JP3470017B2 - Multi-layer dictionary creation method and multi-layer dictionary creation device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a multi-layer dictionary creating method and a multi-layer dictionary creating apparatus for creating a registered multi-layer dictionary by clustering information of characters and voices to be recognized into multi-layer floors.

[0002]

2. Description of the Related Art With the increase in speed and capacity of hardware such as computer memory and CPU (Central Processing Unit), character recognition devices and voice recognition devices have become more sophisticated and have higher performance. In the character recognition device and the voice recognition device as described above, in addition to the realization of a high recognition rate at the beginning when the hardware backup was not as thick as it is now,
The compression of recognition dictionary and recognition information has been a big problem. In order to solve this problem, powerful recognition methods such as the composite similarity method have been developed and are still widely used.

In the above-mentioned composite similarity method, a small number of standard patterns that best approximate the distribution of a large number of learning samples are obtained and used as a dictionary. Then, recognition is performed using the small number of dictionaries. Since such a composite similarity method can realize a certain recognition rate and reduce the dictionary capacity,
It can be said that it is an extremely excellent recognition method that meets the original purpose of compressing the recognition dictionary (large sample → small number dictionary).

However, the composite similarity method (or
In the method of using a minority dictionary that compresses similar recognition information), since each category is represented by a small number of standard patterns, there is an adverse effect that the boundary between different categories is blurred. FIG. 16 is a relationship diagram of neighborhood categories in the method using the above-described information-compressed dictionary. In this figure, actually, a standard sample of a multidimensional space is simulated and plotted on a two-dimensional plane, and the overlapping of the boundary portions of the neighborhood categories is shown by diagonal lines. The larger the shaded area, the lower the recognition performance of similar categories.

As shown in FIG. 16, in the above-mentioned composite similarity method, the boundary portions of the neighborhood categories overlap with each other, and in the case of character recognition, for example, "8" and "B" or "5" and "S". As a result, the discrimination ability at the boundary between similar characters is deteriorated.

Therefore, various solutions have been proposed for such adverse effects. The simplest and most excellent method among them is the nearest neighbor (multi-template) method. In short, the nearest neighbor method is to prepare a large number of standard patterns for each category and recognize them using the large number of standard patterns. FIG. 17 is a relationship diagram of neighborhood categories in the nearest neighbor method. It can be seen that the overlapping of the boundary portions of the neighboring categories is small and the discriminating ability is high as compared with the relational diagram of the neighboring categories in the composite similarity method as shown in FIG.

The concept of the nearest neighbor method has been around for a long time, and its superiority has been proved. On the other hand, there is an adverse effect that the dictionary capacity is significantly increased and the recognition speed is decreased. Therefore, it was excluded from the original recognition device development approach. However, at present, the problem of dictionary memory capacity has been solved by the progress of device technology. However, the problem of the recognition speed still remains because the required speed specifications for the current recognition device have dramatically increased.

In order to execute the above-mentioned nearest neighbor method at high speed, instead of calculating the distance (or similarity) from all dictionary samples (standard pattern) to an unknown input,
It is necessary to effectively narrow down the calculation range on the dictionary sample. As such a recognition device, a character recognition method (Japanese Patent Laid-Open No. 63-263590) using a hierarchical structure dictionary in which each character is classified into a group of multi-layered floors has been proposed.

The hierarchical structure dictionary is created as follows. That is, the feature vector of all registered characters is divided into n1 first-stage classes by clustering, and an average vector is obtained for each first-stage class to obtain a standard pattern. Next, each class in the first stage is divided into n2 classes in the second stage by clustering, and an average vector is obtained for each class in the second stage to obtain a standard pattern. In this way, a hierarchical structure dictionary having a three-layer structure is formed. Then, when recognizing the input character, the distance between the input character and the standard pattern of each class of the first stage is obtained, and p candidate classes of the first stage are obtained from the smaller distance. Next, the distance between the input character and the standard pattern of each class of the second stage existing in the first-stage candidate class is obtained, and q second-stage candidate classes are obtained from the smaller distance. Finally, the distance between the input character and the registered character existing in the second-stage candidate class is obtained, and the registered character having the minimum distance is used as the recognition result.

[0010]

However, the above-mentioned conventional character recognition method using a hierarchical structure dictionary has the following problems. Note that, hereinafter, when simply referred to as "character", it means one in which variations of the font and the binarization level are added. On the other hand, what does not take the above variations into consideration is called a "category".

(1) Low recognition rate The registered characters registered in the final layer (third stage) in the above-mentioned hierarchical dictionary are not all characters in the database, but some characters are compressed. It was done. In particular, regarding the Japanese OCR (optical character reader) database, there are about 4000 categories to be recognized, and the number of characters in the entire database is at least several million in order to provide various fonts and binarization level variations. Will be ordered. In order to register this large number of characters in the final layer of the hierarchical dictionary, the capacity of the dictionary memory becomes enormous. Therefore, some compression must be applied.

A general compression method in that case is, for example, to integrate samples having a very small distance threshold or less within the same category, and a threshold of 1/2 of the distance to the nearest neighbor heterogeneous category, For example, the same powers that fall within the threshold are integrated. Thus, the characters registered in the final layer obtained by compressing the characters in the entire database have already lost the category boundary information in the database.

Then, for the registered characters that have lost the category boundary information on this database, the standard pattern is obtained by performing the clustering as described above and obtaining the average vector for each class. Loss of the category boundary information will result in further information compression. Therefore, in the end, similar to the case of the composite similarity method shown in FIG. 16, the category boundary information on the database is largely lost near the boundary of each cluster. Since this operation is repeated for each layer, it is inevitable that the recognition rate is lowered.

In particular, in actual recognition, it is extremely rare that a character having the same feature vector as the feature vector prepared in the database is input, and in most cases, an unknown character is input. Therefore, even if 100% of the category boundary information on the database is stored in the dictionary, the recognition rate is usually low, and if the category boundary information on the database is lost like the hierarchical structure dictionary. Will further reduce the recognition rate.

(2) The standard pattern of each class in each stage of the high-capacity dictionary memory is an average vector obtained for each class by clustering n1 and n2. Therefore, this obtained standard pattern is
It is different from any of the character feature patterns registered in the final layer. Therefore, it is necessary to register all the standard patterns of each class in each stage in the dictionary memory, which increases the capacity of the dictionary memory.

Since the increase in the dictionary memory capacity in this case increases with the exponential function of the dictionary hierarchy as follows, the capacity of the dictionary memory increases as the high-hierarchical dictionary is constructed to obtain a high recognition rate. It will increase. Dictionary memory capacity = (v ^r ) × s where v: number of representative vectors per node of hierarchical dictionary (= n1, n2) r: number of layers s: storage capacity per representative vector

Therefore, an object of the present invention is to store a 100% category boundary information on a database and to create a multi-layer dictionary with a small storage capacity.
Another object is to provide a multi-layer dictionary creation device.

[0018]

In order to achieve the above object, a multi-layer dictionary creating method according to a first aspect of the present invention provides a feature vector of recognition target information as a learning vector, and obtains the learning vector by clustering. A plurality of representative vectors thus obtained are used as templates, and a plurality of representative vectors obtained by clustering the template and one root node to which the representative vectors belong are set as an initial state of the multi-layer dictionary. , The classification boundary of each category in each representative vector of the multi-layer dictionary is added to the end representative vector which is located at the end of the multi-layer dictionary and is most similar to each of the learning vectors by giving information indicating the category to which the learning vector belongs. If the number of categories given as a result of the learning process and the learning process for learning By repeating the node extension process of extending a new node to which a plurality of representative vectors belong from the terminal representative vector that presents the maximum number, a multi-layer dictionary in which the above-mentioned nodes to which a plurality of representative vectors belong are chained in a tree It is characterized by forming.

According to the above structure, learning for learning the classification boundary of the category in each terminal representative vector of the multi-layer dictionary with respect to the initial state of the multi-layer dictionary consisting of a plurality of representative vectors obtained from the template and one root node. The processing and the node expansion processing for expanding the node from the terminal representative vector having the maximum number of categories that belong to the predetermined number or more are alternately repeated. As a result, the category boundary information based on the learning vector, which is lost when the template is created or when the multi-layer dictionary is constructed, is sharpened again by the learning process. That is, 100% of the category boundary information is stored in the database. Further, the layering is performed only on the region where many categories are densely arranged in the feature space, and the increase in the storage capacity due to the layering of the dictionary is suppressed as much as possible.

The invention according to claim 2 is, in the multi-layer dictionary creating method of the invention according to claim 1, whether or not there is no terminal representative vector whose number of categories given as a result of the learning is equal to or more than the predetermined number. The learning process and the node expansion process are alternately repeated until the total number of nodes reaches a preset number.

According to the above configuration, whether the category identification ability of all the end representative vectors of the multi-layer dictionary is above a certain level,
When the number of nodes reaches a preset number, further multilayering of the multilayer dictionary is stopped. In this way, a multi-layer dictionary in which high recognition accuracy and a small storage capacity are balanced is constructed.

Further, a multi-layer dictionary creating apparatus according to a third aspect of the present invention is a learning vector storage section for storing a feature vector of recognition target information as a learning vector, and a plurality of representative vectors obtained by clustering the learning vectors. , A multi-layer dictionary having a plurality of representative vectors obtained by clustering the templates and one root node to which the representative vectors belong as an initial state, and a multi-layer dictionary located at the end of the multi-layer dictionary. A learning unit that learns the identification boundary of the category in each terminal representative vector of the multi-layer dictionary by adding information indicating the category to which the learning vector belongs to the terminal representative vector most similar to each of the learning vectors, If the number of categories given as a result of the above learning is more than a predetermined number, From the terminal representative vector representing a number, a node expansion unit that expands a new node to which a plurality of representative vectors belong, the learning unit and the node expansion unit are controlled, and the learning is performed on the initial multi-layer dictionary. It is characterized by comprising a multi-layer dictionary creation control unit that forms a multi-layer dictionary in which the above-mentioned nodes to which a plurality of representative vectors belong are alternately formed by repeating node expansion.

According to the above configuration, the learning unit and the node expansion unit are controlled by the multi-layer dictionary creation control unit, and the initial state of the multi-layer dictionary composed of a plurality of representative vectors obtained from the template and one root node is set. Then, the learning process of learning the classification boundary of the category in each terminal representative vector and the node expansion process of expanding the node from the terminal representative vector having the maximum number of categories to which the number of categories belongs are alternately repeated. In this way, the category boundary information due to the learning vector lost at the time of creating the template or at the time of constructing the multi-layer dictionary is sharpened again by the learning process. That is, 100% of the category boundary information is stored in the database. Furthermore, the multi-layering is performed only on the region where many categories are dense in the feature space, and the increase in the storage capacity due to the multi-layering of the dictionary is suppressed as much as possible.

According to a fourth aspect of the present invention, in the multi-layer dictionary creating apparatus according to the third aspect of the present invention, the representative vector of the multi-layer dictionary is stored in the template storage unit of the template corresponding to the representative vector. It is characterized in that it is described by the storage position information indicating the position.

According to the above configuration, the storage capacity is smaller than when each representative vector of the multi-layer dictionary is described by the feature vector itself. In this way, the storage capacity of the multi-layer dictionary is further reduced.

According to a fifth aspect of the present invention, in the multi-layer dictionary creation device according to the third aspect, the learning section is
Information representing the category to which the learning vector of the search source belongs to the most similar representative vector searching means for searching the most representative terminal representative vector from the representative vectors of the multi-layer dictionary, and the searched terminal representative vector. It is characterized by including category information giving means for giving.

According to the above construction, the most similar representative vector searching means and the category information adding means appropriately perform the learning process for learning the identification boundary of the category in each terminal representative vector of the multi-layer dictionary.

According to a sixth aspect of the present invention, in the multi-layer dictionary creating apparatus according to the third aspect of the present invention, the node expansion unit is given as a result of the learning from the end representative vector of the multi-layer dictionary. A target terminating representative vector extracting means for extracting, as a target terminating representative vector, a terminating representative vector having a maximum number of categories which is equal to or larger than the predetermined number, and belongs to a category given to the extracted target terminating representative vector. A template selecting means for selecting all the templates whose representative vectors are the most similar representative vectors among the representative vectors of the above-mentioned multi-layer dictionary and clusters all the selected templates. Clustering means to obtain the average vector of each cluster by Also searches for a similar template, associates the searched template with the most similar representative vector search means that is the representative vector of the cluster, and the extended node to which a plurality of representative vectors belong to the target termination representative vector. The present invention is characterized by including node expansion means for giving the representative vector of each cluster obtained by the above-mentioned most similar representative vector searching means as the representative vector belonging to the expanded node.

According to the above configuration, the end-of-interest representative vector extracting means, the template selecting means, the clustering means,
By the most similar representative vector searching means and the node expanding means, the node expanding processing for newly expanding the node from the terminal representative vector having the maximum number and the number of categories to which it belongs is properly performed.

According to a seventh aspect of the present invention, in the multi-layer dictionary creation device according to the third aspect of the present invention, the multi-layer dictionary creation control section has the number of categories given as a result of the learning is the predetermined number or more. When the terminal representative vector does not exist, or when the total number of nodes reaches a preset number, the operation of the learning unit and the node expansion unit is stopped. .

According to the above configuration, the category discrimination capability of all terminal representative vectors of the multi-layer dictionary is controlled to a certain level or more by the control of the multi-layer dictionary creation control unit, or the total number of nodes is set to a preset number. Then, further multi-layering of the multi-layer dictionary is stopped. In this way, a multi-layer dictionary in which high recognition accuracy and a small storage capacity are balanced is constructed.

[0032]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described in detail below with reference to the embodiments shown in the drawings. FIG. 1 is a block diagram showing an example of the multi-layer dictionary creating device according to the present embodiment.
In the present embodiment, a multi-layer dictionary creation device installed in the print Japanese OCR will be described as an example. Here, the recognition target category of the main print Japanese OCR is 4000,
For the feature vector of each character, the circumscribing frame of the character is set to 8
It is a 64-dimensional vector obtained by equally dividing the norm and
Matching at the time of recognition is based on the similarity method. However, the number of categories to be recognized, the characteristic pattern, and the matching method are not limited to those described above, and any change may not impair the essence of the present invention.

An internal code is used as the category information. That is, both the input learning data and the dictionary data are internally coded (0 to Kanji code of JIS or EUC).
The value converted to the code value of 3999) is used.

The external storage device 1 is composed of a hard disk, a magneto-optical disk, a CD-ROM, or the like, and stores a learning vector in which the features of a large amount of learning data (learning characters) are expressed as vectors. Here, the learning vector is a collection of 4000 categories to be recognized as much as possible from the characters on the order of several millions in which variations of the font and the binarization level are added. Register group 2
Is controlled by the control unit 4 to write and read various data necessary for constructing the multi-layer dictionary. The buffer group 3 is controlled by the control unit 4 and stores the dictionary information of the multi-layer dictionary used at the time of recognition and the characteristic information of the input character. The control unit 4 is composed of a general-purpose CPU (central processing unit) or the like, controls the register group 2 to construct a multi-layer dictionary, and stores it in the buffer group 3.

The multi-layer dictionary created in this embodiment has a structure as shown in FIG. This multi-layered dictionary uses clustering of the above learning vectors.
The representative vector of each cluster when it is classified into (4000 × 50 in this embodiment) is the terminal representative vector.
(Representative vector leading to terminal node ●). Then, the terminal representative vector having no ability to separate from other terminal representative vectors is integrated to expand the node (◯), and one of the integrated terminal representative vectors is set as the representative vector.

Next, the detailed structure of the buffer group 3 will be described. The buffer group 3 is composed of a node buffer 5, a nearest neighbor dictionary buffer 6, an input feature buffer 7, a class average vector buffer 8 and a work memory 9. Hereinafter, each buffer will be described in detail.

FIG. 2 shows the configuration of the node buffer 5. This node buffer 5 is shown in FIG.
2 byte node number part 11 for storing the node number, a 2 byte representative vector number part 12 for storing the number of representative vectors belonging to the node, and a representative vector part for storing information of each representative vector. It is composed of 13. However, the representative vector unit 13 is provided for the maximum number of representative vectors per node. The maximum number of representative vectors per node is a value that can be arbitrarily determined for each recognition device, and is "32" in the main print Japanese OCR.

Further, the representative vector portion 13 is a 2-byte representative vector number portion 14 for storing the number in the nearest neighbor dictionary representing the representative vector.
If it is not the termination representative vector (representative vector that reaches the termination node ●) as shown in 1, the lower node number is stored, while if it is the termination representative vector, it is the hexadecimal number “ffff” indicating the termination. 2 bytes of the next node number portion 15 and 4000 (= 4000/8 bits) bytes of the category information portion 16 storing 4000 bits of binary data associated with 4000 internal codes of the recognition target category. Composed of. If the binary data of the category information section 16 is “1”,
The representative vector means that it belongs to the category associated with the bit (selects the category). It should be noted that the total number of the above nodes is a value that can be arbitrarily determined for each recognition device, and is "25" in the main print Japanese OCR.
6 ”.

FIG. 3 shows a structure of the nearest neighbor dictionary buffer 6. This nearest dictionary buffer 6 is
A 2-byte internal code portion 21 for storing the internal code of the character, a 2-byte population flag portion 22, which will be described in detail later,
A feature vector when the feature of the 64-dimensional mesh of the character is expressed as a vector (hereinafter referred to as the nearest neighbor dictionary vector)
Is composed of a 64-byte vector portion 16 for storing
The nearest-neighbor dictionary vector registered in the nearest-neighbor dictionary buffer 6 is the terminal representative vector of the multi-layer dictionary and is used when finally determining an unknown input character.
The number of characters registered in the nearest dictionary buffer 6 is a value that can be arbitrarily determined for each recognition device, and is “4000 (category) × 50” in the main print Japanese OCR.

The nearest neighbor dictionary stored in the nearest neighbor dictionary buffer 6 is 4000 × 50 characters obtained by previously clustering the learning vectors stored in the external storage device 1 into 4000 × 50 by some method. The standard pattern for minutes.

That is, the template is composed of the nearest neighbor dictionary vector, the template storage unit is composed of the nearest neighbor dictionary buffer 6, and the multilayer dictionary is composed of the node buffer 5.

FIG. 4 shows the configuration of the input feature buffer 7 and the class average vector buffer 8. The input feature buffer 7 is composed of a 2-byte internal code part 25 for storing the internal code of the input character and a 64-byte vector part 26 for storing the feature vector when the feature of the 64-dimensional mesh of the input character is expressed as a vector. To be done. In addition, the class average vector buffer 8 includes a 2-byte internal code portion 31 that stores the internal code of the average vector extracted by clustering (in the present embodiment, the k-average method), the average vector (hereinafter, the class). It is composed of a 64-byte vector part 32 for storing an average vector). The set of the internal code part 31 and the vector part 32 is registered for the maximum class average vector number "32" per node.

As shown in FIG. 5, the control unit 4 has a learning unit 51 for performing a learning process and a node expansion unit 52 for performing a node expansion process, which will be described in detail later.

The learning section 51 has a most similar representative vector searching means 55 and an internal code adding means 56 as the category information adding means. Then, the most similar representative vector searching means 55 causes the external storage device 1 to operate.
Then, the degree of similarity between each learning vector stored in (1) and each representative vector of the node buffer 5 is obtained, and the most similar representative vector is searched. Further, the internal code adding means 5
6, the internal code information (category information) of the learning vector is added to the representative vector most similar to the learning vector. In this way, the information of the category identification boundary, which is lost when the multi-layer dictionary is created, is learned again by the learning vector.

The node expansion unit 52 includes the end-of-interest representative vector extraction means 61, the nearest dictionary vector selection means 62,
It has a clustering means 63, a most similar representative vector searching means 64 and a node expanding means 65.

The noted end representative vector extracting means 61
Refers to the internal code information of each terminal representative vector in the node buffer 5 after learning, and extracts the terminal representative vector for expanding the node as a target terminal representative vector. The nearest neighbor dictionary vector selection unit 62 is the nearest neighbor dictionary vector belonging to the category to which the noted terminal representative vector belongs, and the terminal representative vector that is the most similar among all the terminal representative vectors of the node buffer 5 is noted. Select all nearest neighbor dictionary vectors that are terminal representative vectors. The clustering unit 63 clusters the selected group of nearest neighbor dictionary vectors into a predetermined number of clusters to obtain a class average vector.
The most similar representative vector searching means 64 searches for the nearest neighbor dictionary vector that is most similar to each class average vector. The node expansion means 65 is the end-of-interest representative vector extraction means 6
1, the nearest neighbor dictionary vector selecting means 62, the clustering means 63, and the most similar representative vector searching means 64 expand the node based on the processing result, and the information of the expanded node is written in the node buffer 5.

In the above configuration, the control unit 4 performs the learning process and the node expansion process, which will be described in detail later, on the multilayer dictionary (1 node / 32 representative vector) in the initial state when the total number of nodes is the above. The multi-layer dictionary is created by repeating the process until the total number of nodes reaches "256". The learning process and the node expansion process will be described below.

The learning process is the nearest neighbor dictionary vector (4000 × 50) stored in the nearest neighbor dictionary buffer 6.
This is a process for clarifying the information on the category boundary, which is lost when a multi-layer dictionary is created under a certain condition based on (characters), by learning with a learning vector stored in the external storage device 1. This processing is performed by the learning unit 51. FIG. 6 is a flowchart of the learning processing operation executed by the learning unit 51 of the control unit 4. Below, FIG.
The learning processing operation will be described in detail in accordance with.

This learning process can be executed whether the multi-layer dictionary to be processed is in the initial state of only 1 node / 32 representative vector or in the state of being constructed. The following description will be made assuming a multi-layer dictionary in which node expansion is performed up to a certain n (n ≦ 256).

In step S1, one learning vector is read from the external storage device 1 and the input feature buffer 7
Stored in. In step S2, the root node number "0" is stored in the node number register 35, which constitutes the register group 2 and stores the number of the target node.

In step S3, the number of representative vectors of the target node is read from the representative vector number section 12 of the target node designated by the node number register 35 in the node buffer 5. Then, it is stored in the representative vector number register 36 constituting the register group 2. In step S4, the first register forming the register group 2
The contents of the loop register 37 and the maximum similarity register 28 are initialized to "0".

In step S5, from the representative vector parts 13, 13, ... Of the target node in the node buffer 5,
The representative vector portion 13 having the number (order) indicated by the contents of the first loop register 37 is selected. Then, the number in the nearest neighbor dictionary of the representative vector is read from the representative vector number section 14 of the selected representative vector section 13 and stored in the representative vector number register 39 constituting the register group 2.

In step S6, the nearest neighbor dictionary vector stored in the vector unit 23 in the nearest neighbor dictionary buffer 6 is determined based on the number in the nearest neighbor dictionary of the representative vector stored in the representative vector number register 39. ,
It is read out as a feature vector of the representative vector.
After that, the similarity between the read representative vector and the learning vector of interest stored in the input feature buffer 7 is calculated, and the calculated value is stored in the similarity register 40 constituting the register group 2.

In step S7, the similarity register 40
And the content of the maximum similarity register 28 are compared. As a result, if the content of the similarity register 40 is equal to or larger than the content of the maximum similarity register 28, the process proceeds to step S8.
On the other hand, if not, the process proceeds to step S9. Step S8
Then, the selection number register 41 constituting the register group 2 is
The contents of the first loop register 37 are stored in. Furthermore, the maximum similarity register 3
8 is updated. In step S9, the first loop register 37 is incremented.

In step S10, the contents of the first loop register 37 and the contents of the representative vector number register 36 are compared. As a result, if the content of the first loop register 37 is smaller than the content of the representative vector number register 36, the process returns to the above step S5 and shifts to the processing relating to the next representative vector. Then, until it is determined that the content of the first loop register 37 is greater than or equal to the representative vector number register 36,
The representative vector of the node of interest is sequentially changed and the processing from step S5 to step S10 is repeated. As a result, the selection number register 41 stores the number of the representative vector that is most similar to the focused learning vector in the focused node. Further, the maximum similarity register 38 stores the maximum similarity at that time.

In step S11, the representative vector portion 13 indicated by the contents of the selection number register 41 in the target node indicated by the contents of the node number register 35 is selected from the node buffer 5. Then, “1” is stored in the bit associated with the same internal code as the internal code of the attention learning vector in the category information unit 16 of the representative vector unit 13. In this way, information that the representative vector of the attention node belongs to the same category as the attention learning vector is held.

In step S12, the next node number is read from the next node number portion 15 of the representative vector portion 13 in the node buffer 5, and the next node number register 4 is read.
Stored in 2. In step S13, the next node number stored in the next node number register 42 is a hexadecimal number indicating the end.
By determining whether or not it is (ffff), it is determined whether or not the next node is the terminal node. As a result, if it is not the terminal node, the process proceeds to step S14, while if it is the terminal node, the process proceeds to step S15.

In step S14, the contents of the next node number register 42 are stored in the node number register 35.
After that, the process returns to the step S3 to shift to the process for the next node. If it is determined in step S13 that the next node is the terminal node, the process proceeds to step S15. In step S15, the contents of the node number register 35 are stored in the terminal node number register 43 which constitutes the register group 2. Further, the contents of the selection number register 41 are stored in the termination selection number register 44 which constitutes the register group 2.

In step S16, it is determined whether or not the learning with all the learning vectors stored in the external storage device 1 is completed. As a result, if it is finished, the learning processing operation is finished. On the other hand, if it has not been completed, the process returns to step S1 to proceed to the learning by the next learning vector. Then, when it is determined that the learning with all the learning vectors is completed, the learning processing operation is ended.

That is, in the present embodiment, the above-mentioned steps S3 to S10, S12 to S14, and S16 constitute the above-mentioned most similar representative vector searching means, and the above step S11 constitutes the above-mentioned internal code adding means.

As described above, in the learning process according to the present embodiment, the node buffer 5 stores multiple layers of node information such as node numbers, representative vector numbers, representative vector numbers, and next node numbers. Regarding the dictionary, the terminal representative vector most similar to the learning vector is obtained, and the internal code (category information) of the similar learning vector is added to the obtained category information of the terminal representative vector. Therefore, the information of each category boundary lost at the time of creating the nearest neighbor dictionary and at the time of creating the multi-layer dictionary is sharpened again by the learning vector.

The learning is performed using all learning vectors stored in the external storage device 1. Therefore, no matter how many learning vectors there are, all learning vectors must be selected by any of the terminal representative vectors, and the multilayer dictionary passage rate (the learning vector selected by the terminal representative vector) is It will definitely be 100%.
That is, the multilayer dictionary according to the present embodiment stores 100% of the category boundary information based on a large amount of learning vectors.

Next, the node expansion processing will be described. The node expansion process is a process of expanding a new node in a portion where many categories are dense in the feature space, which is a termination representative vector that does not have a category separation ability from other termination representative vectors. 8 to FIG.
9 is a flowchart of a node expansion processing operation executed by the node expansion unit 52 of the control unit 4. Below, FIG.
The node expansion processing operation will be described in detail with reference to FIG.

The node expansion processing in this embodiment is performed based on the contents of the category information section 16 of each terminal representative vector created by the above-described learning processing. Therefore, as a premise, the learning process for the above-mentioned multilayer dictionary must be completed. Therefore, the following node expansion processing operation will be described assuming that a multi-layer dictionary having a certain n (n ≦ 256) nodes has been created and the learning processing has been completed.

In step S21, the node buffer 5 is searched and the current number "n" of nodes is counted. Then, the total node number register 45 constituting the register group 2
The current number of nodes “n” obtained above is stored in the. In step S22, it is determined whether the content "n" of the total node number register 45 is smaller than the set maximum number of nodes ("256" in the present embodiment). As a result, if it is smaller than the maximum number of nodes "256", the process proceeds to step S23, and if not, the node expansion processing operation ends.

In step S23, the category information part 16 of all terminal representative vectors in the node buffer 5 is referred to based on the contents of all the next node number parts 15 of the node buffer 5. The number of bits in which "1" is stored in the category information section 16 (the number of categories to which it belongs)
Is a set threshold value α (in the present embodiment, α =
200) or more and the terminal end representative vector having the maximum number of bits is extracted and set as the target terminal end representative vector. In this way, when the target termination representative vector is extracted, the contents of the node number portion 11 of the node to which the target termination representative vector in the node buffer 5 belongs is stored in the target termination node number register 46 that constitutes the register group 2. It In addition, the number of the target termination representative vector is the number of the target termination vector number register 4 forming the register group 2.
Stored in 7. On the other hand, if it has not been extracted, the hexadecimal number (ffff) indicating that it has not been extracted is stored in the target terminal node number register 46.

In step S24, the contents of the target terminal node number register 46 are referred to, and it is determined whether or not the target terminal representative vector is extracted in step S23. As a result, if extracted, the process proceeds to step S25, while if not extracted, the node expansion processing operation ends. In step S25, the nearest dictionary buffer 6
The content of the entire population flag section 22 in is cleared to "0". In step S26, the content of the nearest neighbor dictionary counter 48 constituting the register group 2 is cleared to "0". Further, the number of nearest neighbor dictionary vectors stored in the nearest neighbor dictionary buffer 6 is counted, and this count value is set in the nearest neighbor dictionary number register 49 forming the register group 2.

In step S27, the contents of the internal code part 21 and the vector part 23 of the number (order) indicated by the nearest neighbor dictionary counter 48 in the nearest dictionary buffer 6 become the internal code part 25 and the vector of the input feature buffer 7. It is stored in the unit 26.

At step S28, the contents of the noticed end node number register 46 and the noticed end vector number register 4
Based on the contents of No. 7, the category information unit 16 of the target termination representative vector in the node buffer 5 is referred to. Then, it is stored in the input feature buffer 7 by determining whether or not "1" is stored in the bit associated with the internal code stored in the internal code portion 25 of the input feature buffer 7. It is determined whether or not the nearest dictionary vector (character) is selected as the target end representative vector (that is, whether it belongs to the category to which the target end representative vector belongs). As a result, it is selected
If it belongs (belongs to), the process proceeds to step S29. If it is not selected (belongs to), the process proceeds to step S32.

In step S29, the input feature buffer 7 is input.
The nearest neighbor dictionary vector stored in is regarded as a learning vector, and steps S2 to S15 in the learning processing operation shown in FIGS. 6 and 7 are performed.

In step S30, as a result of the learning in step S29, the contents of the end node number register 43 and end selection number register 44 are changed to the end node number register 46 of interest and the end vector number register 4 of interest.
It is determined whether or not the contents of 7 match. In this way, whether or not the nearest neighbor dictionary vector stored in the input feature buffer 7 is terminated by the end-of-interest representative vector (that is, is it most similar to the end-of-interest representative vector among all the end-of-representative vectors of the multi-layer dictionary)? Whether or not) is determined. As a result, if the terminal is the terminal (most similar to the target terminal representative vector), the process proceeds to step S31, while if not terminal (most similar to the terminal representative vector other than the target terminal representative vector). It proceeds to step S32. In step S31, "1" is stored in the population flag unit 22 of the number indicated by the nearest neighbor dictionary counter 48 in the nearest neighbor dictionary buffer 6.

In step S32, the nearest neighbor dictionary counter 48 is incremented. In step S33, the contents of the nearest neighbor dictionary counter 48 and the nearest neighbor dictionary number register 4
The contents of 9 are compared. As a result, if the content of the nearest-neighbor dictionary counter 48 is smaller than the content of the nearest-neighbor dictionary number register 49, the process returns to the step S27 to shift to the process for the next nearest-neighbor dictionary vector. Then, until it is determined that the content of the nearest-neighbor dictionary counter 48 is equal to or larger than the content of the nearest-neighbor dictionary number register, the nearest-neighbor dictionary vector is sequentially changed and the processing from step S27 to step S33 is repeated. As a result, the population flag unit 22 of the nearest-neighbor dictionary vector that belongs to the same category as the noticed end representative vector and ends at the noticed end representative vector.
"1" will be stored in. When the content of the nearest neighbor dictionary counter 48 becomes equal to or larger than the content of the nearest neighbor dictionary number register 49, the process proceeds to step S34.

In step S34, all the population flag parts 22 in the nearest neighbor dictionary buffer 6 are referred to, and all the nearest neighbor dictionary vectors whose contents are "1" are read out. Then, the cluster of the nearest dictionary vectors is subjected to clustering in the maximum “32” class by the k-means method. Then, the obtained average vector of 32 classes is stored in the vector unit 32 of the class average vector buffer 8.

At step S35, the number of all nodes is incremented by 1 due to the subsequent node expansion. Therefore, the content of the total node number register 45 is incremented. In step S36, it is determined whether or not the content of the total node number register 45 is less than or equal to the set maximum number of nodes "256". As a result, if the maximum number of nodes is "256" or less, the process proceeds to step S37, and if not, the new node expansion processing operation ends.

At step S37, the contents of the second loop register 50 constituting the register group 2 are initialized to "0". In step S38, the contents of the first loop register 37 and the maximum similarity register 38 are initialized to "0".

In step S 39, the class average vector stored in the vector portion 32 of the number (order) indicated by the second loop register 50 in the class average vector buffer 8 and the first loop in the nearest neighbor dictionary buffer 6 The degree of similarity with the nearest neighbor dictionary vector stored in the vector part 23 of the number indicated by the register 37 is calculated. Then, the obtained similarity is stored in the similarity register 40.

At step S40, the similarity register 40
It is determined whether or not the content of is greater than or equal to the content of the maximum similarity register 38. As a result, if above, step S41
On the other hand, if not, the process proceeds to step S42. In step S41, the contents of the first loop register 37 are stored in the selection number register 41. Further, the contents of the maximum similarity register 38 are updated with the contents of the similarity register 40.

In step S42, the first loop register 37 is incremented. In step S43, the contents of the first loop register 37 is the nearest neighbor dictionary number register 4
It is determined whether or not it is smaller than the content of 9. as a result,
If it is smaller, the process returns to the step S39 to shift to the similarity calculation using the next nearest neighbor dictionary vector. The content of the first loop register 37 is the nearest dictionary number register 4
Until the content of 9 or more is determined, the nearest neighbor dictionary vector is sequentially changed and the steps S39 to S39 are performed.
The processes up to 43 are repeated. As a result, the selection number register 41 stores the number of the nearest-neighbor dictionary vector having the maximum similarity with the target class average vector. Also,
The maximum similarity degree at that time is stored in the maximum similarity degree register 28. When the content of the first loop register 37 is equal to or larger than the content of the nearest neighbor dictionary number register 49, the process proceeds to step S44.

In step S44, the following processing is performed on the node buffer 5. That is, in the representative vector number portion 14 of the representative vector portion 13 designated by the contents of the second loop register 50 in the node designated by (contents of the total number of nodes register 45-1) (hereinafter referred to as an extended node), The content of the selection number register 41 (the number of the most similar nearest neighbor dictionary vector) is stored. Further, the hexadecimal number “ffff” is stored in the next node number portion 15 in the representative vector portion 13 of the extension node. In addition, the representative vector unit 13 specified by the contents of the second loop register 50 of the node of interest (upper layer node of the expansion node) specified by the contents of the end node number register 46 of interest
The content of the category information section 16 in is cleared to “0”.

In step S45, the second loop register 50 is incremented. In step S46, the content of the second loop register 50 is changed to the maximum class number "32".
It is determined whether or not it is smaller. As a result, if it is smaller, the process returns to the step S38 and shifts to the similarity calculation using the next class average vector. Then, until it is determined that the content of the second loop register 50 is equal to or larger than the maximum number of classes "32", the class average vector is sequentially changed and the processing from step S38 to step S46 is repeated. As a result, the representative vector number part 14 in each representative vector part 13 of the newly expanded extended node stores the number in the nearest neighbor dictionary representing this representative vector. On the other hand, the hexadecimal number "ffff" is stored in each secondary node number section 15.
Is stored. Further, the category information part 16 of all the representative vectors of the upper layer nodes of the expansion node is cleared to "0". When the content of the second loop register 50 becomes equal to or larger than the maximum class number "32", the process proceeds to step S47.

In step S47, the following processing is performed on the node buffer 5. (Total node number register 4
(Content-1 of total node number register 45) is stored in the node number portion 11 of the extended node designated by the content-1 of 5). Further, the maximum number of classes “32” is stored in the representative vector copy number 12 of the expansion node. Further, in the next node number part 15 of the representative vector designated by the contents of the noticed end vector number register 47 in the noticed node (upper node of the expansion node) designated by the contents of the noticed end node number register 46, (all nodes The content of the number register 45-1) (node number of lower layer node) is stored. After that, the process returns to step S22.

Thereafter, the above steps S22 to S47
The processes up to are repeated, and when it is determined in step S22 that the number of nodes "n" is equal to or larger than the maximum number of nodes "256", the node expansion processing operation is ended.

As a result, the node number parts 11 of all nodes including the extended node in the node buffer 5 are
The node numbers 0 to (N-1) are stored. further,
The maximum number of classes “32” is stored in each of the N representative vector number parts 12. Further, the numbers of the most similar nearest neighbor dictionary vectors are stored in the 32 representative vector number parts 14 under the respective node numbers. Then, in the next node number portion 15 of each of the 32 representative vectors of the upper layer node of the node expanded this time, if the representative vector is a terminal representative vector, the hexadecimal number "ff
ff ”is stored, while it is not the termination representative vector (total node number“ N ”−1 (that is, node number of lower layer node)) is stored. Also, the category information section 16
Has been cleared.

Therefore, when the node expansion processing operation is completed, all the category information parts 16 in the node buffer 5 are cleared, and the internal code of each representative vector is undetermined. Therefore, if the above learning process is performed in this state, the internal code information of the learning vector that is most similar to each representative vector is stored in the category information unit 16, and the internal code is determined.

Next, the node expansion processing in the case of constructing the multi-layer dictionary as shown in FIG. 11 will be described with a specific example. Here, in order to simplify the story, the number of learning vectors stored in the external storage device 1 is 50, the number of nearest neighbor dictionary vectors of the nearest neighbor dictionary created based on this learning vector is 22, and The maximum number of nodes is 4, and the threshold value α is 4.

First, a nearest neighbor dictionary having 22 nearest neighbor dictionary vectors is created from 50 learning vectors, for example, by the nearest neighbor method, and stored in the nearest neighbor dictionary buffer 6. Then, the nearest neighbor dictionary is clustered into 4 clusters to create an initial state of a multi-layer dictionary composed of 1 node / 4 representative vectors as shown in FIG. Here, it is assumed that the above initial state is created as follows. One nearest neighbor dictionary vector belongs to the first cluster, and the nearest neighbor dictionary vector most similar to the average vector among them (that is, the only one nearest neighbor dictionary vector) is set as the representative vector (0). Seven nearest neighbor dictionary vectors belong to the second cluster, and the nearest neighbor dictionary vector that is most similar to the average vector is defined as the representative vector (1). One nearest neighbor dictionary vector belongs to the third cluster, and this nearest neighbor dictionary vector is designated as a representative vector (2). 13 for the 4th cluster
The nearest neighbor dictionary vector belongs to the nearest neighbor dictionary vector, and the nearest neighbor dictionary vector most similar to the average vector is the representative vector.
(3) Then, by performing the learning process on the multi-layer dictionary in the initial state, the contents of the node buffer 5 are set as shown in FIG.

Next, the node expansion process is performed on the initial multi-layer dictionary. Where the above representative vector
Since the number of nearest neighbor dictionary vectors belonging to the same cluster as (3) is the largest, the number of “1” stored in the category information unit 16 by the above learning is the largest as well. Become. Therefore, as the end-of-interest representative vector belonging to the most category with the threshold value α = 4 or more,
The representative vector (3) is extracted.
... Step S21 to Step S23

While sequentially changing the nearest neighbor dictionary vector, the nearest neighbor dictionary vector belonging to the category (D) represented by the binary data (D) of the target termination representative vector (representative vector (3)) is a representative vector. (0) -representative vector
In (3), the most similar representative vector is searched for all the nearest neighbor dictionary vectors that are the target termination representative vector (3), and a group of nearest neighbor dictionary vectors is formed. Then, this nearest neighbor dictionary vector group is clustered into 4 clusters by the k-averaging method to obtain the average vector of each class. ... Step S24 to Step S34

In this way, clustering is performed by collecting the nearest neighbor dictionary vectors belonging to the same category as the target termination representative vector (3) having a low category discriminating ability.
However, as described above, when the nearest neighbor dictionary is created from the learning vector, the category boundary information by the learning vector is lost. Therefore, based on the information in the nearest dictionary buffer 6, the target termination representative vector (3)
The nearest neighbor dictionary vector determined to belong to the category (D) to which is belongs is the terminal end representative vector of interest when the maximum likelihood path of the multilayer dictionary is searched with the nearest neighbor dictionary vector.
It does not always lead to (3). Therefore, the maximum likelihood route is actually searched by the multilayer dictionary which has already learned the category boundary, and only the nearest neighbor dictionary vectors reaching the target terminal representative vector (3) are collected.

While sequentially changing the nearest neighbor dictionary vector, the nearest neighbor dictionary vector that is most similar to the average vector of the first cluster is searched. Then, the newly expanded node (second node (1)) is associated with the target termination representative vector (representative vector (3)), and is searched as the first representative vector (0) of the expanded node. In addition to giving the nearest neighbor dictionary vector, information (hexadecimal number ffff) indicating that it is a terminal representative vector is added. Hereinafter, the above-described processing is repeated while sequentially changing the class average vector to associate the nearest neighbor dictionary vector with each of the four representative vectors of the expansion node.
... Step S35 to Step S47

As a result, the contents of the node buffer 5 are updated as shown in FIG. Figure 14 shows the root node
“1” is stored in the next node number portion 15 of the representative vector (3) belonging to (0), and the representative vector (3) is not the terminal representative vector but the node (1) is extended to the lower layer. . Then, it is shown that four termination representative vectors belong to this extended node (1). That is, the above-described node expansion processing forms a two-layer multi-layer dictionary composed of two nodes / 7 representative vectors as shown in FIG. Note that the representative vectors 0, 1, and 3 of the expansion node (1) are associated with one nearest neighbor dictionary vector, and the representative vector 2 is associated with ten nearest neighbor dictionary vectors.

Then, learning processing is performed on the above-mentioned multilayer dictionary to determine the internal code of each representative vector of the extended node (1). After that, by repeating the node expansion process and the learning process, the multi-layer dictionary as shown in FIG. 11 is constructed.

That is, in the present embodiment, the step S23 constitutes the noted end representative vector extracting means, the steps S27 to S33 constitute the nearest neighbor dictionary vector selecting means, and the step S34 describes the above. A clustering means is constituted, and the most similar representative vector searching means is constituted by the steps S38 to S43, S45, S46, and the step S44.
The step S47 constitutes the node expanding means.

As described above, in the node expansion process of this embodiment, as a result of the learning process described above, the number of categories to which it belongs (the number of bits of "1" in the category information section 16) is the threshold value α = 200 or more. , And the terminal representative vector having the maximum value is set as the target terminal representative vector. And
The nearest neighbor dictionary vectors that belong to the category to which this end-of-interest representative vector belongs and that end at the end-of-interest representative vector are collected, and the maximum number of "32" classes are obtained for this nearest neighbor dictionary vector group by the k-means method. Perform clustering. Then, the nearest neighbor dictionary vector that is most similar to the class average vector is searched for each class, and the searched nearest neighbor dictionary vector is set as one terminal representative vector belonging to the node extended from the target terminal representative vector. is there.

As described above, the node expansion is performed only on the terminal representative vector having more than 200 categories and having a low category identification capability, so that the node hierarchy of the multi-layer dictionary is identified as shown in FIG. Will vary depending on the separation capacity of. In other words, only the portion where many categories are dense in the feature space becomes more multilayered. Therefore, in the present embodiment, as compared with the case where the multi-layer dictionary disclosed in Japanese Patent Laid-Open No. 63-263590 is used, clustering is uniformly performed to uniformly construct a multi-layer dictionary having the same node hierarchy. The storage capacity of the dictionary can be reduced.

For example, in Japanese Unexamined Patent Publication No. 63-263590, when a multi-layer dictionary in which the number of representative vectors v per node is 4 and the number of layers r is 5 corresponding to the multi-layer dictionary shown in FIG. Dictionary memory capacity Is. On the other hand, in the dictionary memory capacity of the multi-layer dictionary shown in FIG. 11, when the number of non-terminal nodes is n, Is. Therefore, in the case of the present embodiment and in Japanese Patent Laid-Open No.
Assuming that the storage capacity s per one representative vector is the same as in Japanese Patent Publication No. 263590, the dictionary memory capacity in the case of the present embodiment is approximately 1/35 of the dictionary memory capacity in Japanese Patent Laid-Open No. 63-263590. be able to.

Further, in the present embodiment, the feature vector of each representative vector in the multi-layer dictionary constructed in the node buffer 5 is stored with the number of the corresponding nearest neighbor dictionary vector. Therefore, as compared with the case of Japanese Patent Laid-Open No. 63-263590, in which the representative vector of each layer is stored in the form of a feature vector such as a 64-dimensional vector, the storage capacity s per representative vector is 1/10.
It can be: Therefore, it is possible to significantly reduce the dictionary memory capacity in addition to making the portion where many categories are densely arranged in the above-mentioned feature space more multilayered.

By the way, the multilayer dictionary constructed according to the present embodiment as described above is used as follows when recognizing an unknown character. First, the similarity between the input unknown character feature vector and each representative vector belonging to the root node is calculated. In this case, the similarity calculation is based on the number of the nearest neighbor dictionary vector stored in the representative vector number part 14 of the corresponding representative vector part 13 of the node buffer 5 and the vector part of the corresponding number of the nearest neighbor dictionary buffer 6. The feature vector is read from 23, and this feature vector (nearest neighbor dictionary vector) and the feature vector of the unknown character are used.

As a result, when the representative vector exhibiting the highest similarity is not the terminal representative vector, a plurality of representative vectors belonging to the node extended from the representative vector exhibiting the highest similarity and the feature vector of the unknown character To calculate the similarity.

The above processing is repeated until the representative vector exhibiting the highest degree of similarity becomes the terminal representative vector. Then, based on the terminal representative vector that is most similar to the obtained feature vector of the unknown character, the category information unit 1 of the corresponding representative vector unit 13 in the node buffer 5 described above.
The internal code corresponding to the category information stored in 6 is obtained. Then, when there is one obtained internal code, the category represented by this internal code is set as the recognition result.

On the other hand, when there are a plurality of obtained internal codes, the internal code section 21 of the nearest neighbor dictionary buffer 6 is referred to and the feature vector (nearest neighbor dictionary vector) is calculated from the vector section 23 of the corresponding inner code. Are all read. Then, the similarity between the read feature vectors and the feature vector of the unknown character is calculated to obtain the inner code of the nearest neighbor dictionary vector having the highest similarity, and the category represented by this inner code is recognized. The result.

The above-described processing may be performed by using a representative vector having a similarity rank between the feature vector of the unknown character and the representative vector up to a predetermined rank as a similar representative vector candidate.

In the above embodiment, the feature vector of each representative vector in the multi-layer dictionary is stored as the number of the corresponding nearest neighbor dictionary vector. However,
The present invention is not necessarily limited to this, and even if each feature vector is stored, the memory of the dictionary memory can be stored by layering only the part where many categories are dense in the feature space. The capacity can be reduced.
Further, in the above-described embodiment, the multi-layer dictionary creating device installed in the print Japanese OCR is described as an example. However, it goes without saying that the present invention is not limited to this, and can be applied to handwritten Japanese OCR, English OCR, and the like. Further, the present invention can be applied not only to the recognition dictionary of the character recognition device but also to the recognition dictionary of the voice recognition device.

Further, the procedure for creating the above-mentioned multi-layer dictionary is programmed and recorded in a computer-readable recording medium such as a floppy disk or a CDROM, and if necessary, installed in the memory and the control unit 4 is operated.
It can be used in.

[0105]

As is apparent from the above, the multi-layer dictionary creating method of the invention according to claim 1 clusters the learning vector which is the feature vector of the recognition target information to create the template, and clusters the template. A learning process for creating an initial state of a multi-layer dictionary composed of a plurality of representative vectors and one root node, and learning the category identification boundary in each representative vector with respect to the multi-layer dictionary in the initial state, and the number of categories to which it belongs is predetermined The learning process is performed by alternately repeating node expansion processing that expands a node from the terminal representative vector that is a number greater than or equal to the maximum number to form a multi-layer dictionary in which nodes to which a plurality of representative vectors belong are chained in a tree shape. Re-enhances the category boundary information lost when creating the above template or building a multi-level dictionary. Rukoto can, can build a multi-dictionary where you saved the category boundary information on a database of 100%. Further, by the node expansion processing, multi-layering can be performed only on an area where many categories are dense in the feature space, and an increase in storage capacity due to multi-layering of the dictionary can be suppressed as much as possible.

In the multi-layer dictionary creating method according to the second aspect of the present invention, whether there is no terminal representative vector whose number of categories added as a result of the learning is the predetermined number or more,
Since the learning process and the node expansion process are alternately repeated until the total number of nodes reaches a preset number, a multi-layer dictionary in which the high recognition rate and the small storage capacity are appropriately balanced is constructed. it can.

The multi-layer dictionary creation device of the invention according to claim 3 stores a plurality of templates obtained by clustering the learning vectors stored in the learning vector storage unit in the template storage unit, and clusters the templates. Under the control of the learning unit and the node expansion unit by the multi-layer dictionary creation control unit, the category in each end representative vector is applied to the multi-layer dictionary in the initial state composed of a plurality of representative vectors and one root node obtained by The learning of the identification boundary and the node expansion for expanding a new node from the terminal representative vector having the maximum number and the number of categories belonging to the predetermined number or more are alternately repeated. You can re-clear the category boundary information that was lost when creating The category boundary information of 100%
You can build a saved multi-level dictionary. Further, by the node expansion, the multi-layering can be performed only on the region where many categories are dense in the feature space, and the increase in the storage capacity due to the multi-layering of the dictionary can be suppressed as much as possible.

In the multi-layer dictionary creating apparatus according to the fourth aspect of the present invention, the representative vector of the multi-layer dictionary is described by the storage position information indicating the storage position of the template corresponding to the representative vector on the template storage unit. Therefore, the storage capacity can be reduced as compared with the case where the feature vector itself is described. Therefore, according to the present invention, the storage capacity of the multi-layer dictionary can be further reduced.

Further, the learning unit in the multi-layer dictionary creating device according to the fifth aspect of the present invention, the learning unit searches the representative vector of the multi-layer dictionary for a terminal representative vector most similar to each learning vector. And learning means for learning the identification boundary of the category in each terminal representative vector of the multi-layer dictionary, since it includes category information adding means for adding information indicating the category to which the learning vector of the search source belongs to the searched terminal representative vector. The processing can be performed accurately.

Further, the node extension unit in the multi-layer dictionary creation device according to the sixth aspect of the present invention extracts the noticeable end point representative vector for extracting the noticeable end point representative vector having the maximum number of categories that belong to the predetermined number or more. Means, template selecting means for selecting all templates that belong to the same category as the noticed end representative vector and terminates at the noticed end representative vector, and clustering for clustering all the selected templates Means, and a most similar representative vector searching means for searching a template most similar to the average vector of each cluster to be a representative vector of the cluster,
Since the extension node is associated with the target termination representative vector and the node extension means for giving the representative vector of each cluster obtained by the most similar representative vector search means as the representative vector belonging to this extension node is included, the number of categories to which it belongs It is possible to accurately perform node expansion processing for newly expanding a node from the termination representative vector that is a predetermined number or more and is the maximum number.

Further, the multi-layer dictionary creation control unit in the multi-layer dictionary creation apparatus according to the seventh aspect of the present invention determines whether or not there is no terminal representative vector whose number of categories added as a result of the learning is the predetermined number or more. Alternatively, when the total number of nodes reaches a preset number, the operations of the learning unit and the node expansion unit are stopped, so that the high recognition rate and the small storage capacity are properly balanced. Can build a dictionary.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a multi-layer dictionary creating device of the present invention.

FIG. 2 is a diagram showing a configuration of a node buffer in FIG.

FIG. 3 is a diagram showing a configuration of a nearest neighbor dictionary buffer in FIG.

FIG. 4 is a diagram showing configurations of an input feature buffer and a class average vector buffer in FIG.

5 is a detailed block diagram of a control unit in FIG. 1. FIG.

6 is a flowchart of a learning processing operation performed by a learning unit in FIG.

FIG. 7 is a flowchart of a learning processing operation subsequent to FIG.

8 is a flowchart of a node expansion processing operation performed by a node expansion unit in FIG.

9 is a flowchart of a node expansion processing operation subsequent to FIG.

FIG. 10 is a flowchart of node expansion processing operation subsequent to FIG. 9;

FIG. 11 is an explanatory diagram of a structure of a multi-layer dictionary.

12 is a diagram showing an initial state of the multi-layer dictionary shown in FIG.

FIG. 13 is a diagram showing the contents of a node buffer after the learning process is performed on the initial state of FIG.

14 is a diagram showing the contents of the node buffer after the learning process and the node expansion process have been performed once for the initial state of FIG.

15 is a diagram showing a multi-layer dictionary structure after a node expansion process is performed once with respect to the initial state of FIG.

FIG. 16 is a relationship diagram of neighborhood categories when a dictionary based on the composite similarity method is used.

FIG. 17 is a relationship diagram of neighborhood categories when a dictionary according to the nearest neighbor method is used.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... External storage device, 2 ... Register group, 3 ... Buffer group, 4 ... Control part, 5 ... Node buffer, 6 ... Nearest neighbor dictionary buffer, 7 ... Input feature buffer,
8 ... Class average vector buffer, 11 ... Node number part, 12 ... Representative vector number part, 13 ...
Representative vector part, 14 ... Representative vector number part, 15 ... Next node number part, 16 ... Category information part, 21, 25, 31 ... Internal code part,
22 ... Population flag part, 23, 26, 32 ... Vector part,
51 ... Learning unit, 52 ... Node expansion unit, 55, 6
4 ... Maximum similar representative vector searching means, 56 ... Internal code adding means, 61 ... Target terminal representative vector extracting means, 62 ... Nearest neighbor dictionary vector selecting means, 63 ... Clustering means, 65 ... Node expanding means.

─────────────────────────────────────────────────── ─── Continuation of front page (56) Reference JP-A-60-126784 (JP, A) JP-A-63-263590 (JP, A) JP-A-58-105387 (JP, A) JP-A-6- 76061 (JP, A) JP-A-10-11543 (JP, A) Akiyoshi Ito, et al., "Dictionary Construction Method for Registering Character Types in Multiple Clusters in Hierarchical Printing Kanji Recognition System", IEICE Transactions , Japan, The Institute of Electronics, Information and Communication Engineers, June 25, 1995, Vol. J78-D-II, No. 6, pp. 896-905 (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06N 1/00-7/08 G06K 9/00-9/82 G06T 7/00-7/60 G10L 3/00-3 / 02 G10L 5/06 G10L 7/08 G10L 9/00-9/20 G06F 17/30 JST file (JOIS) CSDB (Japan Patent Office)

Claims (7)

(57) [Claims] 1. A feature vector of recognition target information is prepared as a learning vector, a plurality of representative vectors obtained by clustering the learning vectors is used as a template, and a plurality of representative vectors obtained by clustering the template and this One root node to which the representative vector belongs is set as an initial state of the multi-layer dictionary, and is located at the end of the multi-layer dictionary with respect to the multi-layer dictionary in the initial state, and is a terminal representative vector most similar to each learning vector, A learning process of learning the identification boundary of the category in each terminal representative vector of the multi-layer dictionary by adding information indicating the category to which the learning vector belongs, and the number of categories added as a result of the learning is a predetermined number or more. From the terminal representative vector that exhibits the maximum number of A node expansion processing for expanding the Tana nodes alternately repeated, multilayer dictionary creation method characterized in that the node in which a plurality of representative vectors belong to form a multilayer dictionary chained dendritic. 2. The method for creating a multi-layer dictionary according to claim 1, wherein there is no terminal representative vector whose number of categories added as a result of the learning is the predetermined number or more, or the total number of nodes is preset. A multi-layer dictionary creating method, characterized in that the learning process and the node expansion process are alternately repeated up to a number of times. 3. A learning vector storage unit that stores a feature vector of recognition target information as a learning vector, a template storage unit that stores a plurality of representative vectors obtained by clustering the learning vectors as a template, and the template A multi-layer dictionary having a plurality of representative vectors obtained by clustering and one root node to which the representative vector belongs as an initial state, and a terminal representative located at the end of the multi-layer dictionary and most similar to each of the learning vectors A learning unit that learns the identification boundary of the category in each terminal representative vector of the multilayer dictionary by adding information indicating the category to which the learning vector belongs to the vector, and the number of categories given as a result of the learning is a predetermined number. From the end representative vector that is the above and presents the maximum number, multiple By controlling the node expansion unit for expanding a new node to which the representative vector belongs, the learning unit and the node expansion unit, the learning and the node expansion are alternately repeated for the multilayer dictionary in the initial state, A multi-layer dictionary creation device comprising a multi-layer dictionary creation control unit that forms a multi-layer dictionary in which the nodes to which a plurality of representative vectors belong are chained in a tree shape. 4. The multi-layer dictionary creation device according to claim 3, wherein the representative vector of the multi-layer dictionary is described by storage position information indicating a storage position on the template storage unit of a template corresponding to the representative vector. A multi-layer dictionary creation device characterized in that 5. The multi-layer dictionary creation device according to claim 3, wherein the learning unit searches the representative vector of the multi-layer dictionary for a terminal representative vector that is most similar to each learning vector. And a category information adding means for adding information indicating a category to which the learning vector of the search source belongs to the searched terminal representative vector. 6. The multi-layer dictionary creation device according to claim 3, wherein the node expansion unit determines that the number of categories given as a result of the learning is greater than or equal to the predetermined number from the end representative vectors of the multi-layer dictionary. And a template belonging to a category assigned to the extracted target termination representative vector, which is a template of the multi-layer dictionary. Template selecting means for selecting all the templates whose representative vector most similar among the representative vectors is the target termination representative vector, and clustering all the selected templates to obtain an average vector of each cluster And the template most similar to the mean vector of each cluster above The most similar representative vector search means that searches for the template as the representative vector of the cluster and the extension node to which the plurality of representative vectors belong is associated with the end-of-interest representative vector, and the representative that belongs to this extension node is searched. A multi-layer dictionary creating device, characterized in that it includes node expansion means for giving a representative vector of each cluster obtained by the above-mentioned most similar representative vector searching means as a vector. 7. The multi-layer dictionary creation device according to claim 3, wherein the multi-layer dictionary creation control unit does not exist when there is no terminal representative vector whose number of categories added as a result of the learning is the predetermined number or more. Alternatively, when the total number of nodes reaches a preset number, the operation of the learning unit and the node expansion unit is stopped, so that the multi-layer dictionary creation device is characterized.