JPS60181988A - Standard pattern correction system in pattern recognition - Google Patents

Abstract

PURPOSE:To improve deciding capability of similarity with respect to an input pattern by correcting a standard pattern. CONSTITUTION:A provisional standard pattern 25 of an expansion/contraction normalizing means 5 and a correction learning pattern 20 are sequentially inputted to the 2nd map function generating means 6. Assuming that the 2nd map function of each learning pattern and a standard pattern 14 is j=phi1(i), j=phi2(i)... phi5(i) in terms of the 2nd map function generating means 6, it turns out map functions 141, 142, 143 and 145, which are transmitted to the 2nd expansion/contraction normalizing means 7 as the 2nd map function 26. The 2nd expansion/ contraction normalizing means 7 has the same constitution as that of the expansion/contraction normalizing means 5, and the learning pattern 20 and the 2nd map function 26 are inputted to said means 5. By utilizing a map function j=phis(i), a correction learning pattern j=f1(i)...f5(i) is subjected to expansion/ contraction normalization, and j=fs(i)=fs[phis(i)] can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to a standard slam correction method in button recognition, and particularly in button recognition such as character recognition and voice recognition.
This invention relates to a method for modifying a standard button when recognition processing is performed using a button described as a sequence of characteristic vectors.

(Prior Art) In conventional slam recognition, the most important processing is a feature extraction process and a determination process that includes various bang matching methods. Its importance remains unchanged even today, but as the recognition rate of various types of bangs improves and the idea of commercialization is increasing, the relative importance of the quality of standard slams is increasing. When the device is actually used, a satisfactory recognition rate may not be obtained depending on the environment, and in that case, it is desirable to be able to modify the standard button to adapt it to the environment. As an adaptive correction method, for example, a standard baton of a certain category C or a multidimensional vector XC is given,
The data for standard slam correction is a multidimensional vector t1, ・
..., is given by the father N, and the standard baton xc is M multidimensional vectors xll...

When given by the average vector of XM, the multidimensional vector Xct- obtained by modification.

, the multidimensional vector xc is defined as (M+N) multidimensional vectors x1 . ...I xMI xll...
・.

It is the average vector of xN. So standard slam? If the multidimensional vector shown and the total number of data used to create and modify the standard button are stored, the standard button obtained by modification is always given as the average vector of the data used so far.

On the other hand, if there is a batan Kongo given as a series of feature vectors, it is not possible to simply average it as described above.
Therefore, it was not possible to completely modify the standard slam to meet certain appropriate evaluation criteria.

(Objective of the Invention) The object of the present invention is to perform non-linear matching by expanding and contracting batons given as a series of feature vectors in one direction as described above, in order to always adapt to the initially determined evaluation criteria. An object of the present invention is to provide a standard button correction method that can correct a standard pattern.

(Structure of the Invention) According to the present invention, in a standard batan correction method that corrects a standard batan meeting necessary for recognizing a batan that has been found as a series of feature vectors, using a plurality of m or one learning pattern for modification.

Standard baton storage means for storing old standard battens; cumulative learning batten storage means for storing the number of learning battens used to create the standard pattern; and each of the learning battens for correction that are inputted sequentially. and the standard button, performs expansion/contraction matching between the two, and outputs each mapping function indicating a correspondence when an optimal matching is achieved; each of the mapping functions and the learning pattern for correction; average mapping function generation means for generating a weighted average mapping function determined by the number of training patterns used to create the standard batten and the number of learning patterns used to create the standard batten; A telescoping normalization means that performs telescoping normalization to generate all provisional standard stamps, and the correction learning batt klllj is then manually adapted to perform telescoping matching with the provisional standard bat.

a second mapping function generating means for outputting each second mapping function indicating a correspondence when optimum matching is achieved; a second elastic normalization means for elastically normalizing the learning pattern for correction and outputting all correction elastic normalization buttons corresponding to each of the correction learning patterns; input, outputs a weighted average baton determined by the temporary standard baton and the number of learning patterns used to create the standard baton as a new standard batton, and at the same time outputs the learning pattern used to create the standard baton. and the number of learning patterns for correction as the number of learning patterns used to create all new standard patterns. A standard slam correction method for slam recognition is obtained.

(Example) Next, an example of the present invention will be described in detail with reference to the drawings.

FIG. 1 shows an embodiment of the invention. In FIG. 1, one embodiment of the present invention shows a standard baton 1 required for baton recognition using a baton represented as a sequence of IFF signature vectors.
- In the standard button correction method in which correction is made using a plurality of correction learning buttons or one correction learning button, the standard button storage means 2 for storing the standard patterns 22 and the number of used learning patterns 21t - the stored cumulative learning button; storage means 1;
Each signal 2o of the learning pattern for correction is inputted sequentially.
and the standard button signal 22, performs expansion/contraction matching between the two, and outputs all of the respective mapping functions indicating the correspondence amount when optimum matching is achieved; an average mapping function generating means 4 that generates a weighted average mapping function 24 determined by the number 21 of learning patterns used; and the weighted average mapping function 2411--stretching and contraction normalizing the standard pattern 22 as a mapping function for stretching and contraction normalization;
Temporary standard pattern 25t-generating expansion/contraction normalization means 5;
The learning pattern 20 for correction is firstly inputted, and expansion/contraction matching with the standard pattern 25 of the board is performed, and when the matching is optimal, each second mapping function 26t indicating color correspondence is outputted. The mapping function generating means 6 and each of the second mapping functions 26 are used as mapping functions for expansion/contraction matching to perform expansion/contraction normalization on each of the correction learning patterns 2ot to correspond to the respective correction learning patterns 2°. Correction expansion/contraction normalization button 2
The elastic normalization means 7 of the WJ2 outputs 7, and each of the correction elastic normalization buttons 27 is sequentially inputted, and a weighted average button determined by the standard pattern 25 of the board and the number 21 of the learning patterns used. is output as a new standard pattern 28, and at the same time, the sum of the number of used learning patterns 21 and the number of correction learning patterns is output as the number of learning patterns 29 used to create a new standard pattern. generating means 8.

The cumulative number storage means 1 stores the cumulative number of data used to create and modify the standard pattern, and the standard pattern storage means 2 stores the cumulative number of data used for creating and modifying standard patterns.The standard pattern storage means 2 stores the cumulative number of data used for creating and modifying standard patterns. Memorizes standard patterns. These storage means 1 and 2 can be constructed from various types of storage means commonly used. The mapping function generation means 3 generates a standard pattern 22t manually written by a series of feature vectors from the standard button storage means 2, and a learning pattern 20tl for correction similarly written by a series of feature vectors.
The mapping function 23 is sequentially manually calculated, and the mapping function 23 is sequentially output. The average mapping function generating means 4 inputs the sequentially inputted mapping functions 23 and the cumulative number 21 from the cumulative number storage means 1, and outputs an average mapping function 24 which is a weighted average of the mapping functions. ing.

The expansion/contraction normalization means 5 inputs the standard pattern 22, and calculates the average mapping function 24t--the expansion/contraction normalization mapping function.
It is configured to perform expansion/contraction normalization processing on the standard pattern 22 to generate and output a temporary standard pattern 25.

The second mapping function generating means 6 is configured to input the standard pattern 25 of the board and the learning pattern 20 for correction in order of t-1111N, and sequentially output a second mapping function 26 as a result. .

The second expansion/contraction normalization means 7 inputs the second mapping function as the expansion/contraction normalization mapping function 26, and presses the expansion/contraction normalization button 27 of the learning pattern 20 for correction which is inputted sequentially.
It is configured to generate and output sequentially. The average baton generation means 8 processes the expansion/contraction normalization pattern 27 of the learning pattern for correction that is inputted in sequence, the standard pattern 25 of the board, and the cumulative number of the inputted learning patterns,
Standard pattern 281 that has been modified by generating a weighted average baton
It is configured to output the newly updated cumulative number 29 at the same time.

To begin with, a description will be given of the baton recognition process in an embodiment of the present invention.As shown in FIG.
The two buttons are matched by expansion and contraction, and when an optimal match is achieved, a similar mapping function 12 is set. This slam recognition process has a large influence on manual slams due to the distortion of standard slams. For example, in FIG. 131, 132, 133, 134. and the relationship between the standard pattern 13 and the mapping functions 136 and 137 between sample patterns of different category B. Note that the straight line 130 is a line connecting the starting point and ending point of each mapping relationship.

In the conventional slam recognition method, the judgment process was performed only based on the similarity obtained by stretching/contracting matching, so the above 131
．． ..., when the similarity obtained when calculating the 137 mapping functions is almost the same value, category A and category B
This made it difficult to identify.

Next, in FIG. 3(b), a standard pattern 14 of category A obtained by performing expansion/contraction normalization on the standard pattern 13,
The sample slam of category A and the standard pattern 1
Mapping function with 4 141, 142° 143.144, 14
5 and shows the relationship between the mapping functions 146 and 147 between the category B sample button and the standard pattern 14. In this way, in the button recognition process using the standard pattern 14, even if the degrees of similarity obtained when mapping the mapping functions 141 to 147 are approximately the same, the straight line 140 connecting the starting point and ending point of each mapping function The category A and the category B can be distinguished from each other by the degree of separation of each mapping function, and when calculating the degree of similarity, the four-way function is used to Defining the degree further increases the ability to determine the degree of similarity.

Therefore, in this embodiment, as shown in the example of Figure 3) (b), the button 14 is preferable to the button 13 as the standard pattern.

Next, this embodiment will be explained in more detail.

FIG. 4 shows the mapping function generating means 3 in this embodiment.

In FIG. 4, the mapping function generating means 3 is a distance calculating section 31.
．． Recurrence formula calculation unit 32, cumulative value storage unit 33. It is composed of a minimum value selection unit 34 and a mapping locus storage unit 35, and calculates recurrence formulas 1i, j) in the explanation of the DP matching method,
This is a processing method in which a trajectory hu,j> obtained as a result of recurrence formula calculation is determined, and a mapping function is determined from this trajectory h(i,j).

1-The normal learning button 20 is sequentially inputted with a signal j=,9
tti),...j=,! i'5(1) corresponds to the series of feature vectors A'', A2..., AN, and the standard occurrence 22it of the storage means 2, (i'=Is (i),
Feature vector series A5. A:, . . . is a signal corresponding to Ao.

The distance calculation unit 31 inputs the above two signals and calculates the degree of difference d(i
, j) Perform all calculations and output as a signal 311. The recurrence formula calculation unit 32 calculates the recurrence formula g(1,j)=d(i,j)+m1n(Ji'(i-1
,j).

1i-1,j-1), 9(i-1,j-2)) *d(1,j)t-signal 311°m1n(
, 1il(i-1,j), 1i-i,j-t)+1(
1-1, j-2)> is inputted as a total signal 341 and outputted to the cumulative direct unique unit 33 as a 9(i,j)t-signal 321 as a result of the calculation. The minimum value selection unit 34 selects all signals 331 . Read as signal 332 and signal 333, min(g(it,j
L 1i-t,j-1), 1i-1,j-2)) are output to the mapping locus storage unit 35 as a total signal 341° and a h(t,t)i signal 342. When the recurrence formula calculation is completed, the mapping locus storage section 35 outputs the signal i/ as the mapping function 23.
= ψl mountain, . . . i' = ψN (outputs ri.

Note that the distance calculation unit 31. Recurrence formula calculation unit 32. Cumulative value record room 33. Minimum value selection section 34. The mapping locus storage section 35 is constructed using a storage means, arithmetic means such as addition, and a comparison means.

Next, the average mapping function generation means 4 calculates the number of data vi−M used for creation, and the correction learning pattern j=,! i'1(su.

..., j =,! Mapping function t corresponding to i'*(1)
-i'=91 (su.

..., if i′=ψN mountain, The average mapping function is obtained as .

That is, the button 13 of category IJ A shown in FIG. 3 ta+
Mapping function 131. ..., 135' respectively 1 = 91 (su, 1 = = 92 mountains, ..., 1 = 97 (i)
shall be. The function obtained as the average of the mapping functions of category A is:

Therefore, this ψ (my average mapping function 24) is sent out. Note that this average base one-quadrant function generating means 4 is constructed using a storage means and arithmetic means such as addition, addition, multiplication, and division.

The expansion/contraction normalization means 5 inputs the standard pattern (j''J's k% characteristic vector sequence Aj, A:...A) and the average mapping function (i=ψs[ij) 2
4 f manually, and using this mapping function 241i, the standard pattern 22 is expanded/contracted and normalized.

g8 mountains: 98 (98 mountains) t-obtained.

This results in J=,! The i's mountain is sent as a temporary standard pattern 25.

That is, the expansion/contraction normalization means 5 uses the function i=ψ(ri) to convert the standard pattern 14 from the standard pattern 13 through a straight line 15 indicating the mapping function i=ψ(ri), as shown in FIG.
generate. Expressed as a formula.

If the standard pattern 13 is a mountain j''jta, and the standard pattern 14 is a mountain j=L4, the standard pattern 14 is obtained from the standard pattern 13 by a nonlinear normalization process represented by the mountain ju=La(ψ(su)). The standard pattern 14
is the second temporary standard pattern gs(1)=,! i's(ψI
I(i)).

Furthermore, DP, which is one of the expansion/contraction matching processes in this embodiment,
An example of the matching method will be explained with reference to FIG. In Figure 6, the % standard pattern AO is the M-dimensional vector A
o, Ao, ..., A snow series is completed, Kabatan A
Suppose that it consists of a series of M-dimensional vectors A'', A2゜..., AM.Also, the distance ta(1, j).

If simple matching is done, input button A and standard pattern A (1
The degree of difference with 1) (A, Ao) is 5, for example, by the following formula.

This formula is expressed as the mapping function j=AI and A on the cup in Figure 6(a).
The degree of difference of one car pattern is calculated by making it fully correspond to Δ,
If it is possible to fully correspond bA' and 4 using the mapping function j = ψ() in the same figure, then in order to calculate the degree of difference between the two patterns, the input cover A'fr can be partially expanded and contracted to form the standard pattern A, Consistent with? Can be taken.

1) The P matching method is a method for matching by partially expanding and contracting the input cover.For example, in FIG. 6 [b], from the following initial value and recurrence formula, g(N.

N) t-by finding the mapping function j=ψ(A
Matching can be achieved by making I and AA correspond.

11.1)=d(1,1) g(1,j)=d(j,j)+m1n(,!i'(j-
1,j).

g (i-1, j-B,, 1i-t, j-z)) but +
d(i, j)=■ (i≦0 or j≦0).

Figure 6 (tc) is a diagram showing an example of the DP matching method for calculating the above recurrence formula.
5), and the standard pattern is the same five series (1, 2,
3, 4, 5).

(i, j) is (1, t), (2, 2), (3, 4).

Stretching and contraction alignment is performed on the mapping functions (4, 5) and (5, 5).

Figure 6 [d] shows that the recurrence formula calculation is performed within the range of i-Δ≦j≦i+Δ in order to reduce the total amount of calculation in the above recurrence formula calculation, and generally l) P matching method is used. So, is this range called a matching window, and does it actually improve the efficiency of calculation? I'm trying.

The above recurrence formula is only for calculating the degree of difference.

m1n(1!(11,j), ,li'(j -1,j
-1), g(i-1, j-2)) =,! i'(i-1,
j (+-1)) (where j (1-1) is j, j-x,
j-2).

As h(i,j)=j(+-1), from 1llh(N,N) of h(i,j) after the function h(i, T)% dissimilarity is found. The mapping function can be determined by resolving up to the 11th order h(1,1). For example, in the example of Fig. 6 tc), It(
5,5)h(5,5)=5. h(4,5)=4. h(3
゜4)=2. Since h (2, 2) = 1, the mapping function (i, j) is (1, i, (2, 2), (a, 4), (4゜5), (5
,5) Stop.

Figure 7 is a diagram showing an example of expansion/contraction normalization processing.
(iBi≦l≦16) is the input button, Y(j) (1≦j
≦16) is an expansion/contraction normalization button, and j=ψ(ri is a mapping function for expansion/contraction normalization. In this example, Y(jt is determined by the following rule.

(x)j=ψ(su〉ψ(i-1) and ψmountain ψ(i+1
When Y mountain = X (i) (z)j=ψ(su=?(i-1)+2 and tk
-1) =X(i) (a) When j=ψ(i)=ψ(i-1)<ψ(i+1), y(j)=Xt.Referring to Figure 1 again, the second mapping The function generation means 6 may have the same configuration as the mapping function generation means 3, and the temporary standard pattern 25 (14 (category IJA) in FIG. 3) of the expansion/contraction normalization means 5 and the correction learning button 20 are sequentially generated. is input. In this second mapping function generating means 6, as shown in FIG.
As shown in b), each of the learning buttons and the standard button 1
The second mapping function with 4 is J = 91 (S, J = ψ2ti)
−J = ψS (If done, this becomes the mapping function 141°142.143,145 in Figure 2 [b), and the second
is sent to the second expansion/contraction normalization means 7 as a mapping function 26.

The second expansion/contraction normalization means 7 may have exactly the same configuration as the expansion/contraction normalization means 5, and the learning button 20 and the second
The mapping function 26 is input, and the mapping function j=ψB<
Using
tilF'' stretching normalization, j=fs mountain=fs(ψ
Get S (su).

Furthermore, J=fs[ri obtained in this way is sent to the average button generation means 8 as an expansion/contraction normalization button 27.

That is, the second expansion/contraction normalization means 7 adjusts each learning button tj
= rt mountain, j = h(su,..., j = fs(i),
Each pattern obtained by the second nonlinear normalization process is j ”
f 1 (suIJ"'2[su, ...lJ='s(rit), then fh(su=f, (ψk(su). (k=1.2, 3, 4.5) Therefore, the standard pattern j=fs mountain is determined by the average of the nonlinear normalized pattern j=fk (r) as shown in the following equation.

The pattern to be recognized j = j (%) If il can be treated as a binary pattern, the pattern obtained by taking the average of the nonlinear normalized patterns and the temporary standard pattern are expressed by the mapping function j = ψ (rid The pattern obtained by non-linear normalization is a similar pattern, but the recognition target pattern j=
When j(ri) is treated as a multi-valued bump, it is essential to find the standard pattern by the average of the above pattern j=↑k(ri.Although we have explained here that 1(=5) , k is not specifically set to 5 and is arbitrary.

The average baton generation means 8 is constructed using a storage means and arithmetic means such as addition, multiplication, division, etc., and the temporary standard pattern j=
L(rito correction learning button j=g,,(i)(n=1.
・, N) mapping function ti′=ψ, ti)(n=
1. ..., N), then the new standard pattern j=g,
The mountain is defined as follows.

(Formula 2) The new standard pattern obtained by correcting the standard pattern through this procedure is the average of the patterns used to create and correct the a quasi-pattern.

According to the above explanation, according to the present invention, the cumulative number of data,
Obtained from the learning button for correction and the existing standard pattern,
An average mapping function is determined from a plurality of mapping functions, a temporary standard button is determined from the existing standard button and the average mapping function, each of which is obtained from the temporary standard button and each correction learning button. The second mapping function is calculated, and then the weighted average pattern is completely calculated from each expansion/contraction normalized button, the temporary standard button, and the accumulation, and the corrected standard button is automatically calculated by defining it as the standard button. Obtainable. By using this corrected standard baton, the recognition rate can be improved, and since it can be automatically processed, the number of man-hours required for correcting the standard baton can be significantly reduced.

In the above explanation, we have assumed that the standard batan and large kabatan series have the same length, but if the lengths are different, you can perform processing such as normalizing the length, and pre-processing can change the shape of the batan to be larger. If it changes, it can be treated as a separate subcategory.

Furthermore, the stretch matching process is not limited to l) P matching, and even when IJP matching is used, the recurrence formula used is not limited to the recurrence formula used in the embodiment.

(Effects of the Invention) As explained above, the present invention has the effect of further increasing the ability to determine the degree of similarity with respect to the input button by making all modifications to the standard pattern.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a standard batten correction method in an embodiment of the present invention, FIG. 2 is a diagram showing a mapping function when performing expansion/contraction matching of two battens expressed as a series of feature vectors, and FIG. Figures (a) and (b) are diagrams showing the differences in mapping functions due to differences in standard buttons, Figure 4 is a diagram showing the mapping function generation means used in this example, and Figure 5 is a diagram showing the mapping function generation means using the mapping function. FIG. 6 is a diagram showing an example of the DP matching method, and FIG. 7 is a diagram showing an example of the expansion/contraction sintering process. 1... Cumulative number storage means, 2... Standard pattern storage means, 3... Mapping function generation means, 4.
. . . average mapping function generation means, 5 . . . expansion/contraction normalization means, 6 . . . second mapping function generation means,
7...Second expansion/contraction normalization means, 8...
Average slam generation means, 31... Distance calculation unit, 3
2... Recurrence formula calculation unit, 33... Cumulative value storage unit, 34... Minimum value selection unit, 35...
...Mapping locus record part 1. ￥-1 time'￥-2 times l-5 prompt +1'J (b)

Claims (1)

[Claims] A standard pattern storage means for storing all standard bats in a standard baton modification method for modifying standard batons necessary for baton recognition using batons represented as a series of feature vectors by using a plurality of learning batons or a single modification learning batt;
A cumulative learning button storage means that stores the number of learning buttons used to create all of the standard drums, and inputting each of the 1-pressure learning drums and the standard drum that are sequentially input. mapping function generation means for performing expansion/contraction matching and outputting each mapping function indicating the corresponding value when an optimal matching is achieved; an average mapping function generating means for generating a weighted average mapping function determined by the number of learning patterns used in the standard batten, and an average mapping function generating means for generating a weighted average mapping function determined by the number of learning patterns used in the standard batten, and an average mapping function generation means that uses the weighted average mapping function as a mapping function for stretching/contracting normalization to stretch and normalize the standard batten, and generates a temporary standard batt. The expansion/contraction normalization means that generates all the data is sequentially inputted with the correction learning button, performs expansion/contraction matching with the temporary standard button, and outputs each second mapping function indicating the correspondence when the optimal matching is achieved. a second mapping function generating means for generating a second mapping function, and performing full expansion/contraction normalization of each of the correction learning buttons using each of the second mapping functions as a mapping function for expansion/contraction matching, and performing correction corresponding to each of the correction learning buttons. a second elastic normalization means that outputs all the elastic normalization buttons for correction; and a second elastic normalization means for outputting all the elastic normalization buttons for correction; and a learning pattern used to manually manually generate each of the correction elastic normalization buttons sequentially and to create the temporary standard button and all the standard buttons. In order to create a new standard button, output the weighted average button determined by the number of and the number of standard buttons as a new standard button, and at the same time calculate the sum of the number of learning patterns used to create the standard button and the number of correction learning patterns. The method is characterized in that it includes an average baton generating means that outputs the number of learning patterns used. Standard pattern correction method for baton recognition.