JP2001034709A - Fast recognition and retrieval system, method for accelerating recognition and retrieval used for the same and recording medium with its control program recorded thereon - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a fast recognition and retrieval system capable of fast executing retrieval in a stable required time without accompanying backtrack. SOLUTION: The subset generating part 13 of a learning processing means 1 duplicates a template existing while spreading over an identification boundary for performing separation to a lower subset and registers the template in a slave node. A hierarchical structure storing part 14 stores characteristics degenerated to a leaf node. An identification processing means 2 reads the contribution ratio of each characteristic quantity from leaf node and performs classification by using only a degenerated characteristic. Because stable and also fast retrieval can be performed without the redundant backtrack of a decision tree in this way, the prior probability of a category is calculated and a frequent character can be determined in a shallow hierarchy of the decision tree, it is possible to further realize the improvement of a retrieval speed.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-speed recognition / retrieval system, a high-speed recognition / retrieval method used therefor, and a recording medium on which a control program is recorded. The present invention relates to a method for speeding up a search process.

[0002]

2. Description of the Related Art Conventionally, a decision tree recognition system includes a feature extraction unit, a sample dictionary unit, a discrete decision tree generation unit, a decision tree storage unit, and a backtrack decision tree search unit. I have.

In the above-described decision tree recognition system, when a character pattern is input, a feature vector is generated by a feature extracting unit. The decision tree generated in advance by the discrete decision tree generation unit is stored in the decision tree storage unit.

[0004] The feature vector is identified according to the conditions stored in each node of the decision tree, and subsequent identification is performed after a child node is selected according to the identification result. Similar classification is repeatedly performed until the terminal node is reached.

[0005] At the end node, the final collation with the sample dictionary is performed. If this collation cannot be performed satisfactorily, the backtracking decision tree search unit searches the dictionary by going back to a higher hierarchy. The above-described recognition system using a decision tree is disclosed in Japanese Patent Application Laid-Open No. 6-282687.

[0006]

In the above-described conventional decision tree recognition system, it is generally difficult to optimally implement the backtrack as in the above method, and the actual search efficiency is not improved. There are many. In the worst case, the search may be deteriorated to the same extent as the full search, and the search time may become unstable depending on the search pattern, so that the search efficiency cannot always be improved.

In the construction of the hierarchical structure, a deterministic and uniform criterion such as, for example, evaluation based on a Euclidean distance from a specific position in the feature space is provided. Since it is impossible to modify the structure of a specific decision tree according to its properties, it is usually necessary to prepare a framework that outputs frequently searched elements at high speed in order to improve perceived search efficiency. However, this is not possible with the above method.

SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to solve the above-mentioned problems and to realize a high-speed recognition search system capable of executing a high-speed search with a stable required time without a backtrack, a recognition search speed-up method used therefor, and control thereof It is to provide a recording medium on which a program is recorded.

Another object of the present invention is to provide a high-speed recognition search system capable of optimizing the structure of a decision tree in accordance with the appearance probability of an object and improving perceptual search efficiency, and a method for speeding up a recognition search used therefor. Another object of the present invention is to provide a recording medium recording the control program.

[0010]

SUMMARY OF THE INVENTION A high-speed recognition and retrieval system according to the present invention generates a feature vector from an input character pattern, and stores the feature vector in accordance with a condition stored in each node of a previously generated decision tree. A high-speed recognition and retrieval system for identifying, sequentially selecting child nodes according to the identification result, and repeatedly performing this classification until reaching a terminal node, wherein the recognition is performed from a set of patterns to which a predetermined correct category is assigned. Generating means for generating a template of a multidimensional feature vector stored in a dictionary, template dictionary storing means for storing the template created by the generating means and a pattern contributing to the generation of the template in association with each other, Template and the set of patterns corresponding to each template A subset generation unit that classifies the appearance frequency of the category into a subset and outputs a template belonging to the subset and a threshold value for performing separation into the subset, and the subset generation unit sequentially generates the subset. Hierarchical dictionary means for storing a subset of templates in association with a corresponding subset of pre-separated templates; and inputting the hierarchical structure stored in the hierarchical dictionary storage means in order from the upper level to classify and classify the input pattern. Decision tree classifying means for outputting a child node as a result of the classification, and category determining means for reading out feature amounts effective for determining a template from the leaf nodes of the hierarchical structure and performing a large classification using the feature amounts And

According to the method for speeding up recognition and retrieval according to the present invention, a feature vector is generated from an input character pattern, and the feature vector is identified according to a condition stored in each node of a previously generated decision tree. It is a method for speeding up the recognition and retrieval of a high-speed recognition and retrieval system in which child nodes are sequentially selected according to the result and this classification is repeatedly performed until reaching a terminal node, and is performed from a set of patterns to which a preset correct category is added. Generating a template of a multi-dimensional feature vector stored in the recognition dictionary; classifying a set of a template currently focused on and a pattern corresponding to each of the templates and an appearance frequency of the correct answer category into subsets; Outputs a template belonging to the subset and a threshold for performing separation into subsets Inputting the steps and the hierarchical structure stored in the hierarchical dictionary means for storing the sequentially generated template subsets in association with the corresponding subsets of the template before separation in order from the upper layer, and classifying the input pattern; Outputting a child node as a result of the processing, and a step of reading out feature amounts effective for determining a template from the leaf nodes of the hierarchical structure and performing a large classification using the feature amounts. .

[0012] The recording medium storing the recognition and retrieval speed-up control program according to the present invention generates a feature vector from an input character pattern, and converts the feature vector to a condition stored in each node of a previously generated decision tree. Therefore, a storage medium storing a recognition search speed-up control program for speeding up a recognition search in a recognition search device that performs identification, sequentially selects child nodes according to the identification result, and repeats this classification until reaching a terminal node. The recognition search acceleration control program causes the recognition search device to generate a template of a multidimensional feature vector stored in a recognition dictionary from a set of patterns to which a preset correct category has been added, A set of templates currently focused on and patterns corresponding to the respective templates, and Classifying the frequency of appearance of the solution category into subsets, and outputting a template belonging to the subset and a threshold value for performing separation into subsets, and converting a subset of sequentially generated templates to the corresponding template before separation. The hierarchical structure stored in the hierarchical dictionary means, which is stored in association with the subset of, is input in order from the upper hierarchy, the input pattern is classified, and the child nodes of the classified result are output. In order to determine the effective amount, a feature amount is read out, and a large classification is performed using the effective amount.

That is, a high-speed recognition search system according to the present invention comprises: a dictionary creating section for generating a template of a multidimensional feature vector stored in a recognition dictionary from a set of patterns to which a predetermined correct category has been assigned; A template storage unit that stores the created template and the pattern that has contributed to the generation of the template in association with each other, and inputs a set of templates corresponding to the template of interest and the template, classifies them into subsets, and Outputs the template belonging to the subset and the threshold or identification boundary for separating into the subset, and also reduces the feature vector of the template belonging to the corresponding subset if the output node is a leaf node Subset that selects and outputs only the superior features that are effective for subsequent identification A learning processing unit comprising: a generation unit; and a hierarchical structure storage unit that inputs a subset of templates sequentially generated by the subset generation unit and stores the input in association with a corresponding subset of the template before separation. I have.

Further, the high-speed recognition search system of the present invention
The hierarchical structure stored in the hierarchical structure storage unit of the learning processing unit is input in order from the upper layer, the input pattern is classified, and the child nodes of the classified result are output. It has an identification processing unit composed of a pattern classification unit that ends the classification when the classification is completed up to the lower layer, and a hierarchical structure storage unit that stores the hierarchical structure.

Further, the high-speed recognition and retrieval system of the present invention has a large classification unit which reads a contribution rate of each feature amount from a leaf node having a hierarchical structure and performs a large classification using only features having a low contribution rate. It operates so that subsequent classification can be performed at high speed as the classification progresses. In this case, the subset generation unit generates the decision tree by including the categories existing over the determined identification boundary in the subsets on both sides of the threshold,
It operates so that pattern search can be performed in a stable search time without performing backtracking of redundant decision trees.

Further, the high-speed recognition and retrieval system of the present invention is characterized in that when determining an identification boundary to a subset, the prior probability of the correct category ω _j is calculated and output as a subset accordingly. Since it is designed so that frequently appearing characters can be determined at the earlier hierarchy of the decision tree, it operates to improve the perceived search speed.

[0017]

Next, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a high-speed recognition search system according to one embodiment of the present invention.
In FIG. 1, the high-speed recognition search system according to one embodiment of the present invention includes a learning processing unit 1 and an identification processing unit 2.

The learning processing means 1 comprises a template dictionary creating section 11, a template dictionary storing section 12, a subset generating section 13, and a hierarchical dictionary storing section 14. The identification processing means 2 comprises a decision tree classifying section 21, And a category determination unit 21. Note that the identification processing unit 2 includes the template dictionary storage unit 12 and the hierarchical dictionary storage unit 14 of the learning processing unit 1.

The template dictionary creating section 11 of the learning processing means 1 stores a plurality of patterns stored in the recognition dictionary from a set of patterns to which predetermined correct categories (for example, “A” for “A”) are added. A template (reference pattern) of the dimensional feature vector is generated. The template dictionary storage unit 12 stores the created template in association with the pattern that has contributed to the generation of the template.

The subset generating unit 13 receives a template of interest, a set of patterns corresponding to each template, and an appearance frequency of a correct category, classifies them into subsets, and selects a template belonging to the subset and a subset. A threshold value for performing separation into and is output. The hierarchical dictionary storage unit 14 receives the subsets of the templates sequentially generated by the subset generation unit 13 and stores them in association with the corresponding subsets of the template before separation.

The decision tree classification unit 21 of the identification processing unit 2 sequentially inputs the hierarchical structures stored in the template dictionary storage unit 12, the category determination unit 22, and the hierarchical structure storage unit 14 of the learning processing unit 1 from the upper hierarchy. The pattern is classified, and the child nodes of the classified result are output. If the classification of the pattern into the child nodes is completed to the lowest layer, the classification is ended.

The category determining unit 21 reads out effective features for determining a template from the leaf node (terminal node) of the hierarchical structure from the hierarchical structure storage unit that stores the hierarchical structure, and stores those feature amounts. Classification is performed using

FIG. 2 is a flowchart showing the processing operation of the learning processing means 1 of FIG. 1, and FIG. 3 is a flowchart showing the processing operation of the identification processing means 2 of FIG. 4 to 7
FIG. 3 is a diagram for explaining a processing operation of the high-speed recognition search system according to one embodiment of the present invention.

The overall operation of the high-speed recognition and retrieval system according to one embodiment of the present invention will be described with reference to FIGS. The processing operations shown in FIGS. 2 and 3 are realized by the learning processing unit 1 and the identification processing unit 2 executing a program in a control memory (not shown). As the control memory, a ROM (Read Only Memory) or an IC (Read Only Memory) is used. An integrated circuit) memory or the like can be used.

The learning processing means 1 generates a template from the input learning pattern with the correct answer category (step S1 in FIG. 2). The template is generated by, for example, averaging patterns (feature vectors) having the same correct answer category.

The learning processing means 1 registers a template belonging to the node of interest (step S2 in FIG. 2). In the case of the first loop, the node of interest is the root node, and the templates belonging to it are all templates.

The learning processing means 1 classifies templates belonging to the node of interest into a plurality of subsets by clustering (step S3 in FIG. 2). As a clustering method, for example, existing k-means algorithms (Tou, Gonz
alez, Pattern Recognition Principles, Addiso
n-Weisley Publishing Company, p.90).

The learning processing means 1 obtains an identification plane which is a boundary of the generated subset. Expressing the discrimination plane by simpler calculation is more effective for speeding up discrimination by the decision tree. If the identification surface is, for example, a linear hyperplane, the identification surface of the subset does not match the identification boundary of the template itself.

For this reason, in the vicinity of the identification surface of the subset, there is a possibility that the identification result of the pattern is inconsistent between the identification result of the identification surface of the subset and the identification result of each template (the hatched portion shown in FIG. 4). Area). In this case, the template contributing to the region in the feature space where the contradiction occurred is included in both of the corresponding subsets (the black circle template shown in FIG. 4).

Here, a method of obtaining the identification plane of the subset is, for example, a method in which the average of the feature vectors of the templates belonging to the subset is set as the center of the subset, and the bisector hyperplane between the centers of the subset is set as the identification plane. Can be used.

The learning processing means 1 registers the subsets of the templates generated by the clustering as lower nodes of the target node (step S4 in FIG. 2).
At the same time, the learning processing means 1 registers the identification surface generated in step S3 in association with the node of interest. The condition that the registered node is set as a leaf node can be set, for example, such as “if the number of templates belonging to the node decreases until it becomes less than a specified value”.

The learning processing means 1 determines whether or not the registered node satisfies the condition of a leaf node. If it is determined that the registered node is not a leaf node, the process proceeds to step S6.
If it is determined that the node is a leaf node, step S7
(Step S5 in FIG. 2).

The learning processing means 1 stores the registered child nodes,
Further, it is updated as a node of interest divided into child nodes, and the above-described processing of step S2 and subsequent steps are recursively applied (step S6 in FIG. 2).

The learning processing means 1 selects, for the templates belonging to the nodes registered as leaf nodes, features effective for classifying the templates. As a selection method, for example, the first to n-th principal components when the principal component analysis of the feature vector of the corresponding template is performed. In this way, the selected feature component is stored in association with the leaf node (step S7 in FIG. 2).

When the registration of all the leaf nodes is completed, the learning processing means 1 terminates the above operation, and when not completed, proceeds to the processing of step S9 (step S8 in FIG. 2).
That is, the learning processing means 1 searches for a node for which a leaf node has not been registered, moves the node of interest to that node, and continues the processing from step S2 onward (step S9 in FIG. 2).

On the other hand, the identification processing means 2 compares the feature vector of the input pattern with the identification plane stored in the target node and determines which of the lower nodes is to be classified (step S11 in FIG. 3). As described above, when an identification plane that is a linear hyperplane is used for the identification method, the identification is performed depending on which side of the identification plane the input feature vector exists.

If the lower node determined in step S11 is a leaf node, the process proceeds to step S13. Otherwise, the lower node is replaced with the node of interest, and the processing from step S11 is applied recursively. (Step S12 in FIG. 3).

The identification processing means 2 reads the selected feature element stored in association with the leaf node (FIG. 3).
Step S13). The identification processing unit 2 performs matching between the input pattern and the template below the leaf node using the selected feature element, and outputs a distance value belonging to each template (step S14 in FIG. 3). Finally,
The identification processing means 2 outputs the template category having the minimum distance value as the recognition result (step S1 in FIG. 3).
5).

In the embodiment of the present invention, since the subsequent classification method is optimized according to the subset of leaf nodes determined by the decision tree, a stable and high-speed recognition search can be performed. Further, according to an embodiment of the present invention, when inconsistency occurs at the boundary surface during clustering, a template is included in both nodes and registered, so that it is possible to prevent a complicated backtrack at the time of recognition. , A search result can be presented in a fixed search time.

Next, a description will be given of an example in which a correct category is searched from a character feature pattern by a binary tree. FIG.
It is assumed that the template is distributed in a two-dimensional feature space as shown in FIG. The number of templates belonging to a leaf node is set to less than 2 for simplicity. In the case of two dimensions, the identification plane of the subset is a straight line.

FIG. 6 shows the result of generating an identification plane according to the above algorithm. In FIG. 6, the identification planes are displayed with different thicknesses, but it indicates that the identification planes of the subset are generated in the order of thick → thin.

The corresponding binary tree is as shown in FIG. The templates existing across the identification planes of the subset are registered in duplicate on the binary tree, such as “A” and “R” in FIG.

For this reason, the category of the template can be specified by one vertical search without performing backtracking at the time of recognition. Further, since the feature for optimally identifying the two templates is selected in the leaf node, the calculation efficiency can be reduced.

Unlike the cases shown in FIGS. 6 and 7, since the feature vector is actually composed of multi-dimensions, it is necessary to match templates already used in the process of identification by the binary tree and belonging to leaf nodes. Excluding unnecessary feature vector elements is effective for increasing the efficiency of matching.

FIG. 8 is a block diagram showing a configuration of a high-speed recognition search system according to another embodiment of the present invention. 8, a high-speed recognition and retrieval system according to another embodiment of the present invention shown in FIG. 1 differs from the high-speed recognition and retrieval system according to one embodiment of the present invention shown in FIG. It has the same configuration, and the same components are denoted by the same reference numerals. The operation of the same components is the same as that of the high-speed recognition and retrieval system according to the embodiment of the present invention.

The category appearance frequency measurement unit 15 measures the prior probability of the appearance of a category for a set of data to be recognized and searched, and outputs this to the subset generation unit 13.
The subset generation unit 13 includes a category appearance frequency measurement unit 15.
A subset of the template is generated in consideration of the frequency of appearance of the category output by.

For example, if a template of a frequently occurring category is included in the node of interest, the size of the subset including the corresponding template is modified in accordance with the appearance frequency so as to be reduced. As a result, a frequently occurring template can be made a leaf node at an early stage.

If a frequently occurring template is a leaf node in a higher hierarchy, the amount of calculation required to reach the leaf node can be reduced, and a frequently occurring category can be searched quickly. Here, as a method of controlling the subset size, for example, when generating a lower subset, the identification plane is moved in parallel to the subset with a higher appearance frequency without making the identification plane into a bisecting hyperplane. .

Next, from the target node G ₀ , the subset G
A case where ₁ and G ₂ are generated will be described. Conventionally, when the appearance frequency is not considered (when the appearance frequencies of the categories ω _j belonging to the subsets G ₁ and G ₂ are equal), (x ₁ −x ₂ ) x− (‖x ₁ ‖ ² −‖x ₂ ‖ ² ) / 2 =
It becomes 0. Here, x ₁ , x ₂ , and x are vectors.

Considering that the identification plane is translated in the direction of the subset in which the appearance frequency of each category ω _j of each subset is high, (x ₁ −x ₂ ) x − ｛(A + 1) (x ₁ ) ² − (2A +
1) x ₁ x ₂ + A (x ₂ ) ² ｝ = 0. here,

(Equation 1) And the constant k is 0 <k ≦ 1, and the constant k is 1 / of the appearance frequency ratio.
This is a parameter for controlling the amount of movement of the identification surface that is changed according to the deviation from 2.

The other processing units of the learning processing means 1 and the identification processing means 2 operate in the same manner as the high-speed recognition and retrieval system according to the embodiment of the present invention, and the description of these operations will be omitted.

In another embodiment of the present invention, when the frequency of appearance of a category belonging to a subset is high, the size of the subset in the feature space is controlled in accordance with the frequency of appearance. Therefore, a category having a high appearance frequency can be made a leaf node in a shallow hierarchy of the decision tree. For this reason, the category of a template having a high frequency of appearance can specify the category with a small amount of calculation, so that it is possible to realize a speedy search.

For example, in the case of character recognition, Hiragana occupies most of the document, but if the frequency distribution of the character type in the target document is known, a decision tree corresponding to the distribution can be constructed. At the time of recognition, Hiragana is recorded so that the classification ends at the shallow hierarchy of the decision tree, so that the character recognition over the entire document can be performed at high speed.

As described above, a high-speed recognition search can be performed by optimizing a classification method for specifying subsequent categories according to the distribution of templates belonging to leaf nodes of a decision tree.

Further, when a decision tree is generated, a template existing over a boundary surface of a subset is included in both nodes and registered, so that complicated backtracking at the time of recognition can be prevented. In addition, search results can be presented for any template in a stable search time.

Further, when the frequency of appearance of the category belonging to the subset is high, the category having the high appearance frequency is determined by controlling the size of the subset in the feature space according to the frequency of appearance. Leaf nodes at shallow hierarchies in the tree make it possible to specify a category with a high frequency of occurrence for templates with a small amount of calculation, so that a more sensible search speed can be realized in accordance with the frequency of occurrence of the target category. Can be.

[0057]

As described above, according to the high-speed recognition search system of the present invention, the classification method for specifying subsequent categories is optimized and determined according to the distribution of templates belonging to leaf nodes of a decision tree. When generating a tree,
By registering the template existing over the boundary surface of the subset in both nodes and registering it, there is an effect that a high-speed search can be performed in a stable required time without backtracking.

According to another high-speed recognition and retrieval system of the present invention, when the frequency of appearance of a category belonging to a subset is large, the size of the subset in the feature space according to the magnitude of the appearance frequency Is controlled, the structure of the decision tree can be optimized according to the appearance probability of the object, and the effect of improving the perceptual search efficiency can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a high-speed recognition search system according to an embodiment of the present invention.

FIG. 2 is a flowchart illustrating a processing operation of a learning processing unit in FIG. 1;

FIG. 3 is a flowchart showing a processing operation of an identification processing unit in FIG. 1;

FIG. 4 is a diagram for explaining a processing operation of the high-speed recognition search system according to one embodiment of the present invention;

FIG. 5 is a diagram for explaining a processing operation of the high-speed recognition search system according to one embodiment of the present invention;

FIG. 6 is a diagram for explaining a processing operation of the high-speed recognition search system according to one embodiment of the present invention;

FIG. 7 is a diagram for explaining a processing operation of the high-speed recognition search system according to one embodiment of the present invention;

FIG. 8 is a block diagram showing a configuration of a high-speed recognition search system according to another embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 1 learning processing means 2 identification processing means 11 template dictionary creation unit 12 template dictionary storage unit 13 subset generation unit 14 hierarchical dictionary storage unit 15 category appearance frequency measurement unit 21 decision tree classification unit 22 category determination unit

Claims (11)

[Claims] 1. A feature vector is generated from an input character pattern, the feature vector is identified according to a condition stored in each node of a previously generated decision tree, and child nodes are sequentially selected according to the identification result. A high-speed recognition and retrieval system that repeatedly performs this classification until reaching a terminal node. The system includes a template of a multidimensional feature vector stored in a recognition dictionary from a set of patterns to which a predetermined correct category is assigned. Generating means for generating; a template dictionary storing means for storing a template created by the generating means and a pattern contributing to the generation of the template in association with each other; a template currently focused on and a pattern corresponding to the template. Classifying the set and the occurrence frequency of the correct category into subsets and A subset generation unit that outputs a template belonging to the subset and a threshold value for performing separation into the subset, and a subset of the template that is sequentially generated by the subset generation unit. Hierarchical dictionary means for storing the hierarchical structure stored in the hierarchical dictionary storage means in association with the subset, and classifying an input pattern by sequentially inputting the hierarchical structure stored in the hierarchical dictionary storage means from an upper hierarchy, and outputting a child node of the classified result Means for reading a feature amount effective for determining a template from the leaf node of the hierarchical structure, and a category determining means for performing a large classification using the feature amount, . 2. The method according to claim 1, wherein the subset generating means performs feature reduction of a template included in the subset when the subset is a leaf node, and selects and outputs only valid feature components. 2. The high-speed recognition search system according to claim 1, wherein: 3. The method according to claim 1, wherein the decision tree classifying unit is configured to end the classification when the classification of the pattern into child nodes is completed to the lowest layer.
Or the high-speed recognition search system according to claim 2. 4. The method according to claim 1, wherein the subset generation unit is configured to generate a decision tree by including categories existing over the determined threshold in subsets on both sides of the threshold. Item 4. A high-speed recognition and retrieval system according to any one of Items 3. 5. A category appearance frequency measuring unit for measuring the appearance frequency of the correct category and outputting the appearance frequency of the correct category to the subset generating unit, wherein the subset generating unit includes: The high-speed recognition search according to any one of claims 1 to 4, wherein the classification into subsets is configured to output a threshold value for controlling a boundary for classification according to the prior probability of the correct category. system. 6. A feature vector is generated from an input character pattern, the feature vector is identified according to a condition stored in each node of a previously generated decision tree, and child nodes are sequentially selected according to the identification result. A high-speed recognition / searching method for a high-speed recognition / searching system that repeatedly performs this classification until reaching a terminal node, wherein a plurality of dimensions stored in a recognition dictionary are obtained from a set of patterns to which a predetermined correct category is added. Generating a template of the feature vector of the template, classifying a set of the template currently focused on and the pattern corresponding to each of the templates and the appearance frequency of the correct answer category into subsets, and Outputting a threshold value for performing separation into subsets;
A hierarchical structure stored in a hierarchical dictionary means for storing a subset of sequentially generated templates in association with a corresponding subset of a pre-separated template and inputting the hierarchical structure in order from the upper hierarchy, and classifying the input pattern and classifying the result And a step of reading out effective feature amounts for determining a template from the leaf nodes of the hierarchical structure and performing a large classification using those feature amounts. How to speed up recognition search. 7. A step of outputting a threshold value for separating a template belonging to the subset from a subset, wherein the feature degeneration of a template included in the subset when the subset is a leaf node. 7. The method according to claim 6, wherein only effective feature components are selected and output. 8. The step of outputting the child node as a result of the classification is such that when the classification of the pattern into child nodes has been completed to the lowest layer, the classification is ended. A method for accelerating a recognition search according to claim 6 or 7. 9. The step of outputting a threshold value for performing separation into a template and a subset belonging to the subset is performed by including categories existing over the determined threshold in subsets on both sides of the threshold. 9. The method according to claim 6, wherein a tree is generated. 10. A method for measuring the frequency of appearance of the correct category and outputting the frequency of appearance of the correct category to the subset generating means, for separating the subset belonging to the template from the subset. The step of outputting a threshold value of the template set is output as a threshold value for controlling a boundary for classification in accordance with a prior probability of the correct answer category in classifying the template set into the subset. The method for accelerating recognition search according to any one of claims 6 to 9. 11. A feature vector is generated from an input character pattern, the feature vector is identified according to a condition stored in each node of a previously generated decision tree, and child nodes are sequentially selected according to the identification result. A storage medium storing a recognition search speed-up control program for speeding up the recognition search in the recognition search device that repeatedly performs this classification until reaching the terminal node; A search device is configured to generate a template of a multidimensional feature vector stored in a recognition dictionary from a set of patterns to which a predetermined correct category has been assigned, and to select a template of interest and a pattern corresponding to each of the templates. And the frequency of appearance of the correct answer category are classified into subsets, and A hierarchical structure stored in a hierarchical dictionary means for outputting a template to which the template belongs and a threshold value for performing separation into subsets, and storing the sequentially generated template subset in association with a corresponding subset of the template before separation. Are input in order from the upper hierarchy to classify the input pattern and output the child nodes as a result of the classification, read out effective feature amounts for determining a template from the leaf nodes of the hierarchical structure, and read those feature amounts. A recording medium on which a recognition and retrieval speed-up control program is recorded, wherein a large classification is performed by using a computer.