JPH0638276B2 - Pattern identification device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for creating a dictionary (a set of standard patterns) of a pattern identification device using a pattern matching method.

[Background of the Invention]

Conventionally, there are a pattern matching method and a structural analysis method for character recognition. The former has been mainly used for printing and the latter has been used for handwritten characters. However, if the recognition target is expanded from alphanumeric characters to kanji, handwritten characters also require a pattern matching method. Therefore, the pattern matching method is a mere blurring process (for example, there is a composite similarity method; a hierarchical pattern matching method so that it can endure handwritten characters. For details, see “Introduction to Character Recognition”, edited by Hashimoto, Telecommunications Association, p.35 and p88), a pattern matching method for extracting features such as pattern directionality is being developed (for example, Japanese Patent Laid-Open No. 56-65275, "Pattern Recognition Method").

On the other hand, the pattern matching method has a problem of how to make a dictionary, or how to improve the contents of the dictionary to successively improve the recognition accuracy. In order to put a pattern recognition device such as a character recognition device having sufficient recognition accuracy into practical use, the above problem is actually more important than the detailed contents of feature extraction. More specifically, in a principle recognition method, for example, 95
Even if the accuracy of 10% is obtained, it is necessary to raise the accuracy to 99.9% for practical use. There are three problems here.

The first problem is that recognition accuracy depends on the dictionary as well as the features used for identification. The superiority or inferiority of features cannot always be easily determined. Especially in the case of handwritten characters, improvement of feature extraction alone does not reach a practical level, and it is indispensable to prepare a plurality of standard patterns for each category. It is clear that there are multiple pattern concepts (forming pattern clusters) for the abstract concept of category, if we try to allow deformation to the character shape. That is, the first problem is that improvement of feature extraction alone does not reach a practical level.

The second problem is that the sample patterns used in the experiment are limited, and there is no guarantee for patterns not included therein. Most of the time actually spent in practical application of the character recognition device is spent in improving the dictionary so as to maintain sufficient recognition accuracy even for a new input pattern. The main reason for this is that the measures for improving the accuracy of new input patterns conflict with the measures taken before, and in many cases the fact cannot be detected. Therefore, it is necessary to recognize all the sample patterns and confirm the effect after the measure for improving the accuracy, and eventually it is necessary to repeat the measure and the confirmation.

The third problem is how to make a dictionary based on the pattern matching method, and the problem to solve the second problem.

Conventionally, it has been found that when the number of standard patterns for one category is increased, the recognition accuracy increases exponentially with respect to the number. At that time, a process called clustering is used to obtain a plurality of standard patterns. However, in-category clustering, which is usually performed, has the disadvantages that the number of standard patterns required to reach sufficient accuracy is too large and the number of clusters is not balanced because clustering is performed independently for each category. . Further, generally, there is a drawback that the whole process of clustering has to be redone when a new input pattern is obtained and a dictionary improvement is required.

[Object of the Invention]

An object of the present invention is to solve the above problems and to provide a pattern identification device that can create a pattern identification dictionary with high reliability in a short period of time.

[Outline of Invention]

According to the present invention, in a pattern identification device provided with a pattern identification dictionary based on the pattern matching method, a means for performing feature extraction from a sample pattern to extract a feature pattern, a means for accumulating the feature pattern, and an accumulation means. Hierarchical clustering is performed on the feature patterns with clusters of different sizes, and at this time, at least the lowest hierarchy has a predetermined pattern so that the feature patterns belonging to one cluster are composed of a single category. Clustering with a cluster of a size, expressing the cluster configuration of each layer by a layered structure, creating a standard pattern according to the cluster configuration of each layer according to the layered structure, and a standard according to the cluster configuration of each layer A means for accumulating patterns as the pattern identification dictionary and created according to the cluster configuration of each layer Among the quasi-patterns, the distance or similarity between the input pattern and the standard pattern is sequentially obtained from the standard pattern created according to the cluster structure of the upper layer, and the obtained distance or similarity is evaluated to obtain the standard. Means for narrowing down the pattern candidates and identifying the input pattern, and for the unknown input pattern, the unknown level of each unknown layer above the lowest hierarchical layer subjected to the hierarchical clustering. The feature pattern of the input pattern is sequentially added and re-clustering is performed. As a result of the re-clustering, re-clustering is performed up to a layer in which a cluster other than the cluster to which the feature pattern of the unknown input pattern does not change, Change the structure to modify the standard pattern according to the cluster configuration of each layer, and Characterized by comprising a means for storing in the storage means as use a dictionary.

First, the principle of the present invention will be described with reference to FIGS. 3 categories (A, B, C) for simplification
Are scattered in a two-dimensional space.

FIG. 1 is a schematic diagram when one standard pattern is provided for each category. Dashed lines 100, 200 and 300 are hypercircles (in the case of the simple similarity method) or hyperelliptic bodies (in the case of the compound similarity method) centered on the respective standard patterns. As shown in the figure, in the single standard pattern, the area covered by each standard pattern overlaps, and in the case of handwritten characters, the deformation is large, and thus the overlapping area is large. That is, unrecognizable images often occur.

FIG. 2 is a schematic diagram when clustering is applied independently for each category. In clustering, there is a parameter that specifies the size of each cluster. Generally, there is no definition of a cluster, and a human determines which intermediate point is to be obtained between the case where the whole is imitated as one cluster and the case where each sample is imitated as one cluster. Depends on the parameter. Second
The figure is the result of clustering with certain parameters.

As shown in the figure, since the categories are clustered independently, generally the overlapping area (standard pattern)
Remains. Therefore, in order to increase the recognition accuracy, it is necessary to specify finer clustering and to re-execute the entire clustering. Further, when the whole re-clustering is performed, clusters other than the overlapping standard patterns are also subdivided, and extra standard patterns are generated as a whole. That is, unnecessarily many standard patterns are required to achieve high accuracy. Furthermore, when a new input pattern that cannot be recognized correctly appears after the dictionary is constructed, there is a problem that clustering must be re-executed.

FIG. 3 is a diagram for explaining the principle of the method of the present invention.
In the hierarchical clustering according to the present invention, all samples are clustered without being aware of category boundaries. This clustering method itself may be the same as the clustering method applied to each category unit of the method described in FIG. It is assumed that the result of clustering with a certain parameter is shown in FIG. Generally, when a coarse cluster is designated, the sample patterns forming the cluster are not always occupied by the same category. Fig. 3
In the example of (a), two clusters consist of categories A and B, and B and C, respectively.

For the cluster having the mixed categories as described above, the clustering is applied to the inside thereof. It is a layer of LEVEL-2. This process is recursive and repeats until all clusters consist of sample patterns that belong to only one category. In the example of FIG. 3, LEVEL-3 (3rd
Up to the layer), all clusters are in a single category.

FIG. 4 shows the cluster structure of FIG. 3 in a hierarchical tree. In FIG. 4, S means a sample set.
The “leaf” portion of the hierarchical tree is a cluster, and the symbols there represent sample patterns that are members that make up the cluster. The standard patterns that make up the dictionary for pattern recognition are created from the members of each cluster. A commonly used method is to find the average of the sample patterns (mathematically vectors) that are members. In Figure 4
₁ to ₂₃ represent the average vector of each cluster. Of course, in the pattern matching method for feature extraction, the feature vector represents the sample pattern.

Now, a case where a new pattern is input will be described. First, the case where this input pattern is not correctly identified will be described with reference to FIG.

If there is an initial cluster shown in FIGS. 5 (a) and 5 (b) and a new pattern C ₃ is input thereto, the new pattern C ₃ is determined to be closest to the cluster ₃ from the identification result. However, since the cluster ₃ is a category B cluster, it is misidentified. Here, in this method, the following partial re-clustering is performed.

First, since it is determined that C ₃ belongs to the cluster ₃ , hierarchical clustering is applied to the members {B ₃ , B ₄ , C ₃ }, and in this case, the cluster tree shown in FIG. 6 is obtained as a result. Here clearly the central vector of cluster ₃ is
Move to ₃ ', ₅ to ₅ ', ₀ to ₀ '. Therefore, samples at the level with clusters with a high rate of variation may not be correctly identified. Therefore, it is tested whether or not the lower level can be correctly identified, and if it is not possible, the hierarchical clustering after the same level is performed again.

In the example of FIG. 6, ‖ ₃ - _{3 '‖} compares whether greater than a predetermined threshold. Now, assuming that a large sample position to the level below the hierarchical tree of clusters _{3 '{A 1, A 2} , B 1, B 2, B 3, B 4, C 3}
Is tested for correct discrimination. This test is performed on clusters ₁ , ₂ , and ₃ '. If it is correctly recognized, {B ₃ , B ₄ , C ₃ } should be correctly recognized for the clusters ₇ and ₈ , and the test of the same level is terminated. If some samples are not correctly identified, cluster ₁ , ₂ ,
Redo hierarchical clustering after the level ₃ '.

Once the same level of testing and modification is complete, do the same for the levels above it (clusters ₅ ', ₆ '). That is, the amount of movement of the cluster center vectors ‖ ₅ - ₅ '
‖ Is compared with the threshold value ε, and when it is small, the reclustering is terminated, and when it is large, the sample {A ₁ , A ₂ , B ₁ , B
₂ , B ₃ , B ₄ , C ₃ , C ₁ , C ₂ } are tested correctly for clusters ₅ ′ and ₆ ′. If all are correctly identified, re-clustering is terminated, and if not, hierarchical clustering is performed again for all samples.

The reclustering termination condition may be a condition in which the arrangement of samples in adjacent clusters does not change.

Different from the case of FIG. 5, the case where the new pattern is correctly identified as shown in FIG. 7A will be described. In this case moves cluster ₃ the central vector the ₃ ', it is modified to include members _{{B 3, B 4, B} 5}. Further, the cluster center vector ₅ moves to ₅ '. If the cluster center vector moves, the correct identification of the sample pattern is generally not guaranteed, so the reclustering process described above is performed in exactly the same way.

There can be a transformation algorithm that does nothing if the new pattern is corrected correctly, but there is a risk that the previously correctly identified pattern will not be correctly identified by reclustering due to other causes later. Yes, not desirable.

Example of Invention

Hereinafter, the pattern identification device of the present invention will be described in detail based on examples. FIG. 8 is a system diagram showing the configuration of an embodiment of the present invention. In FIG. 8, the form 1 is converted into an electric signal by the photoelectric conversion device 2. The photoelectric conversion image is scanned by the scanning control circuit 3. This scan output is binarized by the threshold circuit 4. It is effective to provide the threshold circuit 4 with a variable threshold value according to the density distribution to correct the shading. The binarized character group is sent to the character cutting circuit 5 and is cut out character by character. The cut-out characters are subjected to processing such as noise removal and size normalization in the preprocessing unit 6. Information such as strokes required for classification is extracted by the feature extraction unit 7 from the preprocessed characters. It is obvious that the features to be extracted can be applied to this method regardless of the types.
Next, the extracted features are sent to the feature storage unit 10 via the one-character storage unit 8 together with the information indicating the attribute input from the keyboard. When the counter 9 counts a predetermined number of characters by repeating all of the series of processing classification target characters, the pattern creating operation ends. Thus, a set of characteristic patterns to be processed collectively is stored in the characteristic storage unit 10. Here, the information representing the attribute of the feature represents a code representing the type of character and the order in the code. Further, each circuit described above does not necessarily have to be a dedicated logic circuit, and a microcomputer storing a processing procedure may be used.

The characteristic pattern stored in the characteristic storage unit 10 is taken out by the hierarchical cluster creation device 11 described later and sequentially subjected to clustering processing. When the processing for all the characteristic patterns stored in the characteristic storage unit 10 is completed, the hierarchical data structure shown in FIG. 4 is sent to the hierarchical structure storage unit 12. Next, the standard pattern creation device 13 described later synthesizes the standard pattern corresponding to each node shown in FIG. 4 from the features of the feature storage unit 10 according to the information of the hierarchical structure storage unit. The synthesized standard patterns are sequentially sent to the standard pattern storage unit 14 and stored therein. However, from the viewpoint of reducing the storage capacity of the standard pattern storage unit 14, it is acceptable to thin out the nodes that create standard patterns. In this way, the standard pattern by batch processing is constructed. By using the standard pattern configured in this way and adopting a classification method that sequentially matches from the top of the hierarchical tree shown in FIG. 4, at least all learned characters will be correct. Is self-evident.

Next, a method of correcting a standard pattern by sequential processing will be described. It is obvious that input characters learned in advance can be classified as correct answers, but this is not the case with unlearned input characters. Therefore, it becomes necessary to learn unlearned input characters. On the other hand, in the method of constructing a standard pattern by batch processing, hierarchical input must be applied to all inputs to create a standard pattern for each node of the hierarchical tree. It is not desirable from a practical point of view to execute such a huge amount of processing each time an unlearned character is input. Therefore, a method of correcting the standard pattern with a small amount of processing by performing local clustering without impairing the property that all learned characters are correct answers is a standard pattern correction method by sequential learning. That is one of the main points of the present invention.

The sequential learning procedure will be described with reference to FIG.

The form 1 is converted into an electric signal by the photoelectric conversion device 2. The photoelectric conversion image is scanned by the scanning control circuit 3. This scan output is binarized by the threshold circuit 4. The binarized character group is sent to the character cutting circuit 5 and is cut out character by character. The cut-out characters are subjected to processing such as noise removal and size normalization in the preprocessing unit 6. A predetermined feature is extracted from the feature extraction unit 7 for the preprocessed character. The extracted feature is sent to the one-character storage unit 8 together with the information indicating the attribute. The unknown input character to be learned sent to the one-character storage unit 8 is transferred to the matching unit 15 in order to verify which cluster in the hierarchical tree shown in FIG. 4 belongs.
The matching unit 15 obtains a distance from the standard pattern stored in the standard pattern storage unit according to the hierarchical tree held in the hierarchical structure storage unit 12, and sequentially in the same layer from the top of the hierarchical tree. Select the node with the smallest distance in and verify which cluster it belongs to at the bottom of the hierarchical tree. Next, the determination unit 16 determines whether or not the unknown input character is correctly classified. If it is not correctly classified, the hierarchical cluster creation device 11 is activated via the control unit 17. The hierarchical cluster creation device 11 reads a hierarchical tree from the hierarchical structure storage unit 12, and determines the characteristics of the cluster above the lowermost layer including the unknown input character as the characteristic storage unit 1.
Clustering is performed by reading from 0. When the hierarchical clustering of the partial clusters is completed, it is compared with the hierarchical tree accumulated in the hierarchical structure storage unit 12. If there is no change other than in the cluster containing the unknown input character, the process ends. When a change occurs, the same process is performed including the clusters on the higher level and the process is repeated until the change does not occur. When there is no change in the surrounding clusters, a new hierarchical tree is written in the hierarchical structure storage unit 12. When the clustering process is completed, the standard pattern creation device 13 is activated via the control unit 17. Standard pattern creation device 13
Creates a standard pattern to be synthesized from the features of the feature storage unit 10 according to the updated information in the hierarchical structure storage unit 12. Thus, the sequential learning when the unknown input character is not correctly classified is completed. Even if it is correctly classified, the center vector of the cluster moves, so the same process as the process of not being correctly classified is performed. However, a modification process in which no correction operation is performed when the classification is performed correctly as described above is also possible.

Next, the classification device will be described with reference to FIG. The form 1 is converted into an electric signal by the photoelectric conversion device 2. The photoelectric conversion image is scanned by the scanning control circuit 3. This scan output is binarized by the threshold circuit 4. The binarized character group is sent to the character cutting circuit 5 and is cut out character by character. The cut-out characters are subjected to processing such as noise removal and size normalization in the preprocessing unit 6. A predetermined feature is extracted from the feature extraction unit 7 for the preprocessed character. The extracted features are transferred to the matching unit 15 via the one-character storage unit 8. The matching unit 15 obtains a distance from the standard pattern stored in the standard pattern storage unit according to the hierarchical tree held in the hierarchical structure storage unit 12, and sequentially in the same layer from the top of the hierarchical tree. The node with the smallest distance is selected to narrow down the candidates, and the category with the smallest distance among the nodes in the lowest layer is sent to the determination unit 16. Judgment unit 16
Then, the threshold used for the clustering of the bottom layer is used as the determination threshold, and if the candidate distance is within the threshold, the correct answer is given, and conversely, if it is larger than the threshold, the reject is given. However, a method in which the threshold value is not used in the determination unit 16 or a method in which a plurality of candidates are input and the relative threshold value is used can be considered, but any method may be used because it has nothing to do with the principle of the present invention.

FIG. 9 is a system diagram showing the configuration of the hierarchical cluster creation device. The register 111 fetches the characteristic pattern for each character from the characteristic storage unit 10. In addition, in the register 112,
When the characteristic pattern for one character is fetched in the register 111, the central pattern of each cluster in the cluster file 118 is smoothly fetched and the distance calculation circuit 113 obtains the distance between the characteristic patterns of the register 111 and the register 112. The determined distance is sent to the comparator 114 for comparison with a predetermined threshold. The register 115 stores the threshold value determined for each layer in the threshold file 11.
It is taken out from 6 and held. The timing of changing the threshold value is controlled by the counter 119.
The counter 119 counts up each time a characteristic pattern is retrieved from the characteristic storage unit 10, and changes the threshold value of the register 115 after counting a predetermined number of patterns. The comparator 114 compares the threshold value of the register 115 with the distance calculated by the distance calculation circuit 113. If the threshold value is less than or equal to the threshold value, the feature held in the register 111 belongs to the cluster held in the register 112. As a result, an execution command is sent to the arithmetic circuit 117. Arithmetic circuit 1
17 is a register 111 based on a command from the comparator 114.
Is divided by the number of constituent clusters, and added to the cluster center held in the register 112 to obtain a new cluster center. The new cluster center thus obtained is transferred to the cluster file 118 together with the attribute of the characteristic pattern held in the register 111. Further, in the comparator 114, if the distance is equal to or larger than the threshold value, it is determined that the characteristic pattern of the register 111 does not belong to the cluster of the register 112, and another cluster center from the cluster file 118 is set in the register 112. It is sent out and this process is repeated until it becomes less than the threshold value. Further, when the feature pattern held in the register 111 is not captured in any cluster center of the cluster file 118, a new cluster center is newly established. That is, the arithmetic circuit 117
Divides the characteristic pattern of the register 111 by 1, adds it to the 0 pattern held in the register 112, and transfers it to the cluster file 118.

When this process is repeated until all the characteristic patterns are input to the register 111, a cluster configuration table of cluster centers and clusters for one layer is created in the cluster file 118. When the configuration table for one layer is created, it is transferred to the hierarchical structure storage unit 12 via the register 122. At this time, since the center of each cluster held in the cluster file 118 is used as a standard pattern, the standard pattern storage unit 14
It can be transferred to (Fig. 8). When the processing for one layer is completed, the counter 119 detects and the register 11
The threshold value of 5 is updated. Based on the updated threshold value, the clustering in the lower stage is started, and the clustering is completed when all the characteristic patterns forming each cluster have the same character type.

In the case of sequential learning, the required partial hierarchical tree is transferred from the hierarchical cluster storage unit 12 to the register 121. Based on the information in the register 121, the characteristic pattern is extracted from the characteristic storage unit 10 and the same clustering process as described above is performed.

The above control is processed by the control circuit 123. The above-mentioned embodiment is constructed based on a so-called simple method algorithm, but it may be based on another method.

FIG. 10 is a system diagram showing the configuration of the standard pattern creating device 13. The register 132 retrieves the attribute information of one cluster in the hierarchical tree from the hierarchical structure storage unit 12. Further, when the information of one cluster is fetched in the register 132, the register 131 sequentially fetches the feature patterns and their attribute information stored in the feature storage unit 10, and the comparator 13
At 3, the attribute information of the registers 131 and 132 is compared. The attribute information held in the register 131 is the register 13
When it matches the cluster of 2, the latch 134 is opened, and the characteristic pattern of the register 131 is sent to the adder 136. The adder 136 receives the input characteristic pattern and the register 135.
The results held in are added and the result is added to register 1
35. The counter 137 has a register 132
The number belonging to the cluster held in is counted, and when the counting of the predetermined number of characters is completed in the register 138, the number is sent to the divider 139. The divider 139 calculates the average value of the cluster using the addition result of the register 135 and the number of counters 137, and sends it to the standard pattern storage unit 14 via the register 141. Thus, one standard pattern corresponding to the node of the hierarchical tree is obtained. When this operation is performed for all the nodes, the standard pattern creating process ends.

Although one embodiment of the present invention has been described above, it goes without saying that the present invention is not limited to this embodiment. Further, in the above description, the method of recognizing the category having the smallest distance has been described, but the method of recognizing the category having the maximum similarity may be used.

According to the present embodiment, the misread character or the erroneously recognized voice can be learned by the local calculation, so that the sequential learning, which is practically impossible due to the limitation of the calculation amount, can be performed. Further, since it is suitable for the hierarchical pattern matching method, there is an effect that the speed of the recognition device can be greatly improved.

〔The invention's effect〕

According to the present invention, misread characters or erroneously recognized voices can be learned by a local calculation, so that it is possible to perform sequential learning which was not practically possible due to the restriction of the amount of calculation in the related art, and the recognition device classification It can greatly contribute to the improvement of performance. Further, as a derivative effect, the hierarchical pattern matching method can be easily adopted, which can greatly contribute to the speedup of the recognition device. Therefore, by using the learning / classifying method as in the present invention, a device useful for pattern recognition, particularly for recognizing Chinese characters and large narration speech, can be easily realized.

[Brief description of drawings]

1 and 2 are diagrams showing the principle of clustering, FIG. 3 is a diagram for explaining the principle of hierarchical clustering according to the present invention, and FIG. 4 is a schematic diagram showing an example of hierarchical tree representation according to the present invention. , FIG. 5 is a pattern layout diagram and a hierarchical tree diagram thereof before sequential learning, FIG. 6 is a diagram for explaining a correction method when misidentification is made in sequential learning, and FIG.
FIG. 8 is an explanatory diagram of a correction method when a correct answer is obtained in sequential learning, and FIGS. 8 to 10 are system diagrams showing the configuration of an embodiment of the present invention. 100, 200, 300 ... Cluster, 1 ... Form, 2 ... Photoelectric conversion device, 3 ... Scan control circuit, 4 ... Threshold circuit, 5
... character cutout circuit, 6 ... preprocessing unit, 7 ... feature extraction unit,
8 ... Single character storage unit, 9 ... Counter, 10 ... Feature storage unit, 11 ... Hierarchical cluster creation device, 12 ... Hierarchical structure storage unit, 13 ... Standard pattern creation device, 14 ... Standard pattern storage unit, 15 ... Matching unit, 16 ... judgment unit, 17 ... control unit,
111, 112, 115, 121, 122 ... Registers,
113 ... Distance calculation circuit, 114 ... Comparator, 116 ... Threshold file, 117 ... Arithmetic circuit, 118 ... Cluster
File, 119 ... Counter, 123 ... Control circuit, 1
31, 132, 135, 141 ... Registers, 133 ... Comparators, 134 ... Latches, 136 ... Adders, 137, 13
8 ... Counter, 139 ... Divider, 142 ... Control circuit.

Claims (1)

[Claims] 1. A distance or similarity between an input pattern and a standard pattern stored in a pattern identification dictionary is calculated,
In a pattern identification device for identifying the input pattern according to the obtained distance or similarity, a means for extracting a feature pattern by extracting a feature from a sample pattern, a means for accumulating the feature pattern, and a means for accumulating the feature pattern On the other hand, hierarchical clustering is performed on clusters of respective sizes, and at this time, at least the lowest hierarchy has a predetermined size so that the feature patterns belonging to one cluster are composed of a single category. The clustering is performed by clusters, the cluster configuration of each layer is represented by a layered structure, and means for creating a standard pattern according to the cluster configuration of each layer according to the layered structure and the standard pattern according to the cluster configuration of each layer are described above. A means for accumulating as a pattern identification dictionary and created according to the cluster configuration of each layer above. Among the standard patterns, the distance or similarity between the input pattern and the standard pattern is sequentially obtained from the standard patterns created according to the cluster structure of the upper layer, and the obtained distance or similarity is evaluated. To narrow down the standard pattern candidates,
A means for identifying the input pattern, and a feature pattern of the unknown input pattern for each layer with respect to a layer higher than the lowest layer subjected to the hierarchical clustering with respect to the unknown input pattern. In addition, re-clustering is sequentially performed, and as a result of the re-clustering, re-clustering is performed up to a hierarchy in which clusters other than the cluster to which the characteristic pattern of the unknown input pattern does not change, and the hierarchical structure is changed to A pattern discriminating apparatus comprising: a unit that corrects the standard pattern according to a cluster configuration of each layer and stores the standard pattern in a storage unit as the pattern identifying dictionary.