JPH0326877B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a pattern recognition device that improves recognition speed.

Pattern Recognition Device For example, a character recognition device usually has a feature dictionary containing features extracted from standard character patterns for each category (type of character), and selects features that have the highest degree of similarity to the features extracted from the input character pattern, that is, the character pattern to be recognized. A category from which the character pattern can be obtained is selected from the feature dictionary, and this is determined as the category of the character pattern to be recognized.

Furthermore, as the above-mentioned characteristics, for example, in Chinese characters, a plurality of types of characteristics are used, such as the shape of projection, the number of closed loops, and the number of separation strokes.

By the way, in order to select the category with the highest degree of similarity, if we calculate the degree of similarity between the features extracted from the character pattern to be recognized and the features for each category stored in the feature dictionary, for example, the feature There are 10 types and categories.
In the case of 2,000 species, it is necessary to perform similarity calculations 20,000 times.

Therefore, in order to reduce the number of similarity calculations,
Recognition is divided into two stages: primary recognition and secondary recognition. In the primary recognition, multiple categories with high similarity are selected as candidates through rough recognition using only some features, and the secondary recognition For recognition, a pattern recognition device has been proposed that selects the category with the highest degree of similarity to the input pattern from among multiple categories selected in primary recognition through precise recognition using all features. There is.

At this time, it is desired to further improve recognition speed by efficiently selecting a plurality of categories in primary recognition.

[Conventional technology]

FIG. 2 shows a conventional example of a pattern recognition device that aims to reduce the number of similarity calculations between features extracted from an input pattern and features stored in a feature dictionary.

In the figure, 1 is a feature dictionary that stores a plurality of predetermined features extracted from standard character patterns for each category as a feature dictionary, and 2 is a feature extraction unit that extracts the same plurality of predetermined features from an input pattern. , 3 are a predetermined part of the features of the input pattern extracted by the feature extractor 2 and the same part of the standard features for each category stored in the feature dictionary 1. 4 is a selection unit that selects categories up to a predetermined top rank in the order of similarity obtained by the first recognition unit 3; 5 is a selection unit 4 extracts and stores the standard features of each category selected from the feature dictionary 1; 6 extracts and stores the standard features of each category selected by the feature dictionary 1; This is a second recognition unit (secondary recognition) that calculates the degree of similarity with the standard feature for each category of numbers, and determines the category with the highest degree of similarity as the category of the input pattern.

In the above configuration, for example, the number of categories stored in the feature dictionary 1 is 2000, and the types of standard features for each category stored in the feature dictionary 1, that is, the types of features extracted in the feature extraction unit 2, are 10 types. , the number of features used in the first recognition unit 3 is 1 type, the number of features used in the second recognition unit 6 is 10 types, and the categories selected in the selection unit 4 are the top 20, then the first recognition Part 3 calculates the similarity 2000 times, and the second recognition part 6 calculates the similarity (20×10) times, for a total of 2200 times, compared to the 20000 times described above. It can be seen that the number of calculations can be significantly reduced.

[Problem that the invention seeks to solve]

However, in the configuration described above, the selection unit 4 needs to select the top 20 by sorting processing etc. from among the similarities determined by the first recognition unit for 2000 categories. There is a problem that the processing for this requires a long time.

[Means for solving problems]

As mentioned above, the problem is that the present invention stores a first feature dictionary that stores standard features for each category, a category group that includes each category and categories similar to it, and the average value of the features of this category group. a second feature dictionary that extracts the features of the input pattern; and a feature extractor that extracts the features of the input pattern; a first recognition unit that determines the category group to which the pattern belongs; and a category determined by the first recognition unit among the extracted features of the input pattern and standard features stored in the first feature dictionary. The problem is solved by a pattern recognition device characterized by comprising: a second recognition unit that determines the category of the input pattern based on the degree of similarity with the standard feature of the category included in the group.

[Effect]

That is, in the present invention, in addition to a first feature dictionary that stores standard features for each category, a second feature dictionary that stores categories similar to each category and average features for each category group is provided for each category. For example, by determining the category group to which the input pattern belongs based on the degree of similarity between the features of the input pattern and the average feature stored in the second feature dictionary, the category group is stored in the second feature dictionary. Assuming that the number of categories in each category group is 10, the top two categories that are similar to the input pattern are selected from the similarity determined by the first recognition unit. Up to about 20 categories can be selected, and the time required for sorting can be significantly reduced compared to the conventional example.

〔Example〕

Next, the gist of the present invention will be specifically explained using an embodiment shown in FIG.

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention. Reference numerals common to those in FIG. 2 refer to the same objects, and 7 indicates a category similar to each category and a group of these categories. This is a second feature dictionary that stores average features for each.

Further, 3' is a first part that determines the category group to which the input pattern belongs based on the degree of similarity between the features of the input pattern extracted by the feature extractor 2 and the average features stored in the second feature dictionary. 4' is a selection unit that selects categories up to a predetermined top rank in the order of similarity obtained by the first recognition unit 3', and 5' is a selection unit that selects categories selected by the selection unit 4'. This is a buffer that extracts standard features of each category included in a category group from the first feature dictionary 1 and stores them.

To further explain the second feature dictionary, when the categories to be recognized are C ₁ , C ₂ , . . . C _o , similar categories to that category are checked for each category. That is, k similar categories of category Ci are
Let Ci(1), Ci(2), . . . Ci(k) be determined for all i.

Next, the average value of the features of category Ci and categories Ci(1), Ci(2), . . . Ci(k) is used in the major classification dictionary for category Ci.

This corresponds to creating a dictionary of category Gi, which is a temporary category that is the average of all Ci and its similar categories, for example, category Gi.

Such a major classification dictionary is shown in FIG. This major classification dictionary is the second feature dictionary in the present invention.

In the above configuration, the second feature dictionary 7
In addition to setting the number of categories stored in each category group to 10, as in the above embodiment, the number of categories stored in the same feature dictionary 1 is set to 2000 types, and the number of categories stored in the first feature dictionary 1 is set to 10. The types of standard features, that is, the types of features extracted in the feature extraction unit 2 are 10 types, the number of features used in the first recognition unit 3' is 1 type, and the number of features used in the second recognition unit 6 is 10 types. The first recognition unit 3' calculates the similarity 2000 times in the same way as in the conventional example, and as a result, in the example of Fig. 3, the category Ci, that is,
Once a Gi dictionary is selected, it is a total of (k+1) similar categories and Cis that were involved in creating the Gi dictionary,
In other words, 10 categories have been selected. This shows an effect equivalent to selecting 10 categories by selecting one major classification dictionary. By selecting the top two categories from the Gi dictionary selected in this way, it is possible to select the top 20 categories that are similar to the input pattern. The category of the input pattern can be determined by performing similarity calculations approximately (20×10) times in the recognition unit 6.

That is, although the number of similarity calculations is approximately the same as in the conventional example, the time required for sorting in the selection section 4' can be significantly reduced compared to the conventional example.

〔Effect of the invention〕

As described above, according to the present invention, the recognition speed of a pattern recognition device can be significantly improved.

[Brief explanation of drawings]

FIG. 1 shows a configuration diagram of an embodiment of the present invention, and FIG. 2 shows a configuration diagram of a conventional example. FIG. 3 shows a second feature dictionary in the present invention. In the figure, 1 is the first feature dictionary, 2 is the feature extraction unit,
3' is a first recognition section, 6 is a second recognition section, and 7 is a second feature dictionary.

Claims (1)

[Claims] 1. A first feature dictionary that stores standard features for each category; a second feature dictionary that stores, for each category, a category group including each category and categories similar thereto, and the average value of the features of this category group; a feature extraction unit that extracts features of the input pattern, and a category group to which the input pattern belongs is determined based on the degree of similarity between the extracted features of the input pattern and an average value of features stored in a second feature dictionary. a first recognition unit, and standard features of categories included in the category group determined by the first recognition unit among the standard features stored in the extracted input pattern features and the first feature dictionary; a second recognition unit that determines a category of the input pattern based on the degree of similarity between the input pattern and the input pattern.