be a problem. In order to make a determination, several dictionaries are required for comparison.

Therefore, normally, multiple dictionaries are created for several input learning patterns for one category, and when comparing actual human-powered patterns, these multiple dictionaries are used as the target for matching. becomes.

The present invention relates to a method for creating such multiple dictionaries.

[Prior Art and Problems to be Solved by the Invention] As shown in FIG. 3, the conventional multiple dictionary creation system does not take into account the relationship between different characters in the dictionary.

In Figure 3, when creating multiple dictionaries for a certain pattern, for example the numbers "8" and "9", first take the average M of several learning patterns input for "8" and Create one dictionary and divide it into three parts, for example, to create three dictionaries for "8".
, Me'', M% are created.

Next, for "9", three plural dictionaries M9'l Me"l Me"l are similarly created from one dictionary M9.

Using the plurality of dictionaries created in this way, the input pattern is compared with the input pattern.

However, conventionally, as mentioned above, multiple dictionaries related to "8" M s ' + M s '' + M s 3 and "9"
The relationships between the multiple dictionaries Mq', Mq'', M,'' were created completely independently.

Therefore, when the learning pattern P is "8", although it should originally be recognized by the dictionary M81,
Since the distance d is smaller than the distance d, the dictionary M g'
Therefore, it is recognized as "9".

That is, in the past, there was a problem that learning patterns were not classified correctly.

[Means and operations for solving the problem] The present invention solves the above problem and eliminates misclassification.The method is to first assume a plurality of temporary dictionary creation areas for each category, and then perform the following steps. This method consists of determining these overlapping areas as areas for which no dictionary can be created, and further preventing misclassification due to the relationship with the previous temporary area. In this way, the learned pattern will not be misclassified.

〔Example〕

Hereinafter, the present invention will be explained by way of examples with reference to the accompanying drawings.

FIG. 1 is an explanatory diagram of the system of the present invention.

Now, let's assume that the categories in the dictionary are α, β, for example the number r8.
J, r9J, and the learning patterns used to create these dictionaries are αhα2. α31...α7cα, β1. β2.
β1. ...α, 6β.

In the figure, ■ and ■ are singular dictionaries.

Among the above learning patterns, α1. α2
．． Assume that only α3 is recognized.

Here, if we focus on the learning pattern α1, the fact that α1 was recognized by the dictionary @ means that α1 is the center and 1 is the radius r,
This means that there is no other dictionary of other categories in circle U.

Similarly, learning pattern α2. Regarding α1, the radius rg,
There are no other dictionaries of other categories in the circles U+z and U13 of r3.

Therefore, for the correctly classified learning pattern αh α2゜α, the circle Ull, UI! + There are no dictionaries other than ■ in the area U shown by the solid line surrounded by UI3.

That is, if an attempt is made to create a new dictionary within the solid line area U, only a dictionary for category α can be created.

If a dictionary for other categories is created, it will be misclassified by the learning pattern of category α, which was correctly classified up to that point, or by the dictionary of other categories (■ in Fig. 2).

If we apply the same idea to the category (β),
Regarding β, the area U2 surrounded by the dashed-dotted line becomes the category dictionary creation area (■ in FIG. 2).

In addition, the shaded area U4 is an overlapping area between the dictionary creation area for category α and that for β (second
■) in the figure.

If a learning pattern of category β is included in this area U4, it will be recognized as α because the distance to ■ is closer than ■. In the opposite case, α is recognized as β. Therefore, the above U is an area in which no dictionary is created.

Furthermore, the blank area U other than U+, Uz, and U4 mentioned above is an area in which a dictionary of any category can be created (■ in FIG. 2).

Although the dictionary creation area is determined in this way, the above conditions are not necessarily sufficient conditions.

That is, although the dictionary area U1 related to the category α has been determined, the relationship with the area U2 of the other category β is not taken into consideration.

For example, learning pattern β1 belonging to category β may be misclassified by dictionary α. This is because there are no dictionaries other than @ within the radius lfI''-■l=d centered on β''.

Therefore, it is necessary to create a dictionary for category β within this circle U of radius d (■ in Fig. 2).

Moreover, this dictionary for β must be created within the area U2, outside the overlapping area U4, or in the blank area U (■ in FIG. 2).

The flowchart in FIG. 2 shows the flow of the operations up to the above.

〔Effect of the invention〕

As described above, according to the present invention, a temporary multiple dictionary creation area is first assumed for each category, then these overlapping areas are determined as areas in which no dictionary can be created, and furthermore, the overlapping areas with the previous temporary area are determined as areas in which no dictionary can be created. Since it is designed to prevent misclassification based on relationships, the relationships between each area are also fully considered, and learning patterns are no longer misclassified as in the past.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of the method of the present invention, FIG. 2 is a flowchart showing an embodiment of the present invention, and FIG. 3 is an explanatory diagram of the conventional method. U+, Uz, Uz, Us...Dictionary creation area, U
4... Dictionary creation prohibited area, r, + rz l r... radius, ■, ■...
Dictionary of categories α and β, α! , α2. α3...Learning pattern of category α, β1
．． β2. β3: Learning pattern for category β.

Claims (1)

[Claims] In a method of creating a plurality of dictionaries for each category in order to classify patterns correctly, first, representative values of learning patterns belonging to the first category to the Nth category are temporarily used in the dictionary for each category. Then, the area where all the learning patterns belonging to each of the above categories are independently recognized is set as a temporary multiple dictionary creation area, and then the above temporary second category multiple dictionary creation area and the previous temporary first category multiple are set as a temporary multiple dictionary creation area. The overlapping area with the dictionary creation area is determined as a common dictionary creation prohibited area, and furthermore, a distance smaller than the distance between the learning pattern to be correctly classified by the temporary second category dictionary and the temporary first category dictionary is determined. The area where the area with radius vector and the temporary second category multiple dictionary creation area overlap is determined as the true second category multiple dictionary creation area, and the same operation is performed sequentially from the third area to the Nth area. A multiple dictionary creation method characterized by determining in advance a true multiple dictionary creation area for each category by repeating the process, and increasing the number of dictionaries within each area.