JPH064677A - Recognition judgement device - Google Patents

Abstract

PURPOSE:To provide a recognition judgement device where time required for calculating an output category is short for an input feature vector and a parameter can automatically be optimized. CONSTITUTION:At the time of outputting the category to which a twodimensional input feature vector (x, y) belongs, a first threshold comparator outputs the selection signal of a second threshold comparator when a feature element (x) is smaller than a threshold t1, and outputs the selection signal of a second attribute memory B when the feature element (x) is more than the threshold t1. The threshold comparator where the selection signal coincides with the value of the peculiar selection signal or an attribute memory executes the processing of a next stage, and the threshold comparator outputs one of the two selection signals in the same way as the operation of the first threshold comparator. The attribute memory outputs the category and recognizes and outputs the category to which the input feature vector belongs by repeating the comparison of the sizes. Thus, the time required for the calculation of the output category is shortened since the output category can be obtained only by the comparison of the sizes.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a recognition / judgment device for recognizing an object, a phoneme, a situation, etc. with respect to an input image, voice, sensor signal or the like.

[0002]

2. Description of the Related Art As a conventional recognition / judgment device of this type,
For example, it is shown in JP-B-63-55106 or U.S. Pat. No. 4,326,259. FIG. 17 is a block diagram of a pattern type classification device forming a part of the conventional recognition determination device shown in the same patent. As shown in FIG. 17, in the conventional pattern type classification device, vectors S ₁ , S ₂ , ..., S _j ,.
S _N is commonly given to the assemblies 1, 2, ..., I, ..., M, and each assembly i has the structure shown in FIG. In the assembly i of the conventional pattern type classification device shown in FIG. 18, input vectors S ₁ , S ₂ , ..., S _j ,.
_.. , the signals of S _N are weighting stages A _i1 , A _i2 , ..., A _ij , ...,
It is applied to the matrix of A _iN . The coupling element A _ij forming the weighting stage is obtained by multiplying the signal S _j by the transfer function A _ij
i adds the outputs of the weighting stages to produce an intermediate output P _i '. The output P _i 'of adder i is fed to the stage labeled λ _i ,
In this λ unit, the output of adder i is multiplied by some selected coefficient. The result of this multiplication is applied to the next stage, denoted θ _pi . The θ _pi stage produces an output signal P _i when the output of the λ stage is equal to or greater than some selected level. The output of the θ _pi stage is provided to the control element as an inhibit input, and the control element is configured to feed back the output P _i 'of the adder _i to the coupling element when the output of the θ _pi stage does not occur. Based on the feedback P _i ',
The transfer function A _ij of the coupling element is modified based on the algorithm shown in ( _Equation 1).

[0003]

[ _Formula 1] A _ij (t) = A _ij (t-1) + δA _ij (t) However, δA _ij is expressed as in (Formula 2).

[0004]

[Equation 2] δA _ij (t) = ηP _i 'S _{j In} the initial stage of learning, the transfer function A _ij takes a random value, and for the first input pattern S ^k (1) belonging to the type k, The weights A _{ij of the} assembly i (i = 1, 2, ..., M) are corrected by the correction algorithms shown in (Equation 1) and (Equation 2). The repetition causes the output of the λ stage in one of the assemblies to be greater than or equal to the threshold θ _pi , producing the output P _i .

The conventional recognition judging device is constituted by a subordinate connection of such a pattern type classifying device and the pattern type recognizing device shown in FIG. P _i output from the pattern type classification device is applied to a multi-input OR circuit corresponding to the type k of the pattern type recognition device, and an output signal r _k is generated.

By the above operation, the output signal r _k is generated for the first input pattern S ^k (1) belonging to the type k. Similarly, for any input pattern belonging to any type, an output signal corresponding to that type is generated.

[0007]

However, in the above configuration, the input vectors S ₁ , S ₂ , ..., S _j ,.
_N signals are applied to the coupling elements A _i1 , A _i2 , ..., A _ij , ..., A _iN of all the assemblies i (1, 2, ... Since they are respectively multiplied by the transfer function, there is a problem that it takes a long time to calculate the output signal r _k .

Further, the performance of the conventional recognition judgment apparatus depends on the variable η in (Equation 2), the coefficient by which the output of the adder is multiplied in the λ unit, and the threshold value in the θ _pi stage. Therefore, there is a problem that trial and error are required for optimizing these parameters.

In view of the above problems of the conventional recognition / judgment apparatus, the present invention provides a recognition / judgment apparatus in which the time required for calculating an output signal is short and trial and error for parameter optimization are unnecessary. With the goal.

[0010]

The first aspect of the present invention is to provide a threshold comparator of the next stage or M threshold comparators for selecting an attribute memory according to the magnitude of one element of the input feature vector and the threshold. , A recognition determination device including N attribute memories that output categories for input feature vectors.

The second aspect of the present invention is to provide M threshold comparators for selecting a threshold comparator or an attribute memory at the next stage according to the magnitude of one element of the inputted feature vector and the threshold, and the inputted feature vector. To N attribute memories that output categories for, and a means for outputting a category mismatch signal when the categories output by the attribute memory do not match the given teacher category during learning, and in accordance with the category mismatch signals during learning. The recognition judgment device is provided with means for changing the threshold comparator of the next stage selected by the threshold comparator or the attribute memory.

[0012]

According to the structure of the first aspect of the present invention, the feature vector is first input to the first threshold value comparator, and the first threshold value comparator compares the specific element of the feature vector with the specific threshold value. Depending on the magnitude relation, a threshold comparator of the next stage or a selection signal of the attribute memory is output. The second to Mth threshold comparators each store a unique selection signal, and if the selection signal output by the first threshold comparator and the unique selection signal match, then a specific element of the feature vector A selection signal for the threshold comparator of the next stage or the attribute memory is output according to the magnitude relation with a specific threshold. Such an operation is repeated until one of the N attribute memories is selected. Each of the N attribute memories stores a unique selection signal, and if the selection signal output from the threshold comparator at the final stage matches the unique selection signal,
The category signals stored in each are output. As described above, the first invention outputs the category to which the feature vector belongs by repeating the magnitude comparison with the elements of the input feature vector, and in the category calculation, calculations such as transfer function calculation are performed. Since it is not necessary, the output category signal can be obtained in a short time.

The learning operation of the second aspect of the present invention is such that when the recognition operation for calculating the category to which the input feature vector belongs and the category obtained by the recognition operation and the teacher category do not match, the threshold comparator It consists of a threshold comparator in the next stage to be selected or an update operation for changing the attribute memory. In the recognition operation, the magnitude comparison with the elements of the feature vector input by the threshold comparator is repeated, and when the attribute memory is selected, the category to which the feature vector belongs is obtained as the category stored in the attribute memory.
The update operation starts by comparing the attribute memory category selected in the recognition operation with the teacher category.
If the category obtained by the recognition operation and the teacher category do not match, change the selection signal of the final threshold comparator so that one of the unused threshold comparators is selected. The comparator sets its selection signal so as to select the attribute memory selected in the recognition operation and one of the unused attribute memories, and sets the category of the unused attribute memory as the teacher category. As a result, the category of the attribute memory selected for the input feature vector matches the input teacher category. As described above, according to the second aspect of the present invention, the learning operation can update the selection signal output from the threshold comparator so that the category that matches the input feature vector and the teacher category is output. Yes, optimization of parameters is unnecessary.

[0014]

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

FIG. 1 is a block diagram of a recognition judgment apparatus according to an embodiment of the first invention. The recognition determination device of this embodiment is M
It consists of threshold comparators and N attribute memories. In FIG. 1, the feature vector input terminal 7 is connected to threshold comparators 1, 2, and 3, and the outputs of the threshold comparators 1, 2, and 3 are input to attribute memories 4, 5, and 6. , The result is to be output from the category output terminal 8.

FIG. 2 is a PAD diagram (Problem Analysis Diagram) showing the operation of the recognition determining apparatus of this embodiment. An outline of the operation of this embodiment will be described with reference to FIGS.

The feature vector (S ₁ ,
_{S 2, ····, S d,} ····, S D) is inputted to the first threshold comparator 1. The first threshold value comparator 1 outputs a selection signal of the threshold value comparator or the attribute memory at the next stage according to the magnitude relation between the specific element of the feature vector and the specific threshold value. The second to Mth threshold comparators each store a unique selection signal, and if the selection signal output by the first threshold comparator 1 and the unique selection signal match, a specific element of the feature vector According to the magnitude relationship between the threshold value and a specific threshold value, a threshold comparator of the next stage or an attribute memory selection signal is output. Such behavior
Repeat until one of the N attribute memories is selected. Each of the N attribute memories stores a unique selection signal.If the selection signal output from the threshold comparator at the final stage matches the unique selection signal, the category signal stored in each of them is changed to the terminal. Output from 8.

As described above, this embodiment outputs the category to which the feature vector belongs by repeating the magnitude comparison with the input elements of the feature vector. The configurations of the threshold value comparators 1, 2, 3 and the attribute memories 4, 5, 6 for realizing the operation of the present embodiment are shown below, and the operation thereof will be described in detail.

FIG. 3 is a block diagram of a threshold value comparator in the present embodiment which realizes the above-mentioned operation. In FIG. 3, a selection signal storage unit 9, a selection signal input terminal 15, and a comparator 10 are provided.
Is connected to the input terminal of and the input terminal 16 of the feature vector is
Connected to the characteristic element selecting means 11, the characteristic element selecting means
11 and the threshold value storage means 12 are connected to the comparator 13. Further, the comparators 10 and 13 are connected to the selection means 14 of the threshold comparator or attribute memory at the next stage, and the selection means 14 is connected to the output terminal 17 of the selection signal.

The selection signal storage means 9 stores a selection signal unique to each threshold value comparator. Further, the comparator 10 outputs a signal that can be output to the threshold comparator of the next stage or the selection means 14 of the attribute memory when the selection signal input to the terminal 15 and the unique selection signal stored in the storage means 9 match. Is output. The feature element selection means 11 is configured to include feature vectors (S ₁ , S ₂ , ...,
S _d, · · · ·, selecting an element S _d of a specific dimension from S _D). The dimension d selected by the characteristic element selection means 11 is a specific value for each threshold value comparator. The threshold storage unit 12 stores a threshold value to be compared with the selected feature vector element. The comparator 13 outputs a signal indicating the magnitude relationship between the specific element of the input feature vector and the threshold value to the threshold value comparator at the next stage or the selection means 14 of the attribute memory. The threshold comparator or attribute memory selection unit 14 in the next stage stores two threshold comparator or attribute memory selection signals, and when a specific element of the feature vector is smaller than the threshold value, the first selection signal. When the specific element of the feature vector is equal to or more than the threshold value, the second selection signal is output. In this way, the selection signal output by the threshold comparator is
The threshold comparator or the attribute memory having the same unique selection signal performs the next process.

FIG. 4 is a block diagram of the attribute memory in this embodiment which realizes the above-mentioned operation. In FIG. 4, the selection signal storage means 18 and the selection signal input terminal 21 are connected to a comparator 19, which is connected to the category storage means 20 and to the category output terminal 22. The selection signal storage means 18 stores a selection signal unique to each attribute memory. The comparator 19 outputs the output enable signal to the category storage means 20 when the selection signal input from the input terminal 21 and the selection signal stored in the storage means 18 match. Upon receiving the output enable signal from the comparator 19, the category storage means 20 outputs a signal indicating the category from the terminal 22.

The operation of the recognition / judgment device of this embodiment will be described below with reference to a specific example. FIG. 5 shows an example of the operation of the recognition determination apparatus of this embodiment for recognizing which of the three categories “A”, “B”, and “C” the two-dimensional feature vector belongs to. In FIG. 5, the input feature vector is represented by (x, y), and the elements of the input feature vector are limited to the ranges shown in (Equation 3).

[0023]

[Expression 3] 0 ≤ x <x ₀ 0 ≤ y <y _{0 In} the example of the operation of the recognition determination apparatus in FIG. 5, as shown in (a), (Expression 4), (Expression 5), (Expression 6) The category to which the feature vector belongs is recognized as follows.

[0024]

[Equation 4] When 0 ≤ x <t ₁ and 0 ≤ y <t ₂ , category "C"

[0025]

[Formula 5] When 0 ≤ x <t ₁ and t ₂ ≤ y <y ₀ , category "A"

[0026]

When t ₁ ≦ x <x ₀ , category “B” FIG. 5B shows the recognition of the present embodiment which performs the recognition shown in (Expression 4), (Expression 5), and (Expression 6). The elements of the feature vector and the threshold value to be compared by the threshold value comparator of the determination device, the selection signal of the threshold value comparator of the next stage or the attribute memory, and the category stored in the attribute memory are shown. In the recognition determining apparatus of the present embodiment, when a feature vector is input, first the first
Threshold comparator is selected. As shown in FIG. 5 (b),
The characteristic element selection means 11 of the first threshold value comparator 1 selects the element x of the characteristic vector, and the threshold value storage means 12 stores the threshold value t ₁ . The comparator 13 of the first threshold comparator 1 has the element x and the threshold t ₁
Compare with. The threshold comparator 2 next to the first threshold comparator 1
Alternatively, the attribute memory selection means 14 determines that the second means when x <t ₁ .
The selection signal of the threshold value comparator 2 is output, and conversely, when x ≧ t ₁ , the selection signal of the second attribute memory is output. Since the second attribute memory outputs a signal indicating the category "B", the recognition shown in (Equation 6) is possible. The second threshold comparator 2 selected when x <t ₁ is the element y of the feature vector and the threshold t.
₂ is compared, and when y <t ₂ , the selection signal for the third attribute memory is output, and conversely, when y ≧ t ₂ , the selection signal for the first attribute memory is output. Since the third attribute memory outputs a signal representing the category "C" and the first attribute memory outputs a signal representing the category "A", the recognitions shown in (Equation 4) and (Equation 5) are realized. it can.

As described above, according to this embodiment, the category to which the feature vector belongs can be obtained by repeating the magnitude comparison between the input element of the feature vector and the threshold value. Since the calculation of a category does not require calculation of a transfer function or addition, the category can be obtained in a short time.

FIG. 6 is a block diagram of the recognition determining apparatus of the first embodiment of the second invention. The recognition determination device of this embodiment is
It is composed of M threshold comparators, N attribute memories and selection changing means. In FIG. 6, the feature vector input terminal 30 is connected to threshold comparators 23, 24, 25, and the output terminals thereof are attribute memories 26, 27, 28 and selection changing means.
29, while the teacher category input terminal 31 is connected to the attribute memories 26, 27, 28, and 32 is the category output terminal.

In this embodiment, the operation for recognizing the category to which the input feature vector belongs is the same as that of the first aspect of the present invention, and the threshold for comparing the sizes and the selection of the elements of the input feature vector are learned. The point that can be automatically set by is the difference from the first aspect of the present invention. FIG. 7: is a PAD figure (Problem) showing the learning operation | movement of the recognition determination apparatus of the same Example.
Analysis Diagram). An outline of the learning operation of this embodiment will be described with reference to FIGS. 6 and 7.

The learning operation includes a recognition operation for calculating the category to which the input feature vector belongs, and a match / match between the category obtained by the recognition operation and the input teacher category /
The update operation includes changing a threshold value or the like according to the mismatch. The recognition operation is similar to the operation of calculating the category in the first aspect of the present invention, as can be seen by comparing FIGS. 2 and 7. That is, when the attribute memory is selected by repeating the size comparison between the feature element and the threshold value by the threshold value comparator, the category to which the feature vector belongs is obtained as the category stored in the attribute memory. On the other hand, the update operation is performed in the attribute memory selected in the recognition operation,
The process starts by comparing the category with the teacher category input from the teacher category input terminal 31. The attribute memory stores the maximum value and the minimum value for each dimension of the feature vector when each attribute memory is selected,
When the category obtained by the recognition operation matches the teacher category, the maximum value and the minimum value for each dimension are updated. On the contrary, if the category obtained by the recognition operation and the teacher category do not match, the selection signal of the threshold comparator at the final stage is changed so as to select one of the unused threshold comparators, and The threshold signal for use is set to select the attribute memory selected in the recognition operation and one of the unused attribute memories, and the category of the unused attribute memory is set to the teacher category. As a result, the category of the attribute memory selected for the input feature vector matches the input teacher category.

As described above, in the learning operation of the recognition / judgment apparatus of this embodiment, the selection signal of the threshold comparator is updated and the category of the unused attribute memory is set.
The category obtained in the recognition operation matches the teacher category. The configurations of the threshold value comparator and the attribute memory in the recognition and judgment apparatus of this embodiment are shown below, and the operation of the recognition and judgment apparatus of this embodiment will be described in detail.

FIG. 8 is a block diagram of the threshold value comparator in the recognition determination apparatus of this embodiment. In FIG. 8, a selection signal storage means 33 and a selection signal input terminal 39 are connected to a comparator 34, and a feature vector input terminal 40 is a feature element selection means 35.
Further, the comparator 34 and the input terminal 41 of the threshold value and the dimension change signal are connected to the characteristic element selecting means 35 and the threshold value storing means 36, respectively, and further, the characteristic element selecting means.
35 and the threshold value storage means 36 are connected to the comparator 37. Further, the comparator 37 is connected to the threshold comparator of the next stage or the selecting means 38 of the attribute memory. Further, the comparator 34 and the input terminal 42 for the selection change signal are also connected to the comparator 37. 43 is an output terminal for the selection signal.

The recognition operation of the threshold value comparator in this embodiment shown in FIG. 8 is similar to the operation of the threshold value comparator of the first present invention in FIG. That is, when the selection signal input from the input terminal 39 of the selection signal matches the selection signal unique to each threshold value comparator, the magnitude relationship between the specific element of the feature vector input from the feature vector input terminal 40 and the threshold value. In response to this, the selection signal output terminal 43 outputs the selection signal of the threshold value comparator or the attribute memory to be processed next. The difference between the threshold comparator in the present embodiment and the threshold comparator in the first aspect of the present invention shown in FIG. 3 is selected by the feature element selecting means 35 by the threshold and dimension change signal input from the input terminal 41. The threshold value stored in the element and threshold storage means 36 can be set, and the selection change signal input from the input terminal 42 can set the selection signal output by the selection means 38 of the threshold comparator or the attribute memory in the next stage. In point. By updating these parameters, it is possible to change the threshold comparator or attribute memory at the next stage selected by the threshold comparator. The recognition determination apparatus of the present embodiment selects the attribute memory for the input feature vector by repeating the operation of the threshold comparator selecting the threshold comparator or the attribute memory at the next stage. Therefore, the threshold comparator selected by the threshold comparator is selected. Alternatively, the attribute memory for the input feature vector can be changed by changing the attribute memory.

FIG. 9 is a block diagram of the attribute memory in the recognition / judgment device of this embodiment. In FIG. 9, the selection signal storage means 44 is connected to the comparator 45 together with the selection signal input terminal 54, the feature vector input terminal 55 is connected to the comparator 46, and the teacher category input terminal 56 is connected to the comparator 51. It is connected. The comparator 46 is connected to a maximum value and minimum value changing means 47 for each dimension, and the comparator 45 is a category storage means 50.
And the maximum value and minimum value changing means 47 for each dimension, the changing means 47 is connected to the maximum value and minimum value storing means 48 for each dimension, and its output terminal has the maximum difference with the comparator 46. It is connected to the dimension detecting means 52, and is connected to the threshold value and dimension changing means 53 together with the comparator 51. On the other hand, the comparator 51 is connected via the inverter 49 to the maximum value and minimum value changing means 47 for each dimension. The category storage means 50 is connected to the comparator 51. 57 is the category mismatch signal output terminal, 58 is the category output terminal,
Reference numeral 59 is an output terminal for a threshold value and dimension change signal. Figure 9
The operation of recognizing the attribute memory of this embodiment shown in FIG. 4 is the same as the operation of the attribute memory of the first present invention shown in FIG. That is, when the selection signal input from the input terminal 54 matches the selection signal unique to the attribute memory, the category signal is output from the input terminal 58. In the attribute memory in this embodiment,
The difference from the attribute memory in the first present invention shown in FIG. 4 is that the maximum value and the minimum value for each dimension of the feature vector input when each attribute memory is selected are input. When the stored teacher category does not match the stored category, the category mismatch signal, the threshold value, and the dimension change signal are output. The category mismatch signal is output to the selection changing means 29, which enables the selection changing means 29 to output the selection changing signal. The threshold and dimension change signals are output to the M threshold comparators, allowing the selected threshold comparators to change the thresholds and features. The learning operation of the attribute memory according to the present embodiment will be described in more detail in the description of a specific example of the learning operation with reference to FIGS. 11, 12, and 13.

FIG. 10 is a block diagram of the selection changing means 29 of the recognition and judgment apparatus of this embodiment. In FIG. 10, the selection signal input terminal 65 is connected to the selector 60 of the threshold comparator at the final stage, and the selector 60 and the selector of the unused threshold comparator are connected.
61 is connected to the selection signal output means 63, and further, a selector
61 and an unused attribute memory selector 62 are connected to a selection signal output means 63 and a selection change signal output means 64. Further, the selection signal output means 63 is connected to the selection signal output terminal 66, and the selection change signal output means 64 is connected to the selection change signal output terminal 67. Reference numeral 68 is an input terminal for the category mismatch signal.

The selection signal input from the input terminal 65 is sequentially input to the selector 60 of the final threshold comparator. The selector 60 of the threshold comparator at the final stage identifies whether the input selection signal is a signal for selecting the threshold comparator or a signal for selecting the attribute memory, and a selection signal before the selection signal for selecting the attribute memory. As a result, the selection signal of the final threshold comparator is obtained. The unused threshold value comparator selector 61 and the unused attribute memory selector 62 are set so as to be selected by the threshold value comparator or attribute memory selecting means 38 in the next stage of any threshold value comparator. It stores the threshold comparator and the selection signal of the attribute memory. The selection change signal output means 64 outputs an unused threshold value comparator or an attribute memory selection signal,
If any threshold comparator is set to select that threshold comparator or attribute memory, the selection signal will be stored by the unused threshold comparator selector 61 or the unused attribute memory selector 62. The selected signal is deleted from the memory. The selection signal output means 63 and the selection change signal output means 64 are provided with an output terminal 66 and an output terminal when the category mismatch signal input from the input terminal 68 is active.
A selection signal and a selection change signal are output from 67. The operation of the selection changing unit 29 of this embodiment will be described in more detail in the following description of a specific example of the learning operation using FIGS. 11, 12, and 13.

The learning operation of the recognition / judgment device of this embodiment will be described with reference to a specific example. 11, 12 and 13 are 2
Three-dimensional feature vector has three categories "A", "B", "C"
An example of the learning operation of the recognition determination device of the present embodiment for recognizing which one of them belongs will be shown. 11, 12 and 1
In 3, the input feature vector is represented by (x, y), and x and y are each limited to the range shown in (Equation 3). At the beginning of learning, the characteristic elements and the threshold values set in the characteristic element selection means 35 and the threshold value storage means 36 of the first threshold value comparator may be arbitrary, but FIG.
1, FIG. 12 and FIG. 13, the characteristic element is shown as x, and the threshold value is shown as 0. At the beginning of learning, the first threshold value comparator is set to output the selection signal of the first attribute memory regardless of the size of the feature element and the threshold value. At the beginning of learning, all the selection signals of the M threshold comparators are stored in the unused threshold comparator selector 61, and all the selection signals of the N attribute memories are stored in the unused attribute memory. Stored in the selector 62 of. FIG. 11 shows the operation when the recognition determination apparatus of this embodiment is given a two-dimensional feature vector (x ₁ , y ₁ ) and a teacher category “A” at the next time point. At this point, the first threshold comparator is set to select the first attribute memory regardless of the value of the input feature vector.
Attribute memory is selected. At this point, set the category of the first attribute memory to "A", and set the maximum value of x dimension (x _max ).
And the minimum value (x _min ) to x ₁ and the maximum value in the y-dimension
Set (y _max ) and minimum (y _min ) to y ₁ . At this point, the first attribute memory select signal is removed from the unused attribute memory selector 62.

FIG. 12 shows that at the next time point, the recognition and judgment apparatus of the present embodiment has a feature vector (x ₂ , y ₂ ) and a teacher category "
The operation when B "is given is shown. At this point, the first threshold value comparator is set to select the first attribute memory regardless of the value of the input feature vector. Therefore, the first attribute memory is selected.The category "A" stored in the first attribute memory is the teacher category.
Since it does not match B ", the comparator 51 of the first attribute memory
Outputs the category mismatch signal. At the same time, in the first attribute memory, the detection means 52 for the dimension having the maximum difference is the maximum value or the minimum value for each dimension stored by the means 48, and the feature vector (x ₂ , y ₂ ) for each dimension. The difference is calculated and the dimension with the maximum difference is output. That is, the difference in the x dimension is (x _min -x ₂ ) and (x _2-
x _max ), whichever is larger, and the difference in the y dimension is (y _min-
The larger value of y ₂ ) and (y ₂ -y _max ) is set as the larger value, and the larger dimension of the difference between the x-dimensional difference and the y-dimensional difference is output by the detection means 52 having the maximum difference. In the example of FIG. 12, (x ₂ -x ₁ )
Is larger than (y ₂ -y ₁ ), the detecting means 52 of the dimension having the largest difference outputs the dimension x. At the same time, the detection unit 52 for the dimension with the largest difference outputs the median value of the differences in the dimension with the largest difference as the threshold. In the example of FIG. 12, the detection means 52 of the dimension having the largest difference is the difference in the dimension x having the largest difference.
(x ₂ - x ₁₎ median _{(t 1 = (x 2 +} x 1) / 2) and outputs a. In this way, the first attribute memory outputs the threshold value and dimension change signal from the terminal 59. At the same time, the threshold value and dimension of the _first threshold value comparator are changed to t ₁ and x as the selection signal of the first threshold value comparator is output by the operation of the selection changing unit 29 described later. At the time of FIG. 12, the selection changing unit 29 outputs the selection signal of the unused first threshold value comparator as a selection signal from the terminal 66, and the unused second attribute memory. The selection signal of is output from the terminal 67 as a selection change signal.
This causes the second selection signal of the first threshold comparator to
Changed to 2 attribute memory selection signal. at this point,
The selection signal of the first threshold comparator is deleted from the selector of the unused threshold comparator, and the selection signal of the second attribute memory is deleted from the selector of the unused attribute memory. With the above operation, at the time of FIG. 12, the recognition determination device that outputs the category according to (Equation 7) for the input feature vector can be realized.

[0039]

Equation 7] When 0 ≦ x <t _1, when the category _{"A" t 1 ≦ x <} x 0, category "B" 13, at the next time point, the recognition judgment apparatus of the present embodiment, wherein The operation when a vector (x ₃ , y ₃ ) and a teacher category "C" are given is shown. In the example of FIG. 13, when x ₃ <t ₁ ,
Since the category "A" recognized according to (Equation 7) and the teacher category "C" do not match, the first attribute memory outputs a category mismatch signal. At the same time, in the first attribute memory, the detecting means 52 of the dimension having the largest difference is (x ₃ , y ₃ ) and (x
_It outputs y as the dimension with the largest difference from ₁ , ₁ , y ₁ ) and (t ₂ = (y ₃ + y ₁ ) / 2) as the median of the differences. From this point onward, the update operation shown in the latter half of the PAD diagram of FIG. 7 is executed. That is, the first selection signal output from the final-stage (first) threshold comparator is changed to an unused (second) threshold comparator selection signal. The threshold value and the characteristic element of the unused (second) threshold value comparator are set according to the threshold value t ₂ and the change signal of the dimension y output from the first attribute memory. Unused (second)
The first selection signal of the threshold comparator is set to the selection signal of the unused (third) attribute memory, and the second selection signal is set to the selection signal of the first attribute memory. Set to "C" category attribute memory unused (3), sets the maximum value / minimum value for each dimension (x _3, y _3). At this point, the selection signal of the second threshold comparator is deleted from the selector of the unused threshold comparator, and the selection signal of the third attribute memory is deleted from the selector of the unused attribute memory. By the above operation, FIG.
At the time point of, for the input feature vector,
It is possible to realize a recognition determination device that outputs categories according to (Equation 4), (Equation 5), and (Equation 6).

In this embodiment, when the attribute memory is selected at the time of recognition, the attribute memory outputs the category stored in the means 50. However, even if the attribute memory is selected, it is input. When the element of the feature vector is not included in the range of the maximum value and the minimum value for each dimension stored in the means 48, a signal of category "unknown" may be output. FIG. 14 shows a specific example of the recognition operation of the recognition determination device of this embodiment to show this operation. In FIG. 14, for the input vector (x, y) in the range represented by (Equation 3),
The first attribute memory is selected, and the first attribute memory stores the category "A" in the means 50, and the means 48 stores x _max as the maximum and minimum values for each dimension. ,
It is assumed that x _min , y _max , and y _min are stored. In this case, when the feature vector includes the feature vector in the shaded area in FIG. 14, that is, the area of (Equation 8), the category “A” is output to the area of (Equation 8). If it is not included, a signal of category "unknown" may be output.

Further, in the present embodiment, if the category stored in the attribute memory does not match the teacher category during learning, the selection signal of the threshold comparator of the next stage is changed by the operation of FIG. The threshold of the threshold comparator at the final stage may be changed. FIG. 15 shows a specific example of the learning operation of the recognition determination device of this embodiment to show this operation. FIG. 15 shows the operation when the teacher category “C” is given to the input feature vector (x ₄ , y ₄ ) at the next time point in FIG. In FIG. 15, since the magnitude relationship of (Equation 8) is established between each element of the input feature vector and the threshold value of the threshold value comparator, the first attribute for the input feature vector (x ₄ , y ₄ ). The memory is selected, and the category "A" stored in the first attribute memory and the teacher category "C" do not match.

[0042]

X ₄ <t ₁ y ₄ ≧ t _{2 In} this case, the threshold value t ₂ of the second threshold value comparator may be changed to the threshold value t ₂ ′ that satisfies (Expression 9).

[0043]

By Equation 9] t ₂ '> y ₄ which can without increasing the number of threshold comparator and attributes memory used, and categories and teacher category attribute memory is stored in the match.

In this embodiment, if the category stored in the attribute memory does not match the teacher category during learning, the selection signal of the threshold comparator in the next stage is changed by the operation of FIG. The threshold of the threshold comparator at the final stage may be changed. FIG. 16 shows a specific example of the learning operation of the recognition and determination apparatus of this embodiment to show this operation. FIG. 16 shows the operation when the teacher category “C” is given to the input feature vector (x ₅ , y ₅ ) at the next time point in FIG. 13. In FIG. 16, since the magnitude relationship of (Equation 10) is established between each element of the input feature vector and the threshold value of the threshold value comparator, the first attribute for the input feature vector (x ₅ , y ₅ ). The memory is selected, and the category "A" stored in the first attribute memory and the teacher category "C" do not match.

[0045]

X ₅ <t ₁ y ₅ ≧ t _{2 In} this case, the element selected by the second threshold value comparator from the feature vector is changed to x, and the threshold value t ₂ satisfies the threshold value t satisfying (Equation 11).
You may change to ₃ .

[0046]

By Equation 11] t _3> x ₅ which can without increasing the number of threshold comparator and attributes memory used, and categories and teacher category attribute memory is stored in the match.

Each means of the present invention may be realized by software using a computer, or may be realized by using a dedicated hardware circuit having each of these functions.

[0048]

As described above, according to the first aspect of the present invention, the category to which the feature vector belongs can be output by repeating the magnitude comparison with the input elements of the feature vector. Since an operation such as calculation of a function is not required, it has an advantage that an output category signal can be obtained in a short time.

According to the second aspect of the present invention, the selection signal output from the threshold comparator is updated so that the learning operation outputs a category that matches the input feature vector and the teacher category. And has the advantage that parameter optimization is unnecessary.

Since the recognition operation of the second aspect of the present invention is similar to the recognition operation of the first aspect of the present invention, it naturally has an advantage that an output category signal can be obtained in a short time.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a recognition determination device according to an embodiment of the first invention.

FIG. 2 is a PAD diagram showing an operation of the recognition determination device of the embodiment.

FIG. 3 is a configuration diagram of a threshold value comparator in the embodiment.

FIG. 4 is a configuration diagram of an attribute memory in the embodiment.

FIG. 5 is a configuration diagram showing a specific example of the operation of the recognition determination device of the embodiment.

FIG. 6 is a configuration diagram of a recognition determination device according to a first embodiment of the second invention.

FIG. 7 is a PA showing a learning operation of the recognition determination device of the embodiment.
FIG.

FIG. 8 is a configuration diagram of a threshold value comparator in the recognition determination device of the embodiment.

FIG. 9 is a configuration diagram of an attribute memory in the recognition determination device of the embodiment.

FIG. 10 is a configuration diagram of selection changing means of the recognition determination device of the embodiment.

FIG. 11 is a diagram showing a specific example of a learning operation of the recognition determination device of the same embodiment.

FIG. 12 is a diagram showing a specific example of a learning operation of the recognition determination device according to the embodiment.

FIG. 13 is a diagram showing a specific example of a learning operation of the recognition determination device according to the embodiment.

FIG. 14 is a diagram showing a specific example of the recognition operation of the recognition determination device of the embodiment.

FIG. 15 is a diagram showing a specific example of a learning operation of the recognition determination device of the example.

FIG. 16 is a diagram showing a specific example of a learning operation of the recognition determination device of the example.

FIG. 17 is a block diagram of a pattern type classification device that constitutes a conventional recognition determination device.

FIG. 18 is a block diagram of an assembly i in a conventional recognition determination device.

FIG. 19 is a block diagram of a pattern type recognition device that constitutes a conventional recognition determination device.

[Explanation of symbols]

1, 2, 3 ・ ・ ・ ・ M threshold comparators (means) 4, 5, 6 ・ ・ ・ ・ N attribute memories 11 ・ ・ ・ ・ Feature element selectors (means) 12 ・ ・ ・ ・ Thresholds Memory (means) 14 ・ ・ ・ ・ Next threshold comparator (means) or attribute memory selector (means) 20 ・ ・ ・ ・ Category memory (means) 23, 24, 25 ・ ・ ・ ・ M Threshold comparators (means) 26, 27, 28 ・ ・ ・ ・ N attribute memories 29 ・ ・ ・ ・ Changer (means) for selection 35 ・ ・ ・ ・ Characteristic element selector (means) 36 ・ ・ ・ ・Threshold memory (means) 38 ... Threshold comparator (means) at the next stage or attribute memory selector (means) 48 ..... Maximum value and minimum value memory (means) for each dimension

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Taiji Yuki, 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (8)

[Claims] 1. A threshold comparing means of the next stage or M threshold comparing means for selecting an attribute memory according to the magnitude of one element of the inputted feature vector and the threshold, and a category for the inputted feature vector are output. A recognition and determination device comprising N attribute memories. 2. A threshold comparing means of the next stage or M threshold comparing means for selecting an attribute memory according to the magnitude of one element of the inputted feature vector and the threshold, and a category for the inputted feature vector are output. N attribute memories, a means for outputting a category mismatch signal when the category output by the attribute memory does not match the given teacher category at the time of learning, and the threshold value comparing means according to the category mismatch signal at the time of learning. A recognition / judgment device comprising a threshold comparing means in the next stage for selection or a means for changing the attribute memory. 3. The recognition / judgment device according to claim 2, further comprising means for changing an element of a feature vector to be compared by said threshold value comparison means in response to said category mismatch signal during learning. 4. The recognition / judgment apparatus according to claim 2, further comprising means for changing a threshold value to be compared by said threshold value comparing means in response to said category mismatch signal at the time of learning. 5. A means for storing the maximum value and the minimum value for each dimension of the feature vector when the category output from the attribute memory and the teacher category match at the time of learning. The recognition determination device described. 6. When learning, the feature vector elements compared by the threshold value comparing means according to the category disagreement signal have the maximum difference between the maximum value and the minimum value for each dimension and the input feature vector element. 6. The recognition / judgment device according to claim 5, further comprising means for changing to the dimension of. 7. The threshold value compared by the threshold value comparing means according to the category non-matching signal at the time of learning is the element of the input feature vector in the dimension having the maximum difference, and the dimension in the dimension having the maximum difference. 7. The recognition and determination device according to claim 6, further comprising means for changing the value to a value between the maximum value and the minimum value. 8. When recognizing, when a feature vector which is not included in the range of the maximum value and the minimum value for each dimension of the feature vector stored by the means for storing the maximum value and the minimum value is input, the category is input. 6. The recognition and determination device according to claim 5, further comprising means for outputting a signal indicating that is unknown.