JPH0467277A - Pattern recognition dictionary generating device - Google Patents

Abstract

PURPOSE:To easily execute the additional registration of a sub-category, while securing the improvement of a recognition rate by deciding the propriety of the additional registration to a recognition dictionary under the condition of a decrease of the number of erroneous recognition patterns. CONSTITUTION:By an erroneous recognition pattern classifying device 5, an erroneous recognition pattern assembly is classified into each separate assembly by combination of an exact category name and an erroneous recognition category name. From in the erroneous recognition pattern assembly being an assembly of the erroneous recognition category name classified in such a way, one piece each is registered temporarily in a recognition dictionary 2 as a sub- category candidate in a suitable sequence by a sub-category candidate selecting device 6. In the case the erroneous recognition pattern decreases as a whole, the sub-category candidate is registered additionally and formally to the recognition dictionary, and the erroneous recognition pattern assembly is updated. On the other hand, in the case the erroneous recognition pattern increases as a whole, with regard to the next sub-category candidate, as well, whether the number of erroneous recognition patterns decreases or not is decided in the same way.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a pattern recognition dictionary creation device, and is particularly applicable to computer-based pattern recognition, such as recognition within multiple categories, in the field of character/speech recognition and image recognition. A general-purpose method that provides additional registration of subcategories that compensates for the improvement in recognition rate by setting a condition that the number of misrecognized patterns is reduced by additional registration of subcategories when performing the process based on the distance or similarity between patterns from the representative pattern of the category. The present invention relates to a pattern recognition dictionary creation device.

[Conventional technology]

When performing pattern recognition within multiple categories using a computer,
A method of constructing an identification tree based on a combination of judgment logic for identifying between categories, or a pattern matching method using a representative pattern registration dictionary, in which judgment is made based on the distance or degree of similarity between patterns from the representative pattern of each category, are known. ing.

Also known are the hierarchical clustering method and the LBG method, which are methods for determining subcategories by dividing a set of learning patterns in the same category into patterns with a small distance between them or patterns with a large degree of similarity.

Furthermore, there is also a method of selecting a core pattern of a subcategory from a set of misrecognition patterns for each category and registering it in a recognition dictionary.

[Problem to be solved by the invention]

However, when the number of categories is large, it is difficult to construct a huge number of decision logics in the discrimination tree using a method of configuring a discrimination tree by combining decision logics for discriminating between categories.

In addition, in the pattern matching method using a representative pattern registration dictionary, in order to reduce the number of misrecognized patterns, it is necessary not only to correct the representative pattern of each category but also to register additional subcategories. There is no viable method.

Furthermore, since the number of subcategories is unknown in the hierarchical clustering method and the LBG method, it is difficult to grasp the scope of application. Furthermore, when used for pattern recognition, it may be necessary to consider similar categories, but clustering within individual categories does not guarantee improvement in recognition rate.

Furthermore, in the above-mentioned method of selecting the core pattern of a subcategory from a set of misrecognition patterns for each category and registering it in the recognition dictionary, it is not possible to remedy the misrecognition pattern of the own category by devising the selection of subcategories. However, there is no efficient method to avoid the risk that correct recognition patterns for similar categories turn into incorrect recognition.

As mentioned above, methods for registering additional subcategories when performing recognition based on the distance or similarity between patterns from the representative pattern of each category include a method of clustering within each category to form subcategories, and a method of forming subcategories by clustering within each category. A method has been considered that selects patterns for each subcategory from a set. However, each method has its own drawbacks, and adding additional subcategories does not necessarily guarantee improved recognition rates.

The present invention has been made in view of the above points, and an object of the present invention is to provide a general-purpose pattern recognition dictionary creation device that sequentially registers additional subcategories so that the number of erroneously recognized patterns is definitely reduced.

[Means to solve the problem]

A pattern recognition dictionary creation device that creates a recognition dictionary including subcategories from a learning pattern set including a plurality of categories, comprising: a learning pattern storage device that stores the learning pattern set; a recognition dictionary storage device that stores representative patterns of each category; a first calculation device that calculates an inter-pattern distance between each learning pattern stored in the learning pattern storage device and each representative pattern stored in the recognition dictionary storage device; and an inter-pattern distance calculated by the first calculation device. a recognition result storage device that stores the minimum category name, its minimum distance between patterns, and the correct category name; a recognition pattern classification device, a subcategory candidate selection device that extracts the classified misrecognition pattern set one by one in an arbitrary order and selects a representative pattern therein as a subcategory candidate of a certain category name, and a representative pattern of the subcategory candidate. and each learning pattern, and a second calculation device that calculates the inter-pattern distance between each learning pattern and the inter-pattern distance value of the minimum pattern stored in the recognition result storage device for each learning pattern. A recognition result correction device that corrects the recognition category name of the learning pattern by comparing the inter-pattern distance values calculated, and a recognition dictionary that calculates the reduction in the number of erroneously recognized patterns based on the correction results by the recognition result correction device. and a category registration determination device that determines whether or not to additionally register the subcategory candidate, and the subcategory candidate selection device determines whether all learned patterns are correctly recognized and no subcategory candidates appear, or whether all subcategory candidates are registered. The recognition dictionary is updated by sequentially adding subcategories until it is determined that it is not possible.

[Effect]

In the present invention, the misrecognition pattern classification device classifies misrecognition pattern sets into separate sets based on combinations of correct category names and misrecognition category names. The subcategory candidate selection device tentatively registers one subcategory candidate as a subcategory candidate in the recognition dictionary in an appropriate order from among the misrecognition pattern set, which is a set of misrecognition category names classified in this way. If the number of misrecognized patterns decreases as a whole, the subcategory candidate is formally added to the recognition dictionary and the misrecognition pattern set is updated. Furthermore, if the number of misrecognized patterns increases as a whole, it is similarly determined whether the number of misrecognized patterns decreases for the next subcategory candidate.

If the number of misrecognized patterns becomes zero due to additional registration of a subcategory, or if it is determined that all subcategory candidates cannot be registered and the subcategory is not additionally registered, the recognition dictionary obtained at that point. is output as the final pattern recognition dictionary, and the operation ends. In this way, the present invention does not guarantee an improvement in the recognition rate.

You can easily register subcategories.

〔Example〕

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention. In the figure, ■ is a learning pattern storage device that stores learning patterns. 2 is a recognition dictionary storage device that stores representative patterns of each category and additionally registered subcategories. 3 is an inter-pattern distance calculation device a that calculates an inter-pattern distance as a recognition measure. Reference numeral 4 denotes a recognition result storage device that stores the recognition category name and the minimum inter-pattern distance for each learning pattern. Reference numeral 5 denotes a misrecognition pattern classification device that classifies misrecognition pattern sets into separate sets by combining correct category names and misrecognition category names. Reference numeral 6 denotes a subcategory candidate selection device that selects subcategory candidates one by one in an appropriate order from the set of misrecognition patterns classified by the misrecognition pattern classification device 5. 7 is an inter-pattern distance calculation device that calculates an inter-pattern distance between each learning pattern and the representative pattern of the subcategory candidate. 8
compares the minimum inter-pattern distance value stored in the recognition result storage device 4 with the inter-pattern distance calculated by the inter-pattern distance calculation device 7 for each learning pattern, and if the inter-pattern distance calculation device b7 is smaller, the learning pattern is This is a recognition result modification device that modifies recognition category names. 9 counts the decrease Δ in the number of incorrectly recognized patterns as a whole due to correction of recognition results, and if Δ is positive, the subcategory candidate is additionally registered in the recognition dictionary, and if Δ is negative, the next subcategory candidate is This is a subcategory registration determination device that determines whether to proceed to processing.

The learning pattern storage device 1 stores a certain number of pattern sets for each category, with the total number of categories being recognized. It is assumed that each pattern to be stored is stored as an expression suitable for recognition processing, for example, a vector expression in which values of a plurality of recognition features effective for recognition are arranged (learning pattern set).

The recognition dictionary storage device 2 creates a certain representative pattern Mn (l≦n≦K) from a fixed number of learning pattern sets for each category stored in the learning pattern storage device 1 in advance. This representative pattern Mn is stored in the recognition dictionary storage device 2 as an initial recognition dictionary. For example, an average pattern may be used as the representative pattern Mn, and if each pattern is expressed by a vector, an average vector is used.

The inter-pattern distance calculation device a3 calculates the inter-pattern distance between the representative pattern of each category stored in the recognition dictionary storage device 2 and each representative pattern of the subcategory group additionally registered for the category. The calculated minimum value of the inter-pattern distance is sent to the recognition result storage device 4 as the inter-pattern distance value for the category. Also, since the total number of categories is K, eight inter-pattern distance values are applied to each learning pattern.

Further, as the distance between patterns, for example, a general-purpose Euclidean distance in a vector space may be used.

In addition, when using similarity instead of inter-pattern distance as a recognition measure, it can be expressed by replacing "inter-pattern distance" with "similarity" and "minimum inter-pattern distance" with "maximum similarity" in this example. I can do it.

The recognition result storage device 4 stores the minimum value of the inter-pattern distance values for each learning pattern sent from the inter-pattern distance calculation device a3, the recognition category name, and the correct category name.

The misrecognition pattern classification device 5 extracts the recognition category name of each learning pattern stored in the recognition result storage device 4 or those different from the correct category name as misrecognition patterns, and uses the combination of the correct category name and the misrecognition category name. Classify into multiple misrecognition pattern sets. The operation of this misrecognition pattern classification device 5 will be explained with reference to the conceptual diagram of FIG. The extracted total erroneous recognition patterns 20 are expressed by a combination of the correct category name 1 and the erroneous recognition category name j. This reduces the set of multiple patterns to 5ij (1≦L j≦k, j≠1)
21a, 21b, 21c, 21d, and each S
The erroneous recognition patterns belonging to ij are stored in the learning pattern storage device l.
Read from and store. Here, the number of classifications is C(k, 2)=
It is about k(k-1)/2.

The subcategory candidate selection device 6 arranges the plurality of misrecognition pattern sets classified by the misrecognition slam classification device 5 in an appropriate order and selects them one by one as a subcategory candidate in that order. Further, the representative pattern of the subcategory candidate is sent to the inter-pattern distance calculation device b7. The sent subcategory candidates are sent to inter-pattern distance calculation devices b7 to be described later.
The subcategory registration determination device 9 determines whether additional registration to the recognition dictionary is possible. At this time, the subcategory candidate selection device 6 continues sending subcategory candidates until a registrable subcategory appears. If it is determined that all subcategory candidates cannot be registered and the subcategory candidates are exhausted,
Alternatively, when the number of erroneously recognized slams becomes zero and the number of subcategory candidates also becomes zero, the recognition dictionary obtained at that time is used as the final pattern recognition dictionary, and the operation ends.

The pattern/distance calculation device b7 creates a representative pattern M' for the set of misrecognized patterns Sij of subcategory candidates sent from the subcategory candidate selection device 6.

The inter-pattern distance calculated with M' for each learning pattern read from the learning pattern storage device 1 is sent to the recognition result correction device 8. However, incorrect recognition pattern S
Since iJ is a subcategory candidate of category name i, the inter-pattern distance calculated above corrects the inter-pattern distance for category name i.

The recognition result correction device 8 compares the minimum inter-pattern distance value stored in the recognition result storage device 4 for each learning pattern with the inter-pattern distance value sent from the inter-pattern distance calculation device b7, and calculates the latter inter-pattern distance. If the value is small, the new recognition category name of the learning pattern is stored as i along with the old recognition category name and the correct category name. The corrected contents of the stored recognition results are sent to the subcategory registration determination device 9.

FIG. 3 shows a conceptual diagram for explaining the operation of the subcategory registration determination device 9. The subcategory registration determination device 9 compares the old and new recognition category names of each learning pattern with the correct category name from the recognition result modification contents sent from the recognition result modification device 8. In the figure, let P be the pattern set 35 of the pattern set of category l that has been moved from the correct recognition pattern 31 to the pattern set 35 recognized as subcategory candidate M'.
The movement from the correctly recognized pattern 32 of the category 2 pattern set to the recognized pattern set 35 to subcategory candidate M' is designated as P2, and the movement from the incorrectly recognized pattern 33 of the category 1 pattern set to M' is the recognized pattern set 3.
Let Q be the pattern that has moved to M', and let PK be the pattern that has moved from the misrecognized pattern 34 of the pattern set in the category to the pattern set 35 that has been recognized to M'.

As a result, the number of correct recognition patterns for category name j (≠i) that become the recognized category name i, and the number of cases that changed to incorrect recognition patterns, is Pj, and the number of cases that changed from incorrect recognition patterns for category name 1 to correct recognition patterns. is counted as Qi, and the decrease Δ in the number of misrecognized patterns is calculated using the following formula.

Δ=Qi−ΣPj formula (lj If the decrease in the number of misrecognized patterns Δ is positive, it is determined that the subcategory can be registered. In other words, if Δ is positive, M′ is recognized as a subcategory of category name i. 2, and repeat the processing in inter-pattern distance calculating device a3 and below. Also, if the value of Δ in equation (1) is negative, the operation is returned to subcategory candidate selection device 6, and the next subcategory candidate is selected. are selected one after another, and the processes from the inter-pattern distance calculation device b7 onward are repeated.

〔Effect of the invention〕

As described above, according to the present invention, when pattern recognition is widely performed by a computer based on the distance between patterns or the degree of similarity from the representative pattern of each category, misrecognized patterns are classified by a combination of correct category name and incorrect recognition category name. By creating subcategory candidates based on the above criteria and providing a means to determine whether or not to add them to the recognition dictionary on the condition that the number of misrecognized patterns is reduced, it is possible to create a pattern recognition dictionary that guarantees an improvement in the recognition rate. . Furthermore, most of the processing consists of calculating the distance or similarity between patterns, and the content is simple and the amount of calculation is small.

As described above, the present invention provides a general-purpose and effective technique for creating a dictionary for pattern recognition because additional registration of subcategories can be easily performed without guaranteeing an improvement in the recognition rate. For this reason, it is particularly advantageous when applied to recognition targets where the number of categories is large and the number of patterns belonging to each category is insufficient, such that statistical clustering theory would be invalidated.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, FIG. 2 is a conceptual diagram for explaining the operation of an erroneous recognition pattern classification device according to an embodiment of the present invention, and FIG. 3 is a block diagram showing the configuration of an embodiment of the present invention. FIG. 2 is a conceptual diagram for explaining the operation of the subcategory registration determination device of one embodiment. l... Learning pattern storage device, 2... Recognition dictionary storage device, 3... Inter-pattern distance calculation device a, 4... Recognition result storage device, 5... Erroneous recognition pattern classification device, 6
...Subcategory candidate selection device, 7. Inter-pattern distance calculation device, 8. Recognition result correction device, 9. Subcategory registration determination device. Patent applicant: Nippon Telegraph and Telephone Corporation Block diagram N1 showing the configuration of an embodiment of the present invention Figure 2vA Conceptual diagram for explaining the operation of the misrecognition slam classification device 5

Claims (1)

[Scope of Claim] A pattern recognition dictionary creation device that creates a recognition dictionary including subcategories from a learning pattern set including a plurality of categories, comprising: a learning pattern storage device that stores the learning pattern set; and a learning pattern storage device that stores representative patterns of each category. a recognition dictionary storage device; a first calculation device that calculates an inter-pattern distance between each learning pattern stored in the learning pattern storage device and each representative pattern stored in the recognition dictionary storage device; a recognition result storage device that stores the category name for which the inter-pattern distance calculated by the calculation device is the minimum, the minimum inter-pattern distance, and the correct category name; A misrecognition pattern classification device that classifies into a plurality of misrecognition pattern sets, and a subcategory that extracts the classified misrecognition pattern sets one by one in an arbitrary order and selects a representative pattern among them as a subcategory candidate of a certain category name. a candidate selection device; a second calculation device that calculates an inter-pattern distance between the representative pattern of the subcategory candidate and each learning pattern; and a minimum pattern stored in the recognition result storage device for each learning pattern. a recognition result correction device that corrects the recognition category name of the learning pattern by comparing the inter-pattern distance value and the pattern-to-pattern distance value calculated by the second calculation device; and a category registration determination device that calculates a decrease in the number of patterns and determines whether to be additionally registered in the recognition dictionary based on the calculation result, and the subcategory candidate selection device has a category registration determination device that calculates a decrease in the number of patterns and determines whether or not to additionally register the patterns in the recognition dictionary. A pattern recognition dictionary creation device that updates a recognition dictionary by sequentially registering additional subcategories until no subcategories appear or it is determined that all subcategory candidates cannot be registered.