JP3180477B2 - Pattern recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for recognizing an input pattern. In particular, the present invention relates to an apparatus for recognizing a pattern by performing hierarchical recognition such as performing a large classification and then a fine classification on an input pattern.

[0002]

2. Description of the Related Art Conventionally, for an input pattern,
First, a large classification is performed to select a category group to which the input pattern belongs (here, the category group indicates a set of patterns having similar pattern feature vectors), and then the sub-classification is performed on the selected category group. As an example of a pattern recognition apparatus that recognizes an input pattern by performing the following, for example, IEICE Transactions D-II Vol.J
75-D-II No.3 pp545-553 "Large-scale neural network" C
ombNET-II "".

FIG. 9 is a block diagram of this conventional pattern recognition apparatus. Reference numeral 110 denotes a feature extraction unit which extracts a feature vector used to identify a category from an input pattern. Reference numeral 111 denotes a large classification unit, which largely classifies the input pattern into each category group. 112
Denotes a sub-classification unit, which sub-classifies the input pattern in each category group. 113 is a group selection unit, and a large classification unit 1
A plurality of category groups are selected from the eleven output values (hereinafter, referred to as “fitness”). 114 is a sub-classification unit input signal selection unit, which inputs a feature vector of an input pattern based on the group selection information obtained by the group selection unit 113
Select 112. Reference numeral 115 denotes an identification unit that identifies an input pattern based on the degree of conformity of the category group selected by the group selection unit 113 and the output value of the subclassification unit 112.

In the large classification unit 111, an input unit 116 inputs a feature vector of an input pattern. Reference numeral 117 denotes a multi-input / one-output signal processing unit which calculates the degree of conformity of each category group to an input pattern.

In the sub-classifying section 112, reference numeral 118 denotes an input section for inputting a feature vector of an input pattern outputted from the sub-classifying section input signal selecting section 114. 119 is a multi-input / one-output signal processing unit, and the output of the lower-layer input unit 118 or the multi-input / one-output signal processing unit 119 connected thereto is multiplied by a weight coefficient which is a degree of the connection. The sum is output after threshold processing. Here, the plurality of multi-input / one-output signal processing units have a layered structure, and are connected to a network so that there is no mutual connection in each layer and the signal propagates only to the upper layer. Is determined for each category. Reference numeral 120 denotes a maximum value selection unit which selects the maximum value from the outputs of the plurality of multi-input / output signal processing units in the uppermost layer.

In the identification unit 115, a similarity calculation unit 121 is used to determine the similarity of each category from the fitness of the category group selected by the group selection unit 113 and the output value of the subclassification unit 112 corresponding to the category group. Calculate the degree. A category identification unit 122 identifies the category of the input pattern by obtaining the maximum value of the similarity of each category obtained from the similarity calculation unit 121.

The operation of the conventional pattern recognition device configured as described above will be described below. First, the feature extraction unit 110 generates a feature vector X composed of n pieces of feature data for an input pattern.

[0008]

(Equation 1)

Is obtained. Next, the feature vector X is input to the input unit 116 of the large classification unit 111. The input unit 116 is equal to n pieces prepared wherein the number of data patterns, each feature data x _i is inputted to the input unit 116 corresponding.
Each multi-input first output signal processing section 117 of the major classification section 111, an input x _j and the weighting coefficient _{v ij (1 ≦ i ≦ m} r is the degree of connection of the input section 111 connected thereto; the m _r After calculating the sum of the product of the number of category groups and 1 ≦ j ≦ n), the sum is calculated as the feature vector X and the weight coefficient vector V _{i of} each multi-input / one-output signal processing unit 117.

[0010]

(Equation 2)

Norm | X |, | V_iDivide by the product of |
Output That is, the weight coefficient vector shown in FIG.
Tor V_iOutput value sim of the multi-input one-output signal processing unit 117 having
(X, V _i)

[0012]

(Equation 3)

Can be expressed as The weight coefficient vector V _i is designed in advance so that a determined multi-input / one-output signal processing unit generates a maximum output for a set of patterns having similar feature vectors X.

According to the prior art, these weight coefficient vectors V _i are designed by the following method. First, in the first process, every time a feature vector X of a pattern for weight coefficient vector design is input, V _C having the largest sim (X, V _i ) is set.
The calculated (this time, X is that optimally matched to V _C.), Closer to V _C to X. When the number of input patterns optimally matched to one weight coefficient vector exceeds a certain number, the area assigned to the weight coefficient vector is divided into two. In the second step, the total pattern for weighting coefficient vectors designed to obtain the V _i for optimum matching to see if it has changed from the previous. And, if there is a change, to correct the V _i. At this time, similarly to the first process, the division of the weight coefficient vector is also performed. This is repeated until the correction and division of the weight coefficient vector are eliminated.

[0015] In this way, by performing the design of the weighting coefficient vectors, each weight coefficient vector V _i divides the feature vector space of the input pattern, it can be quantized. That is, the input pattern by each weighting coefficient vector V _i, a set of a plurality of patterns feature vectors are similar, that is, into categories groups. The output value of each multi-input / one-output signal processing unit 117 is used as the degree of conformity of each category group with respect to the input pattern.
Output to 3.

The group selecting unit 113 selects an arbitrary number of category groups in descending order of the degree of matching obtained by the large classifying unit 111, and outputs group selecting information indicating which category group has been selected and the corresponding degree of matching. I do.

Based on the group selection information obtained from the group selection unit 113, the sub-classification input signal selection unit 114 selects the sub-classification unit 112 for inputting the feature vector X of the input pattern, and replaces X with these sub-classification units. Output to 112.

In each of the sub-classification units 112 corresponding to the category group selected by the group selection unit 113 (ie, the sub-classification unit 112 to which the feature vector X of the input pattern is input from the sub-classification input signal selection unit 114) First, the input unit 118 receives the feature vector X. The input unit 118 is equal to the N prepared wherein the number of data patterns, each feature data x _i is inputted to the input unit 118 corresponding. Each multi-input / one-output signal processing unit 119 of the sub-classification unit 112 multiplies the output of the lower-layer input unit 118 connected thereto or the output of the multi-input / one-output signal processing unit 119 by a weight coefficient which is a degree of connection thereof. After converting the sum of the data by a threshold function, the value is output to the upper layer. Here, the multi-input one-output signal processing unit 119 of the top layer of each sub-classification unit 112 is set to the same number as the number of categories of the pattern included in each category group, and the multi-input one-output signal of the top layer is set. The processing unit 119 corresponds to each of these categories. The maximum value selection unit 120 selects the largest one of the output values of each multi-input one-output signal processing unit 119 of the highest layer, and the category corresponding to the multi-input one-output signal processing unit 119 and the maximum output value Is output.

The weight coefficient of each multi-input / one-output signal processing unit 119 is determined based on the feature vector X of the pattern having each category in the category group, and the multi-input / one-output of the top layer corresponding to each category. The signal processing unit 119 is previously learned so as to generate the maximum output.

Specifically, such a learning method of the weight coefficient is performed by a learning algorithm called an error back propagation method. For the error backpropagation method, for example, DE Ru
"Lea" by melhart, GE Hinton and RJ Williams
rning Representations by Back-Propagating Errors, "
Nature, vol. 323, pp. 533-536, Oct. 9, 1986.

Hereinafter, an outline of the error back propagation method will be described. First, the feature vector X of the pattern for weight coefficient learning
Is input to the input unit 118 of the sub-classification unit 112. As described above, each multi-input one-output signal processing unit 119 includes a lower-layer input unit 118 connected thereto, or the output of the multi-input one-output signal processing unit 119 and a weighting factor indicating the degree of connection thereof. Is converted by a threshold function, and the value is output to the upper layer. It is where the error E with all of the output signal o _k and the desired output signal t _k multiple-input first output signal processing section 119 of the top layer (referred to as a teacher signal) is obtained as equation (4) .

[0022]

(Equation 4)

Here, Σ _p is the total sum regarding the number of patterns of the teacher signal. The purpose of the learning is to determine the value of the weight coefficient that minimizes the error E, and the amount of change △ w _ij of the weight coefficient of each multi-input / one-output signal processing unit 119 is calculated based on ( _Equation 5). Is done.

[0024]

(Equation 5)

Here, ε is a positive constant called a learning rate. The error E can be reduced by repeating such updating of the weighting coefficient based on (Equation 5) every time the feature vector of the learning pattern is input. When the error E becomes sufficiently small, the learning is terminated on the assumption that the output signal has become sufficiently close to the desired value.

According to such a weight coefficient learning method,
For a pattern having each category in the category group, the multi-input / one-output signal processing unit 119 of the highest layer corresponding to each category can generate the maximum output.
Therefore, a plurality of multi-input one-output signal processing units 11 in the uppermost layer
By selecting the one that generates the maximum output from among the nine in the maximum value selection unit 120, the category of the input pattern can be identified in each category group, that is, in each sub-classification unit.

In the identification section 115, first, in the similarity calculation section 121, the degree of conformity of the category group selected by the group selection section 113 and the output value of the subclassification section 112 corresponding to the category group are expressed by (Equation 6). The similarity of each category obtained by the sub-classification unit 112 is calculated using the equation, and these similarities are calculated by the category identification unit.
Output to 122.

[0028]

(Equation 6)

Here, a and b are real constants. Finally, the category identifying unit 122 compares the similarities of the respective categories obtained from the similarity calculating unit 121, and outputs a category corresponding to the maximum similarity among them as an identification result.

[0030]

In pattern recognition, features that are effective in identifying a pattern are extracted as feature vectors. However, in general, it is difficult to realize a sufficient recognition ability with only a single feature vector. By using a plurality of types of feature vectors, a high identification function can be realized. In other words, by using a plurality of types of features, even a pattern in which a certain feature vector can only be ambiguously or erroneously identified may be able to be accurately identified with a different feature vector. This is because there is a possibility that the discrimination performance can be improved by placing importance on.

However, in a pattern recognition device that hierarchically recognizes a pattern having many categories,
In a configuration in which only one feature extraction unit and recognition unit are provided, as in the conventional example, if a plurality of types of feature vectors are used, they must be collectively input and recognized. Absent. In this case, it is possible to certainly realize a somewhat higher discrimination performance than when only a single feature vector is used, but it is not possible to make full use of the advantages obtained by using different feature vectors as described above. There is a problem. In other words, when a plurality of types of feature vectors are collectively used, the identification performance can be improved, but using the recognition result,
With a certain feature vector as described above, it is very difficult to correctly recognize a pattern that is erroneously recognized based on the identification result of a different feature vector.

Further, when a plurality of types of feature vectors are collectively used, there is a problem that a large amount of calculation time is required for recognition as the number of dimensions of the feature vectors increases.

In view of the problem of the conventional pattern recognition apparatus, the present invention provides a highly accurate recognition method that makes full use of the advantage of using a plurality of types of feature vectors together when a plurality of types of feature vectors are used in the recognition apparatus. It is another object of the present invention to provide a pattern recognition device that can realize the above-described method and can shorten the time required for recognition.

[0034]

According to the present invention, there are provided a plurality of feature extraction units for obtaining different feature vectors from an input pattern, and a similarity which is a degree to which the input pattern belongs to each category is obtained from each of the feature vectors. A plurality of single feature recognition units to be obtained, and a category identification unit that identifies the input pattern using similarities obtained from the plurality of single feature recognition units, wherein the single feature recognition unit is A group dictionary in which a plurality of group reference feature vectors representing a group of patterns having similar feature vectors is stored, and an input pattern is formed using the group reference feature vector and the feature vector of the input pattern. A fuzzy large classification unit that calculates a group belonging degree, which is a degree belonging to each category group, and the input pattern is obtained by using a feature vector of the input pattern. A plurality of sub-classification units for obtaining intra-group similarity, which is a degree belonging to each category included in the category group, a group selection unit for selecting a plurality of category groups from the group belonging degree, and the group selection unit A sub-classification unit input signal selection unit that selects a sub-classification unit that inputs the feature vector of the input pattern based on group selection information; and a group membership degree of the category group selected by the group selection unit and the sub-classification unit. A single feature similarity calculation unit that obtains a similarity of each category with respect to the input pattern using the obtained intra-group similarity; wherein the single feature similarity calculation unit includes a category group selected by the group selection unit. A plurality of multipliers for multiplying all the intra-group similarities obtained from the sub-classification unit input with the feature vector of the input pattern from the sub-classification unit input signal selecting unit and the sub-classification unit; The multiplier A category similarity calculation unit for selecting a plurality of output values having a large value and calculating the sum of these output values is provided, and the category identification unit is obtained from the single feature recognition unit corresponding to each of the feature vectors. A plurality of integrated similarity converters for converting the similarity of each category into an integrated similarity by normalizing the maximum similarity among the similarities, and integrated similarities obtained from the plurality of integrated similarity converters A category determining unit for determining the final similarity of each category using the degrees and identifying the final input pattern.

[0035]

According to the present invention, first, each feature extracting unit extracts a different feature vector from an input pattern. Next, these feature vectors are respectively input to a plurality of corresponding single feature recognition units.

Each single feature recognition unit first inputs a feature vector to the fuzzy large classification unit. The fuzzy large classification unit reads out all the group reference feature vectors stored in the group dictionary, and performs a large classification on the feature vector in which the boundary of each category group in the feature vector space is ambiguous. , The degree of group membership, which is the degree to which the input pattern belongs to each category group, is determined. Using these group membership degrees, the group selection unit selects a plurality of category groups, outputs group selection information to the subclassification unit input signal selection unit, and outputs the corresponding group membership degrees to the single feature similarity calculation unit. I do. The sub-classification unit input signal selection unit outputs the feature vector to the sub-classification unit corresponding to the group selection information, and each sub-classification unit uses the feature vector to include the input pattern in each category group. In-group similarity, which is the degree of similarity to each category, is calculated and output to the single feature similarity calculation unit.

The single feature similarity calculation unit is obtained from the group classification degree of the category group selected by the group selection unit and the subclassification unit that receives the feature vector from the subclassification unit input signal selection unit. All intra-group similarities are multiplied by a multiplier and output to the category similarity calculator. In the category similarity calculation unit, a plurality of large output values of the multiplier are selected for each category, and the sum of these output values is calculated.
This is defined as the similarity of each category to the input pattern. That is, each single feature recognition unit outputs this similarity to the category identification unit as a similarity obtained from each feature vector for the input pattern.

Lastly, in the category identification section, each of the integrated similarity conversion sections firstly calculates the similarity of each category obtained from the single feature recognition section corresponding to each of the feature vectors. Convert to the integrated similarity by normalizing with the maximum value. Next, the category determination unit obtains the final similarity of each category with respect to the input pattern using these integrated similarities, and identifies the final input pattern.

As described above, the present pattern recognition apparatus performs a classification process on each of a plurality of types of feature vectors in separate single feature recognition units to determine the similarity of each category to each feature vector. Integrates the identification results obtained from each single feature recognition unit using all these similarities, and performs final input pattern recognition. Therefore, a pattern that is erroneously recognized in a certain feature vector can be correctly recognized based on an identification result by a different feature vector, and highly accurate recognition can be realized. Further, in the conventional method, the time required for recognition increases due to the use of a plurality of types of feature vectors. However, in the present pattern recognition apparatus, since each single feature recognition unit performs the identification processing in parallel, the time required for recognition is Compared to the conventional example,
Can be shorter.

[0040]

Embodiments of the present invention will be described below with reference to the drawings.

In general, pattern data input to a pattern recognition device includes time-series patterns such as voices,
Although there are spatial patterns such as images, etc., any pattern data may be used in the present invention.

FIG. 1A is a block diagram of a pattern recognition device according to an embodiment of the present invention. FIG.
In FIG. 7A, reference numeral 11 denotes a feature extraction unit for extracting a feature vector used to identify a category from an input pattern. However, each of the plurality of feature extraction units extracts a different feature vector. Reference numeral 12 denotes a single feature recognizing unit, which is prepared in the same number as the feature extracting unit 11 and obtains a similarity in which an input pattern belongs to each category from a corresponding feature vector. Reference numeral 13 denotes a category identification unit that identifies an input pattern by using the similarity of each category to each feature vector of the input pattern obtained from each single feature recognition unit 12.

In the category discriminating unit 13, reference numeral 14 denotes an integrated similarity converting unit. The same number is prepared as the feature extracting unit 11 and the single feature recognizing unit 12, and each category is obtained from the corresponding single feature recognizing unit 12. Are normalized into the integrated similarity by normalizing the similarity with the maximum value among the similarities. Reference numeral 15 denotes a category determination unit that obtains the final similarity of each category with respect to the input pattern using all the integrated similarities, and identifies the final category.

The category determining unit 15 shown in FIG.
Is a first example of the category determination unit 15. FIG.
In (a), reference numeral 16 denotes an adder, which obtains the final similarity of each category by adding the integrated similarity of each category obtained from the plurality of integrated similarity calculators 14 for each same category. . Reference numeral 17 denotes a maximum value selection unit for selecting a maximum value from the final similarities of the respective categories.

FIG. 1B is a block diagram specifically showing the configuration of the single feature recognition unit 12. In FIG. 1B,
Reference numeral 21 denotes a group dictionary in which a plurality of group reference feature vectors representing a category group including a set of patterns having similar feature vectors are stored. Reference numeral 22 denotes a fuzzy large classification unit that calculates a group belonging degree, which is a degree at which an input pattern belongs to each category group, using a group reference feature vector stored in the group dictionary 21. Reference numeral 23 denotes a sub-classification unit which calculates the intra-group similarity, which is the degree to which the input pattern belongs to each category included in the category group. Reference numeral 24 denotes a group selection unit for selecting a plurality of category groups based on the degree of group membership. Reference numeral 25 denotes a sub-classification unit input signal selection unit, which is a sub-classification unit 23 that inputs a feature vector of an input pattern based on group selection information obtained from the group selection unit 24.
Is to select. Reference numeral 26 denotes a single feature similarity calculation unit, which calculates the similarity of each category to the input pattern from the group membership of the category group selected by the group selection unit 24 and the intra-group similarity obtained by the subclassification unit 23 Things.

In the single feature similarity calculating section 26, reference numeral 27 denotes a multiplier, which calculates the group membership degree of the category group selected by the group selecting section 24 and the feature vector of the input pattern from the sub-classifying section input signal selecting section 25. This is to multiply the input intra-group similarity obtained from the sub-classification unit 23. Reference numeral 28 denotes a category similarity calculator for selecting a plurality of output values of the multiplier 27 having a large output value for each category and calculating the sum of these output values.

FIG. 2 is a block diagram specifically showing the configuration of the fuzzy large classification unit 22. In FIG. 2, an input unit 31 inputs a feature vector of an input pattern. Reference numeral 32 denotes a distance calculation unit that calculates the distance between all the group reference feature vectors of the group dictionary 21 and the feature vectors. 33 is a divider for calculating the reciprocal of the output of the distance calculator 32. 34 is an adder for calculating the sum of the outputs of the respective dividers 33. A multiplier 35 multiplies the output of the adder 34 by the output of the distance calculator 32. A divider 36 calculates the reciprocal of the output of the multiplier 35.

FIG. 3 is a block diagram specifically showing the configuration of the first embodiment of the subclassification unit 23. In FIG. 3, reference numeral 41 denotes an input unit for inputting a feature vector of an input pattern output from the sub-classification unit input signal selection unit 25.
Reference numeral 42 denotes a category dictionary which stores a plurality of category reference feature vectors indicating representative values of each category of the pattern. Reference numeral 43 denotes a distance calculation unit that calculates the distance between all the category reference feature vectors of the category dictionary 42 and the feature vectors of the input pattern. 44 is a divider for calculating the reciprocal of the output of the distance calculator 43. Reference numeral 45 denotes an adder for calculating the sum of outputs from the respective dividers 44. 46 is a multiplier and an adder 45
Is multiplied by the output of the distance calculation unit 43. 47
Is a divider for calculating the reciprocal of the output of the multiplier 36.

The operation of the pattern recognition device configured as described above will be described below.

First, a plurality of prepared feature extraction units 11 extract different feature vectors X _k (k = 1 to N _{F, where} N _F is the number of types of feature vectors) from the input pattern. I do. Here, each feature vector X _k is composed of n _k pieces of feature data, and is represented as (Equation 7).

[0051]

(Equation 7)

As examples of a plurality of different feature vectors, for example, if a pattern is a character, the following can be considered. For example, if the image information of a character is 8
The image is divided into 64 parts, and the image information of a character or a 64-dimensional shading feature (mesh feature) using the frequency of a black pixel as an element is divided into four.
There is a 64-dimensional contour direction density feature in which the frequency of appearance of the direction elements (horizontal, vertical, right diagonal, left diagonal four directions) of the character outline points in each of the four divided areas is an element.

Each feature vector X _k of the input pattern extracted by each feature extraction unit 11 is stored in the corresponding single feature recognition unit 1.
Entered in 2. Multiple single feature recognition units 12
Performs the following identification processing independently,
The similarity of each category to each feature vector X _k is determined. Note that a feature vector X _k used in the following description of the operation of the single feature recognition unit 12 is a feature vector obtained by a certain feature extraction method, and hereinafter, X _k and _nk are described as X and n for simplicity. I do.

Hereinafter, a specific operation of the single feature recognition unit 12 will be described. First, in the single feature recognition unit 12, the feature vector X is input to the fuzzy large classification unit 22. In the fuzzy large classification unit 22, the input unit 31 inputs the feature vector X and outputs X to the m _r distance calculation units 32. Each distance calculator 32
Is a group reference feature vector V _i (1 ≦ i ≦ m _r ; _mr is the number of group reference feature vectors, that is, the number of category groups) representing each category group stored in the group dictionary 21. readout, (8) calculates the distance d _i between X and V _i shown in the expression, and outputs each corresponding divider 33, and to the multiplier 35.

[0055]

(Equation 8)

Here, f is a real number satisfying f> 1. Each divider 33, the reciprocal of the distance d _i is calculated, and its output to the adder 34. The adder 34 calculates the sum of the outputs of all the dividers 33 and outputs the output to the m _r multipliers 35. Each multiplier 35 multiplies the output of the corresponding distance calculator 32 and the output of the adder 34, and inputs the output to the corresponding divider 36. Each divider 36 calculates the reciprocal of the output of the corresponding multiplier 35. Finally, in the fuzzy large classification unit 22, the output of each divider 36 is the group membership μ _i (1 ≦ i ≦ 1) of each category group with respect to the feature vector X of the input pattern.
m _r ) is output to the group selector 24. That is, the group membership degree μ _i (1 ≦ i ≦ m _r ) of each category group is

[0057]

(Equation 9)

Can be expressed as follows. Note that a group reference feature vector representing each category group stored in the group dictionary 21 is determined in advance by a conventional clustering method, for example,
IEICE, edited by Makoto Nagao, "Pattern Information Processing"
K-means algorithm shown in (Corona), and
Isodata algorithm, Y.Linde, A.Buzo, and RMGr
"An Algorithmfor Vector Quantizer desig by ay
n, "IEEE Trans. Commun., COM-28, 1, pp.84-95, Jan.1
980 is designed using the LBG algorithm.

Hereinafter, a group dictionary using the K-means algorithm will be described.
A brief description of how to design 21. (1) From a set of patterns for designing a group dictionary of a recognition target, _mr (where _mr is a predetermined number of category groups)
Are appropriately selected, and these feature vectors are referred to as m _r
It is assumed that the group reference feature vectors V _i (1 ≦ i ≦ m _r ). (2) For the feature vectors X of all the group dictionary design patterns, distances d _i shown in (Equation 10) are respectively given.

[0060]

(Equation 10)

Is determined to minimize V _i . At this time, X
Belong to the category group S _i (1 ≦ i ≦ m _r ). (3) The average value of the feature vector X of the pattern belonging to each S _i is obtained, and this is set as V _i ′. (4) if holds for V _i '= V _i all i, stores the feature vector V _i for reference group of this time the group dictionary 21. Otherwise, V _i ′ is set as a new group reference pattern signal V _i and the process returns to (2).

By designing a group reference feature vector in this way, all patterns can be divided into some subsets (category groups) of patterns having similar feature vectors. Note that the Isodata algorithm and the LBG algorithm are basically almost the same as the K-means algorithm.

The group selecting unit 24 selects a plurality of category groups in descending order of the degree of group membership obtained by the fuzzy large classifying unit 12, and outputs group selection information indicating which category group has been selected to the sub-classified input signal selection. The unit 25 outputs the group membership degree corresponding thereto to the single feature similarity calculation unit 26. As a method of selecting a category group, a category group having a group membership degree equal to or higher than a certain threshold value may be selected.

Based on the group selection information obtained from the group selection unit 24, the sub-classification input signal selection unit 25 selects the sub-classification unit 23 for inputting the feature vector X of the input pattern, and converts the feature vector X into these sub-categories. Output to the classification unit 23.

Each of the sub-classification units 23 corresponding to the category group selected by the group selection unit 24 (that is, the sub-classification input signal selection unit 25
Sub-classification unit that receives the feature vector of the input pattern from
In 23), first, the input unit 41 inputs the feature vector X of the input pattern, and outputs the X to m _c-number of distance calculator 43.
Each distance calculation unit 43 calculates a category reference feature vector W _i (1 ≦ i ≦ _mc ; _mc is the number of category reference feature vectors) indicating a representative value of each category stored in the category dictionary 42. Then, the distance d _i between X and W _i shown in Expression (11) is calculated, and the corresponding divider 44 and multiplier 46 are respectively assigned.
Output to

[0066]

[Equation 11]

Here, f is a real number satisfying f> 1. Each divider 44, the reciprocal of the distance d _i is calculated, and its output to the adder 45. The adder 45 calculates the sum of outputs of all of the divider 44, and outputs the output to m _c multipliers 46. Each multiplier 46 multiplies the output of the corresponding distance calculator 43 and the output of the adder 45 and inputs the output to the corresponding divider 47. Each divider 47 calculates the reciprocal of the output of the corresponding multiplier 46. Finally, in the sub-classifying unit 23, the output of each divider 47 is defined as a single feature similarity ν _i (1 ≦ i ≦ _mc ) of each category with respect to the feature vector X of the input pattern. It is output to the calculation unit 26. That is, in-group similarity ν _i (1 ≦ 1) of each category in each sub-classification unit 23
i ≦ m _c )

[0068]

(Equation 12)

Can be expressed as follows. The category reference feature vector indicating the representative value of each category stored in the category dictionary 42 is output from the distance calculation unit 43 corresponding to each category for an input pattern having each category in the category group. Is designed in advance to generate a minimum output as compared with the output of the other distance calculation units 43.

The design method of these category reference feature vectors is performed, for example, by a learning algorithm called a learning vector quantization method (LVQ). For the learning vector quantization method, see, for example, "Le
arning Vector Quantization for Pattern Recognitio
n ", Helsinki University of Technology, Report TKK-
This is shown in F-A601 (1986.11).

Hereinafter, the learning vector quantization method will be briefly described. First, a category reference feature vector W _i having m _c category labels is prepared. The initial value for this W _i, from the pattern for the category dictionary design consisting pattern set in each category group, a feature vector and the pattern selected arbitrarily for each category, in the design of the group dictionary 11 A reference feature vector obtained from a conventional clustering method such as the K-means algorithm described above is used. Then, the selected feature vector X of the pattern with any one category C _X from the pattern for the category dictionary design, this X sequentially, the following steps are repeated. (1) Select the category reference feature vector W _C closest to X. However, the category label of this W _C is C _C. (2) If C _X = C _C ,
W _C is close to the direction of the X. On the other hand, if C _X ≠ C _C , W _C is moved away from X. Also, the category reference feature vector other than W _C is not updated.

The updating of the above-mentioned category reference feature vector when X is presented is repeatedly performed for all the prepared patterns for category dictionary design.

By designing the category reference feature vector of the category dictionary 32 in this manner, the category reference pattern having the label of each category is always added to the feature vector of the pattern having each category in the category group. The feature vector is located at the closest distance. Therefore, the category of the input pattern can be recognized in each category group by selecting the distance calculation unit 43 that generates the minimum output from all the distance calculation units 43.

In the single feature similarity calculation unit 26, first, the multiplier 27 determines the group membership of the category group selected by the group selection unit 24 and the sub-classification units 23 corresponding to the category group (ie, , Multiply all the intra-group similarities obtained from the sub-classification unit 23) that has been input with the feature vector of the input pattern from the sub-classification input signal selection unit 25, and output their outputs to the category similarity calculation unit 28. That is, the number of multipliers 17 is prepared (the number of feature vectors for group reference × the sum of the number of feature vectors for category reference in each sub-classification unit), and the number of multipliers 17 of a certain category group p selected by the group selection unit 24 is prepared. group membership μ _p (1 ≦ _p ≦
m _r; m _r category group number) and the group in similarity ν _pq (1 ≦ q ≦ a certain category q obtained from the fine classification section 13 corresponding to the category group p m _c; m _C pattern data Category number) is input. That is, the output value ξ _pq of the multiplier 17 is

[0075]

(Equation 13)

Is expressed as follows. The category similarity calculation unit 28 classifies the output values of all the multipliers 27 collectively for each category, and selects a plurality of output values having a large output value.
Then, for each category, the total sum of these selected output values, which the similarity _{r i (1 ≦ i ≦ N} C of each category to the input pattern in a single recognition unit 12; the N _C category The number is output to the category identification unit 13. In addition, as a method of selecting a plurality of output values of the multiplier 27 for each category, a method in which the output value of the multiplier 27 is equal to or larger than a certain threshold value may be selected.

As described above, the single feature recognition unit 12
Performs a hierarchical classification on the feature vector obtained from a certain feature extracting unit 11 after performing a large classification, and finally integrating their outputs, thereby obtaining each of the input patterns for a certain feature. Find the similarity between categories.

The category discriminating unit 13 first inputs the similarity of each category to each feature vector of the input pattern obtained from each single feature recognizing unit 12 to the corresponding integrated similarity converting unit 14, Is converted into an integrated similarity. Next, the category determination unit 15 performs final recognition of the input pattern using the integrated similarities of all the categories obtained from each integrated similarity conversion unit 14. In other words, the category determination unit 15 performs classification by finally integrating classification results obtained by recognizing each feature vector of the input pattern in each single feature recognition unit 12, that is, similarity. It is. In addition, the integrated similarity conversion unit
14 converts the similarity obtained from the single feature recognizing unit 12 into an integrated similarity that allows the category discrimination unit 15 to successfully integrate the discrimination results for each feature vector.

Hereinafter, specific operations of the integrated similarity conversion unit 14 and the category judgment unit 15 will be described.

In the integrated similarity conversion unit 14, as shown in (Equation 14), the similarity r _ki (1 ≦ k ≦ N _F : N _F) of each category obtained from the corresponding single feature recognition unit 12 is a characteristic. Normalization is performed by dividing the number of types of vectors, 1 ≦ i ≦ N _C ; N _C is the number of categories) by the maximum value r _{km ax} of these similarities, which is then integrated into the integrated similarity t _ki ( Where 0 ≦ t _ki ≦
1, 1 ≦ i ≦ N _C ; N _C is the number of categories) and outputs the result to the category determination unit 15.

[0081]

[Equation 14]

In the category determining section 15, the adder 17 adds the integrated similarity t _ki of each category obtained from each integrated similarity converting section 14 for each same category as shown in ( _Equation 15). , The final similarity u _i (where 1 ≦ i ≦ N _C ; N _C is the number of categories). Finally, the maximum value selection unit 18, the final similarity u _i compared, and outputs as a final recognition result category corresponding to the similarity as the largest of them.

[0083]

(Equation 15)

The most important value for the category discriminating unit 15 in the category discriminating unit 13 to perform the final discrimination is the maximum similarity of each category obtained from each single feature recognizing unit 12 and its vicinity. , That is, the similarity values of the first, second,... Candidates obtained by each single feature recognition unit 12.
However, the maximum similarity obtained from each single feature recognition unit 12 is affected by the sum of the number of output units of each subclassification unit in the single feature recognition unit 12 (specifically, the output unit of the subclassification unit). The larger the sum of the numbers, the smaller the maximum similarity tends to be.) Therefore, if the similarity obtained from each single feature recognition unit 12 is directly output to the category determination unit 15, the maximum similarity is always averaged. Single feature recognition unit with high degree
The identification result from 12 has priority. Therefore, in this state, it is difficult to use a plurality of types of feature vectors, that is, it is difficult to correctly recognize a pattern that is erroneously recognized by a certain feature vector based on an identification result by a different feature vector.

However, the integrated similarity conversion unit 14 normalizes the similarity of each category obtained from each single feature recognizing unit 12 by the maximum similarity, so that the maximum of each single feature recognizing unit 12 is obtained. The similarity is always 1. Therefore, in the final category identification in the category determination unit 15,
The identification result of each single feature recognition unit 12 can be equally contributed, and the identification result from a certain single feature recognition unit 12 is not always given priority.

Here, the results obtained experimentally are shown. The pattern data to be recognized was multi-font (Mincho, Gothic, etc.) 3390 character types (symbols, alphanumeric characters, hiragana, katakana, Greek characters, JIS first-level kanji characters). The feature vectors for feature extraction are a 64-dimensional contour direction density feature and a 32-dimensional background density feature. The contour direction density is a feature amount indicating the direction and complexity of the character stroke in the local area, and the background density feature is a feature amount indicating the complexity of the character background. The method of detailed feature extraction is described in "Document Copy System (3)-Character Recognition Algorithm and its Hardware-" by Waki, Fujiwara, Takenouchi, Yokoe, Shimizu, IEICE General Conference, 1986, 1512,
6-154.

Single feature recognition unit corresponding to each feature vector
The group dictionary 21 in 12 and the category dictionary of each subclassification unit 23
42 was designed with 27 fonts in 8 fonts. here,
The fuzzy large classification unit 22 in each single feature recognition unit 12 sets the number of category groups (the number of feature vectors for group reference) to 7.
Also, the number of category reference feature vectors of each subclassification unit 23 is 10 in the category dictionary 42 corresponding to the contour direction density feature.
70,949,957,562,1029,920,996, the total number is 6483, and in the category dictionary 42 corresponding to the background density feature, 112
The total number of 2,902,633,1287,751,1090,984 was 6,769.

At the time of recognition, f in the fuzzy large classification unit 22 in (Equation 8) in each single feature recognition unit was 1.2, and f in (Equation 11) in the fine classification unit 23 was 1.3. The number of category groups selected by the group selection unit 24 was 7, that is, each feature vector was input to all the sub-classification units 23. In the category similarity calculator 28, the number of output values of the multiplier 27 selected for each category is set to four.

For comparison, a single feature recognizing unit 12 for recognizing the 96-dimensional feature vector by combining the two feature vectors into one was also designed. The same learning pattern was used for the design, the number of category groups was 7, the number of category reference feature vectors in the category dictionary 42 was 905,590,1255,382,933,1272,1083.
20 pieces. At the time of recognition, the same value was used for each parameter.

As character patterns for evaluation, 27120 characters in 8 fonts different from the data used for designing each dictionary were used.

As a result of the experiment, a recognition rate of 98.33% was obtained in this embodiment. In addition, the recognition rate of a recognition device that performs two feature vectors into one and performs recognition for comparison was 97.98%. That is, as in the present embodiment, it is more efficient to individually identify a plurality of feature vectors and perform recognition by integrating the identification results than to perform recognition processing by combining a plurality of feature vectors into one. It can be seen that is improved.

As described above, according to the present embodiment, a plurality of types of feature vectors are discriminated by the separate single feature recognizing unit 12 to determine the similarity of each category with respect to each feature vector. The unit 13 integrates the identification results obtained from each single feature recognition unit using all these similarities,
Recognize the final input pattern. Therefore, it is difficult to collectively use a plurality of types of feature vectors as in the conventional example. That is, "a pattern that is erroneously recognized with a certain feature vector is correctly recognized based on the identification result of a different feature vector". Can be easily done,
Highly accurate recognition can be realized.

In the conventional method, the time required for recognition increases due to the use of a plurality of types of feature vectors. However, in the present embodiment, since each single feature recognition unit performs the identification processing in parallel, the recognition process is not performed. The required time can be shortened as compared with the conventional example.

FIG. 4 shows a second example of the integrated similarity conversion unit 14 of the present invention.
FIG. 3 is a block diagram specifically showing the configuration of the embodiment. FIG.
, 51 is a similarity normalization unit, and a single feature recognition unit
The similarity of each category obtained from 12 is normalized by the maximum value among the similarities. Reference numeral 52 denotes an integrated similarity calculation unit that non-linearly converts the normalized similarity of each category and divides each of the non-linearly converted normalized similarities by the sum thereof to obtain an integrated similarity of each category. It is.

The integrated similarity converter configured as described above
The operation of the second embodiment will be described below.

As in the first embodiment, the similarity normalizing unit 51 first obtains the similarity r _ki (1 ≦ 1) of each category obtained from the corresponding single feature recognizing unit 12 as shown in (Equation 16). k ≦ N
_F: N _F is the number of types of feature _{vectors, 1 ≦ i ≦ N C;} N C performs normalization by dividing the number) of the category by the maximum value r _kmax in these similarities, normalizing it Similarity s
_ki (where 0 ≦ s _ki ≦ 1, 1 ≦ i ≦ N _C ; N _C is the number of categories) and outputs the result to the integrated similarity calculation unit 52.

[0097]

(Equation 16)

In the integrated similarity calculating section 52, the similarity normalizing section
The normalized similarity s _ki of each category obtained from 51 is nonlinearly transformed, and the normalized similarity of each category subjected to the nonlinear transformation is normalized (divided) by the sum of the normalized similarities. The degree t _ki (where 1 ≦ i ≦ N _C ) is output to the category determination unit 15. That is, if the nonlinear function is f (), the integrated similarity t _ki is

[0099]

[Equation 17]

Can be expressed as follows. Note that, when 0 ≦ x ≦ 1, the non-linear function f (x) is 0 ≦ f (x) ≦ 1, f
This is a monotonically increasing function that satisfies (0) = 0 and f (1) = 1. For example, as an example, there is a function as shown in (Expression 18). However,
0 ≦ a ≦ 1.

[0101]

(Equation 18)

Although the integrated similarity calculating section 52 obtains the integrated similarity t _ki for all the categories, a plurality of candidate categories having a large normalized similarity s _ki of each category are used as candidate categories. Individuals may be selected and an integrated similarity may be determined for these. At this time, if the set of candidate categories is ω, the integrated similarity t _ki

[0103]

[Equation 19]

Can be expressed as follows. As a selection method, a method in which the normalized similarity is equal to or larger than a certain threshold may be selected. By selecting a plurality of candidate categories in this way, the number of categories to be non-linearly converted is limited, so that the amount of calculation regarding the integrated similarity can be reduced. Usually, only the category having a large similarity has an effect on the recognition performance. Therefore, even if the integrated similarity is calculated for only the candidate categories in this way, the recognition processing can be sped up without substantially affecting the recognition performance.

By the way, in each single feature recognition unit 12,
When closely observing the similarity of each category when misrecognizing, (1) the magnitude of the maximum similarity is not so related to misrecognition. (2) The difference between the maximum similarity and the second largest similarity is It can be seen that there are many false recognitions when the child is small. Therefore, in the present embodiment, after normalizing the similarity obtained from each single feature recognition unit 12 by the maximum similarity, processing for emphasizing the difference between the maximum similarity and the second largest similarity, that is, the maximum similarity When the difference between the degree and the second largest similarity is small, the difference is made smaller, and the magnitude of the similarity itself is made smaller, so that a single feature that may cause erroneous recognition is obtained. A process of lowering the reliability (similarity) of the identification result obtained from the recognition unit 12 is performed. Therefore, when the category discrimination unit 15 performs the final discrimination, the result of the other single feature recognition unit 12 is regarded as important.

Here, results obtained experimentally are shown. The experiment was the same as in the first embodiment. However, in the integrated similarity calculation unit 52, the nonlinear function used is (Equation 1)
8), where a = 0.5.

As a result of the experiment, a recognition rate of 98.38% was obtained in this embodiment, and the recognition rate was improved by 0.05% as compared with the first embodiment.

As described above, according to this embodiment, by performing the above-described conversion on the similarity obtained from each single feature recognizing unit 12, a pattern that is erroneously recognized in a certain feature vector is different from the feature vector. Correct recognition based on the identification result of the first embodiment can be performed better than in the first embodiment, and more accurate recognition can be performed.

FIG. 5 is a block diagram specifically showing the configuration of the second embodiment of the category judging section 15 of the present invention. In FIG. 5, reference numeral 61 denotes an input unit for inputting the integrated similarity of each category obtained from the integrated similarity converter 14.
Reference numeral 62 denotes a multi-input one-output signal processing unit, and a lower-layer input unit 61 or a multi-input one-output signal processing unit 62 connected thereto.
Is multiplied by a weighting factor which is a degree of the connection, and the sum is summed and output. Here, the plurality of multi-input / one-output signal processing units 62 are connected to a network so as to have a layered structure, have no mutual coupling in each layer, and propagate signals only to the upper layer, so that the Is finally identified. Reference numeral 63 denotes a maximum value selection unit for selecting the maximum value from the outputs of the plurality of multi-input / output signal processing units 62 in the uppermost layer.

FIG. 6 is a block diagram specifically showing the configuration of the multi-input one-output signal processing unit 62. In FIG.
An input unit 71 inputs a signal. Reference numeral 72 denotes a weight coefficient storage unit which stores a weight coefficient for weighting a plurality of input signals from the input unit 71. A multiplier 73 multiplies the weight coefficient of the weight coefficient storage unit 72 by the input signal from the input unit 71. An adder 74 sums the output values of all the multipliers 73. A threshold processing unit 75 limits the output value of the adder 74 to a value within a certain range.

The category judging section 15 configured as described above
The operation of the second embodiment will be described below.

In the category determination section 15, first, the input section 61
Then, the integrated similarity t _{ki of} each category obtained from each integrated similarity converter 14 is input. That is, as many input units 61 as (the number of categories × the number of types of feature vectors) are provided,
Integrated similarity t of each category for each feature vector
_ki (1 ≦ k ≦ N _F ; N _F is the number of feature vector types, 1 ≦ i
≦ N _C; N _C is the number of categories) are input to the input unit 61 corresponding. Each multi-input / one-output signal processing unit 62
Is a weighting coefficient w which is the degree of the connection stored in the weighting coefficient storage section 72 and the output of the lower-layer input section 61 connected thereto or the multi-input / one-output signal processing section 62 as shown in FIG. _ij is multiplied by a multiplier 73, and each multiplier
After the sum of the outputs from 73 is calculated by the adder 74, it is converted by the threshold value processing unit 75 and the output value is output to the upper layer. That is,
The i-th multi-input one-output signal processing unit 62 of a certain layer shown in FIG.
The output value I _i of the input value I _j to the input unit 71, the weighting factor is the degree of the input and the connection that is connected to it w
_ij (the connection weight between the i-th multi-input one-output signal processing unit and the j-th input)

[0113]

(Equation 20)

Can be expressed as follows. The input / output characteristics of the threshold processing section 75 are shown in FIG. For example, the input / output characteristics of the threshold processing unit 65 for limiting the output to the range of (0, 1) are as follows:

[0115]

(Equation 21)

It can be expressed mathematically as shown below. Here, a is an input of the threshold value processing unit 65. The threshold processing unit
As the input / output characteristics of 75, a threshold function other than the above may be used.

The multi-input / one-output signal processing unit 62 in the uppermost layer
The number of categories of the pattern is set, and each multi-input / one-output signal processing unit 62 in the uppermost layer corresponds to each category of the pattern. The maximum value selection unit 63 selects the maximum value among the output values of the multi-input / one-output signal processing unit 62 in the uppermost layer, and outputs the corresponding category as the final recognition result of the input pattern.

The weighting factor of each multi-input / one-output signal processing unit 62 is determined in advance by using the integrated similarity of each category with respect to each feature vector of the input pattern, using the highest level layer corresponding to the category of the input pattern. Multi-input / one-output signal processing unit
Learning is performed so that 62 generates the maximum output. As a result, the category of the input pattern can be identified in the category determination unit 15. The design method of these weight coefficients is performed by a learning algorithm called an error back propagation method. Regarding the back propagation method, for example, DE Rumelhart, GE Hinton and
"Learning Representations by RJ Williams
Back-Propagating Errors, "Nature, vol.323, pp.533
-536, Oct. 9, 1986).

As described above, according to this embodiment, the similarities of all the categories obtained from the integrated similarity conversion units 14 are used, that is, all the similarities obtained as a result of identifying each feature vector are calculated. To calculate the final similarity of each category. Therefore, compared to the first embodiment, the category can be identified more flexibly by using a large amount of information, so that a more accurate recognition performance can be obtained.

FIG. 8 is a block diagram specifically showing the configuration of the third embodiment of the category judging section 15 of the present invention. 8, an input unit 91 includes a plurality of integrated similarity calculation units 14.
Are integrated into one and input as an integrated similarity vector. Reference numeral 92 denotes an integrated similarity dictionary in which a plurality of reference integrated similarity vectors indicating representative values of each category of the integrated similarity vector are stored. Reference numeral 93 denotes a distance calculation unit which calculates the distance between the integrated similarity vector and all the reference integrated similarity vectors stored in the integrated similarity dictionary 92. Reference numeral 94 denotes a minimum value selection unit that selects the minimum value from the distances obtained from the distance calculation unit 93.

The category judging section 15 configured as described above
The operation of the third embodiment will be described below.

In the category judging section 15, the input section 91 first integrates the integrated similarity obtained from each integrated similarity calculating section 14 into one as an integrated similarity vector T as shown in (Equation 22). It is input and output to N _C (N _C is the number of categories) distance calculation units 93.

[0123]

(Equation 22)

Each distance calculation unit 93 is provided with an integrated similarity dictionary 92
, The reference integrated similarity vector Z _i (1 ≦ i ≦ N _C ) indicating the representative value of the integrated similarity of each category stored in the equation (23) is read, and the distance d between T and Z _i shown in Expression (23) is read. _i is calculated and output to the minimum value selection unit 94.

[0125]

(Equation 23)

The minimum value selecting unit 94 selects the smallest one of the distances obtained from the distance calculating unit 93, and outputs the corresponding category as the final recognition result of the input pattern.

Note that the reference integrated similarity vector indicating the representative value of the integrated similarity of each category stored in the integrated similarity dictionary 92 is the same as the integrated similarity obtained from the input pattern. Distance calculator corresponding to
The output of 93 is designed in advance so as to generate a minimum output as compared with the outputs of other distance calculation units 93.

The design method of these reference integrated similarity vectors is performed by, for example, a learning algorithm called a learning vector quantization method (LVQ). For the learning vector quantization method, for example, the sub-classification unit 23 shown in FIG.
As explained in the design method of the category dictionary 42, T. Kohone
"Learning Vector Quantization for PatternRe
cognition ", Helsinki University of Technology, Rep
ort TKK-F-A601 (1986.11).

As described above, according to this embodiment, the similarities of all the categories obtained from the integrated similarity converters 14 are used, that is, all the similarities obtained as a result of identifying each feature vector are calculated. To calculate the final similarity of each category. Therefore, compared to the first embodiment, the category can be identified more flexibly by using a large amount of information, so that a more accurate recognition performance can be obtained.

[0130] In the single recognition unit 12 of this embodiment,
The sub-classification unit 23 is not configured as shown in FIG. 3, but has an input unit for inputting a feature vector of an input pattern output from the sub-classification unit input signal selection unit 25 as shown in FIG. Alternatively, a plurality of multi-input / one-output signal units connected to a network such that a signal propagates may be provided only in an upper layer without mutual coupling between layers.

[0131]

According to the present invention, a plurality of types of feature vectors are discriminated by separate single feature recognizing units to determine the similarity of each category to each feature vector. The recognition result obtained from each single feature recognition unit is integrated by using the similarity of, and the final input pattern is recognized. Therefore, it is difficult to collectively use a plurality of types of feature vectors as in the conventional example. That is, "a pattern that is erroneously recognized with a certain feature vector is correctly recognized based on the identification result of a different feature vector". This makes it easy to realize highly accurate recognition.

In the conventional method, the time required for recognition increases due to the use of a plurality of types of feature vectors. However, in the present embodiment, since each single feature recognition unit performs the identification processing in parallel with each other, the recognition is not performed. The required time can be shortened as compared with the conventional example.

[Brief description of the drawings]

FIG. 1A is a block diagram illustrating an embodiment of a pattern recognition device according to the present invention; FIG. 1B is a block diagram illustrating an embodiment of a single feature recognition unit of the pattern recognition device according to the present invention;

FIG. 2 is a block diagram showing an embodiment of a fuzzy large classification unit of the pattern recognition device according to the present invention.

FIG. 3 is a block diagram showing a first embodiment of a fine classification unit of the pattern recognition device according to the present invention.

FIG. 4 is a block diagram showing a second embodiment of the integrated similarity conversion unit of the pattern recognition device according to the present invention.

FIG. 5 is a block diagram showing a second embodiment of the category determination unit of the pattern recognition device according to the present invention.

FIG. 6 is a block diagram showing an embodiment of a multi-input / one-output signal processing unit of a category determination unit according to a second embodiment of the present invention;

FIG. 7 is an input / output characteristic diagram of a threshold value processing unit in a multi-input / one-output signal processing unit of a category determination unit according to a second embodiment of the present invention;

FIG. 8 is a block diagram showing a third embodiment of the category determination unit of the pattern recognition device according to the present invention.

FIG. 9 is a block diagram showing an embodiment of a conventional pattern recognition device.

[Explanation of symbols]

 11 Feature extraction unit 12 Single feature recognition unit 13 Category identification unit 14 Integrated similarity conversion unit 15 Category judgment unit 16 Similarity normalization unit 17 Adder 18 Maximum value selection unit 21 Group dictionary 22 Fuzzy large classification unit 23 Fine classification unit 24 Group selector 25 Sub-classifier Input signal selector 26 Single feature similarity calculator 27 Multiplier 28 Category similarity calculator

──────────────────────────────────────────────────の Continuing on the front page (72) Inventor Taiji Takagi 1006 Kazuma Kadoma, Osaka Prefecture Inside Matsushita Electric Industrial Co., Ltd. In-company (72) Inventor Toshio Niwa 1006 Kadoma, Kazuma, Osaka Prefecture Inside Matsushita Electric Industrial Co., Ltd. (56) References JP-A-5-303642 (JP, A) JP-A-4-31979 (JP, A (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06T ⁷ /00-7/60 G06K 9/00-9/03 G06K 9/46-9/52 G06K 9/62-9/82 JICST File (JOIS)

Claims (8)

(57) [Claims] 1. A plurality of feature extracting units for obtaining different feature vectors from an input pattern, and a plurality of single feature recognizing units for obtaining, from each of the feature vectors, a degree of similarity of the input pattern belonging to each category. And a category identification unit that identifies the input pattern using similarities obtained from a plurality of the single feature recognition units. The single feature recognition unit includes a set of patterns having similar feature vectors. A group dictionary in which a plurality of group reference feature vectors representative of a category group consisting of: and a degree to which an input pattern belongs to each category group using the group reference feature vector and the feature vector of the input pattern. A fuzzy large classifying unit for calculating the degree of group membership, and using the feature vector of the input pattern as a function of each category in which the input pattern is included in a category group. A plurality of sub-classification units for obtaining intra-group similarity that is a degree belonging to the category, a group selection unit for selecting a plurality of category groups from the group belonging degree, and the group selection information obtained from the group selection unit. A sub-classification unit for selecting a sub-classification unit for inputting a feature vector of an input pattern; an input signal selection unit; a group membership degree of the category group selected by the group selection unit; and an intra-group similarity obtained by the sub-classification unit. A single feature similarity calculating unit for calculating the similarity of each category with respect to the input pattern by using the single feature similarity calculating unit. A plurality of multipliers for multiplying all the intra-group similarities obtained from the sub-classification unit input with the feature vector of the input pattern from the classification unit input signal selection unit, and an output value of the multiplier for each category Multiple large ones A category similarity calculation unit for selecting the sum of the output values, and the category identification unit calculates the similarity of each category obtained from the single feature recognition unit corresponding to each of the feature vectors. A plurality of integrated similarity conversion units that convert to the integrated similarity by normalizing with the maximum value of the similarities among the similarities, and for each category using the integrated similarity obtained from the plurality of integrated similarity conversion units A pattern recognition apparatus, comprising: a category determining unit that determines a final similarity and identifies a final input pattern. 2. A similarity normalizing unit for normalizing the similarity of each category obtained from the single feature recognizing unit by a maximum value among the similarities; An integrated similarity for non-linearly converting the normalized similarity of each category obtained from the normalization unit and dividing the non-linearly converted normalized similarity by the sum thereof to obtain an integrated similarity of each category. The pattern recognition device according to claim 1, further comprising a calculation unit. 3. The method according to claim 1, wherein the category determination unit obtains the final similarity of each category by adding the integrated similarity of each category obtained from the plurality of integrated similarity calculation units for each same category. Item 3. The pattern recognition device according to item 1 or 2. 4. The category judging section comprises a plurality of multi-input one-output signal processing sections which have a layer structure, are not connected to each other, and are network-connected so that signals propagate only to the upper layer. The multi-input / one-output signal processing unit includes a weight coefficient storage unit that holds a plurality of weight coefficients, an input unit that inputs a plurality of input signals, and a weight coefficient stored in the weight coefficient storage unit. Multiplication means for weighting the input signal, addition means for adding the plurality of input signals weighted by the multiplication means, and a threshold processing section for limiting the output of the addition means to a value within a certain range. The pattern recognition device according to claim 1 or 2, wherein 5. A category judging section, wherein a plurality of reference integrated similarity vectors each representing a representative value of each category of the integrated similarity vector obtained by integrating the integrated similarities obtained from the plurality of integrated similarity calculating sections. A stored integrated similarity dictionary; and a plurality of distance calculation units for calculating distances between the integrated similarity vector and all of the reference integrated similarity vectors stored in the integrated similarity dictionary. Claim 1.
Or the recognition device according to 2. 6. A large fuzzy classification unit includes: a plurality of distance calculation units for calculating distances between a feature vector of an input pattern and all group reference feature vectors stored in a group dictionary; A plurality of dividers for calculating the reciprocal of the output, an adder for adding each output of the divider, a plurality of multipliers for multiplying the output of the adder and the output of the distance calculation unit, The pattern recognition device according to claim 1, further comprising a plurality of dividers for calculating a reciprocal of an output of the multiplier. 7. A category dictionary storing a plurality of category reference feature vectors indicating a representative value of each category of a pattern, and all categories stored in the feature vector and the category dictionary. A plurality of distance calculators for calculating the distance to the reference feature vector, a plurality of dividers for calculating the reciprocal of the output of the distance calculator, and an adder for adding each output of the divider. And a plurality of multipliers for multiplying an output of the adder and an output of the distance calculation unit, and a plurality of dividers for calculating a reciprocal of an output of the multiplier. 7. The pattern recognition device according to any one of claims 1 to 6. 8. The sub-classification unit comprises a plurality of multi-input / output signal processing units which have a layer structure, are not connected to each other, and are network-connected so that signals propagate only to upper layers. The multi-input / output signal processing unit includes a weight coefficient storage unit that holds a plurality of weight coefficients, an input unit that inputs a plurality of input signals, and a weight coefficient stored in the weight coefficient storage unit. Multiplication means for weighting the input signal;
7. An apparatus according to claim 1, further comprising an adding means for adding a plurality of input signals weighted by said multiplying means, and a threshold processing unit for limiting an output of said adding means to a value within a certain range. The pattern recognition device according to any one of the above.