JPH0531792B2 - - Google Patents

Description

Detailed Description of the Invention (Technical Field) The present invention relates to a standard pattern correction method in pattern recognition, and particularly in pattern recognition such as character recognition and speech recognition, in which recognition processing is performed using a pattern described as a series of feature vectors. This invention relates to a method for modifying standard patterns when performing

(Prior Art) In conventional pattern recognition, the most important processing has been feature extraction processing and determination processing including various pattern matching methods. Its importance remains unchanged even today, but as the recognition rate of various pattern recognition improves and people become more aware of its practical use, the relative importance of the quality of standard patterns is increasing. When a pattern recognition 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 pattern to adapt to the environment. As an adaptive correction method, for example, if a standard pattern of a certain category C _is a multidimensional vector
The standard pattern correction data is given by multidimensional vectors X^ ₁ , ..., X^ _N , and the standard pattern X _C is the average vector of _M multidimensional vectors X ₁ , ..., When given, the multidimensional vector X^ _C obtained by modification is defined as X^ _C = 1/M + N (M・X _C + _N 〓 ⁱ⁼¹ X^i), then the multidimensional vector X^ _C is (M+N)
It is the average vector of the multidimensional vectors X ₁ , ..., X _M , X^ ₁ , ..., X^ _N. Therefore, if the multidimensional vector representing the standard pattern and the total number of data used to create and modify the standard pattern are stored,
The modified standard pattern is always given as the average vector of the data used so far.

On the other hand, when using patterns given as a series of feature vectors, it is not possible to simply average them as described above, and therefore it is not possible to modify a standard pattern that satisfies a certain appropriate evaluation standard.

(Objective of the Invention) The object of the present invention is to constantly adapt to the initially determined evaluation criteria in pattern recognition that performs non-linear matching by expanding and contracting patterns given as a series of feature vectors in one direction as described above. An object of the present invention is to provide a standard pattern modification method capable of modifying a standard pattern.

(Structure of the Invention) According to the present invention, in a standard pattern correction method that corrects a standard pattern necessary for pattern recognition using a pattern expressed as a series of feature vectors using a plurality of learning patterns or one learning pattern for correction. ,
standard pattern storage means for storing old standard patterns; cumulative learning pattern storage means for storing the number of learning patterns used to create the standard patterns; and each of the learning patterns for correction that are sequentially input. a mapping function generating means that inputs the standard pattern, performs expansion/contraction matching between the two, and outputs each mapping function indicating a correspondence when an optimal matching is achieved;
average mapping function generating means for generating a weighted average mapping function determined by each of the mapping functions, the number of learning patterns for correction, and the number of learning patterns used to create the standard pattern; and the weighted average mapping function. an elastic normalization means for elastically normalizing the standard pattern using the elasticity normalization mapping function to generate a temporary standard pattern; a second mapping function generating means for outputting each second mapping function indicating the correspondence when the optimum matching is obtained; a second elastic normalization means for elastically normalizing the correction learning pattern and outputting a correction elastic normalization pattern corresponding to each of the correction learning patterns; and sequentially inputting each of the correction elastic normalization patterns. Then, a weighted average pattern determined by the temporary standard pattern and the number of learning patterns used to create the standard pattern is output as a new standard pattern, and at the same time, the number of learning patterns used to create the standard pattern is output. and average pattern generation means for outputting the sum of the number and the number of learning patterns for correction as the number of learning patterns used to create a new standard pattern. can get.

(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 is a standard pattern modification that modifies a standard pattern necessary for pattern recognition using a pattern expressed as a series of feature vectors using a plurality of learning patterns or one learning pattern for modification. In this method, standard pattern storage means 2 stores standard patterns 22, cumulative learning pattern storage means 1 stores the number 21 of used learning patterns, and signals 20 of each of the learning patterns for correction are sequentially input. and the standard pattern signal 22, performs expansion/contraction matching between the two, and outputs each mapping function indicating the correspondence when optimum matching is achieved.
and a weighted average mapping function 24 determined by each of the mapping functions 23 and the number 21 of learning patterns used.
an average mapping function generating means 4 that generates a stretching normalization means 4 that generates a temporary standard pattern 25 by stretching and normalizing the standard pattern 22 using the weighted average mapping function 24 as a stretching normalization mapping function; The correction learning patterns 20 are sequentially input, and expansion/contraction matching with the temporary standard pattern 25 is performed, and each second mapping function 2 indicates a correspondence when optimal matching is achieved.
A second mapping function generating means 6 outputting a second mapping function 6 outputs each of the second mapping functions 26 as a mapping function for expansion and contraction matching, and expands and contracts the respective correction learning patterns 20 to perform expansion and contraction normalization. A second expansion/contraction normalization means 7 outputs a correction expansion/contraction normalization pattern 27 corresponding to the learning pattern 20, and each of the correction expansion/contraction normalization patterns 27 is sequentially input, and the provisional standard pattern 25 and the use The weighted average pattern determined by the number of learned patterns 21 and the new standard pattern 28
and at the same time output the sum of the number of learning patterns used 21 and the number of correction learning patterns as the number 29 of learning patterns used to create a new standard pattern. .

The cumulative number storage means 1 stores the cumulative number of data used to create and modify standard patterns, and the standard pattern storage means 2 stores the cumulative number of data used to create and modify 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 inputs the standard pattern 22 described by the series of feature vectors from the standard pattern storage means 2, and sequentially inputs the learning pattern 20 for correction also described by the series of feature vectors, A mapping function is determined and the mapping function 23 is sequentially output. The average mapping function generating means 4 receives the mapping function 23 inputted sequentially and the cumulative number 2 from the cumulative number storage means 1.
1 is input, and an average mapping function 24, which is a weighted average of the mapping functions, is output. The expansion/contraction normalization means 5 inputs the standard pattern 22, performs expansion/contraction normalization processing on the standard pattern 22 using the average mapping function 24 as a mapping function for expansion/contraction normalization, and generates and outputs a temporary standard pattern 25. is configured to do so. The second mapping function generation means 6 is configured to sequentially input the provisional standard pattern 25 and the correction learning pattern 20, and sequentially output a second mapping function 26 as a result.
The second expansion/contraction normalization means 7 inputs the second mapping function as a mapping function 26 for expansion/contraction normalization, and generates expansion/contraction normalization patterns 27 of the learning patterns 20 for correction that are sequentially input. is configured to print. The average pattern generation means 8 processes the expansion/contraction normalization pattern 27 of the learning pattern for correction that is inputted sequentially, the temporary standard pattern 25, and the cumulative number of the inputted learning patterns, and generates a weighted average pattern. Generated and modified standard pattern 2
Outputs 8 and at the same time newly updated cumulative number 29
is configured to output.

First, to explain the pattern recognition process in one embodiment of the present invention, a standard pattern 1 expressed as a series of feature vectors is
0 and the input pattern 11, the two patterns are matched by expansion and contraction, and a mapping function 12 for optimal matching is determined. This pattern recognition process has a large influence on the input pattern due to the difference in standard patterns. For example, in FIG. 3a, mapping functions 131, 132, 133, 134, and mapping functions 136, 137 between the standard pattern 13 and a different category B sample pattern. Note that the straight line 130 is a line connecting the starting point and ending point of each mapping relationship.

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

Next, in FIG. 3b, a standard pattern 14 of category A obtained by stretching and normalizing the standard pattern 13, a mapping function 141, 142 between the sample pattern of category A, and the standard pattern 14,
143, 144, 145 and mapping functions 146, 147 between the category B sample pattern and the standard pattern 14 are shown. In this way, in pattern recognition processing using the standard pattern 14, even if the degrees of similarity obtained when calculating the mapping functions 141 to 147 are approximately the same, the straight line 140 connecting the starting point and the 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 determining the similarity, the similarity is determined so that a penalty is added as the mapping function moves away from the straight line 140. and,
This will further increase the ability to determine similarity. Therefore, in this embodiment, pattern 14 is preferable to pattern 13 as the standard pattern, as shown in the examples of FIGS. 3a and 3b.

Next, this embodiment will be explained in more detail.

FIG. 4 shows mapping function generation means 3 in this embodiment.
shows. In FIG. 4, the mapping function generation means 3 is composed of a distance calculation section 31, a recurrence formula calculation section 32, an accumulated value storage section 33, a minimum value selection section 34, and a mapping locus storage section 35. A processing method that calculates the recurrence formula g(i, j) in be.

The correction learning pattern 20 is based on a signal j that is sequentially input.
= g ₁ (i), ...j = g _N (i), and the series of feature vectors A ¹ ,
Corresponding to A ² _, A ³ , ^. . .
These are signals corresponding to A ¹ ₀ , A ² ₀ , ..., A ^N ₀ . The distance calculation unit 31 inputs the above two signals and calculates the phase latitude d(i,
j) is calculated and output as a signal 311. The recurrence formula calculation unit 32 calculates the recurrence formula g(i, j)=d(i, j)+min(g(i-1,
j)), g(i-1, j-1), g(i-1, j-
2)), and d(i, j) is the signal 311,
min(g(i-1, j), g(i-1, j-1), g
(i-1, j-2)) as the signal 341,
The calculation result g(i, j) is output to the cumulative value storage section 33 as a signal 321. The minimum value selection unit 34 selects g(i-1, j) and g(i-
1, j-1) and g(i-1, j-2) as signal 331, signal 332 and signal 333, min(g(i-1, j), g(i-1, j-1) ,
g(i-1, j-2)) is output as a signal 341, and h(i, j) is output as a signal 342 to the mapping locus storage unit 35. When the recurrence formula calculation is completed, the mapping locus storage section 35 outputs a signal as the mapping function 23.
Output i′=ψ ₁ (i), ...i′=ψ _N (i).

The distance calculation section 31, the recurrence formula calculation section 32, the cumulative value storage section 33, the minimum value selection section 34, and the mapping locus storage section 35 are all configured using storage means, calculation means such as addition, and comparison means. be done.

Next, the average mapping function generation means 4 sets the number of data used for creation to M, the learning pattern for correction j=g ₁ (i),
…, the mapping function corresponding to j=g _N (i) is i′=ψ ₁ (i),
..., i′=ψ _N (i), then the average mapping function is obtained as ψ _S (i)=1/M+N {M+ _N 〓 ⁿ⁼¹ ψ _o (i)} ...(Formula 1).

That is, the pattern 13 of category A shown in FIG.
Mapping _{functions 131 ,} . Assuming that the function obtained as the average of the mapping functions of category A is i=ψ(i), ψ(i) becomes ψ(i)=1/5 ₅ 〓 ^k=1 ψ _k (i).

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

The expansion/contraction normalization means 5 uses the standard pattern (j=g _S
(i)) as a series of feature vectors A ¹ ₀ , A ² ₀ , ..., A ^N ₀ , and the average mapping function (i=ψ _S (i)) 2
4 is input, and the standard pattern 22 is expanded/contracted and normalized using this mapping function 24 to obtain ψg^(i)=g _S (ψ _S (i)).

The resulting j=ψg^ _S (i) is sent out as a temporary standard pattern 25.

That is, the expansion/contraction normalization means 5 uses the function i=ψ(i) to calculate the standard pattern 13 as shown in FIG.
The standard pattern 14 is generated from the straight line 15 indicating the mapping function i=ψ(i). Expressed in the formula,
Standard pattern 13 is j=j ₁₃ (i), standard pattern 1
4 as j= ₁₄ (i), the standard pattern 14 is obtained from the standard pattern 13 by the nonlinear normalization process expressed as j ₁₄ (i)=j ₁₃ (ψ(i)). The standard pattern 14 is a second temporary standard pattern ψg^ _S (i)=g _S
(ψ _S (i)).

Furthermore, an example of the DP matching method, which is one of the expansion/contraction matching processes in this embodiment, will be explained with reference to FIG. In Figure 6, the standard pattern
A ₀ consists of a series of M-dimensional vectors A ¹ ₀ , A ² ₀ , ..., A ^N ₀ , and the input pattern A consists of M-dimensional vectors A ¹ , A ² ,
..., A ^N series. Also, let d(i, j) be the distance between an arbitrary vector A ^j ₀ of the standard pattern and an arbitrary vector A ⁱ of the input pattern.
By simple matching, input pattern A and standard pattern
The degree of difference D (A, A ₀ ) from A ₀ is determined, for example, by the following formula.

Dissimilarity D (A, A ₀ ) = _N 〓 ⁱ⁼¹ d (i, j) This formula is expressed as A ⁱ on the mapping function j = i in Figure 6a.
^{A i} _{0 and} ^A
If it is possible to make the input pattern A correspond to the standard pattern A0, the input pattern A can be partially expanded or contracted to match the standard pattern _A0 in order to find the degree of difference between the two patterns.

The DP matching method is a method for matching input patterns by partially expanding or contracting them.
In Figure b, g(N,
N), on the mapping function j=ψ(i)
Matching can be achieved by making A ⁱ and A ^j ₀ correspond.

g(1,1)=d(1,1) g(i,j)=d(i,j)+min{g(i-1,
j), g(i-1, j-1), g(i-1, j-2)} However, d(i, j)=∞(i≦0 or j≦
0).

Figure 6c is a diagram showing an example of the DP matching method for calculating the above recurrence formula, and the input pattern is 5.
one-dimensional vectors, that is, a sequence of scalar quantities {1, 2, 4, 5, 5}, and the standard pattern is also a sequence of 5 {1, 2, 3, 4, 5},
(i, j) is (1, 1), (2, 2), (3, 4),
Stretching and contraction matching is performed on mapping functions that are (4, 5) and (5, 5).

Figure 6d shows that the recurrence formula calculation is performed within the range of i-Δ≦j≦i+Δ in order to reduce the amount of calculation in the above recurrence formula calculation, and generally, in the DP matching method, within this range is called a matching window, and it actually aims to improve the efficiency of the calculation amount. The above recurrence formula is only for calculating the degree of dissimilarity, but min(g(i-1, j), g(i-1, j-1),
g (i-1, j-2)) = g(i-1, j _(i-1) ) (where j _(i-1) is either j, j-1, or j-2) By calculating the function h(i, j) with h(i, j)=j _(i-1) , after calculating the degree of dissimilarity, the value of h(i, j) can be calculated as h(N , N) to h(1, 1), the mapping function can be obtained. For example, in the example in Figure 6c, h (5, 5) h (5, 5) = 5, h (4, 5) = 4, h (3, 4)
= 2, h = (2, 2) = 1, so the mapping functions (i, j) are (1, 1), (2, 2), (3, 4), (4, 5),
(5, 5) is found.

FIG. 7 is a diagram showing an example of expansion/contraction normalization processing, where X(i) (1≦i≦16) is the input pattern, Y(j)
(1≦j≦16) is a stretching normalization pattern, and j=ψ(i)
is a mapping function for stretch normalization. In this example, Y(j) is determined by the following law.

(1) j＝ψ(i)＞ψ(i−1) and ψ(i)＜ψ(i+1
When Y(j)=X(i) (2) When j=ψ(i)=ψ(i-1)+2, Y(j-
1)=X(i) (3) When j=ψ(i)=ψ(i-1)<ψ(i+1) Y(j)=X(i) Referring again to Figure 1, the second The mapping function generation means 6 may have the same configuration as the mapping function generation means 3, and the temporary standard pattern 25 of the expansion/contraction normalization means 5
(14 (Category A) in FIG. 3) and the correction learning pattern 20 are sequentially input. As shown in FIG. 3b, the second mapping function generating means 6 generates a second mapping function between each learning pattern and the standard pattern 14 as j=ψ^ ₁ (i), j=ψ^ ₂ (i)...j=ψ^ ₅ (i), this is the mapping function 141, 14 in Figure 2b
2,143,145, and the second mapping function 26
It is sent to the second expansion/contraction normalization means 7 as a.

The second elastic normalization means 7 may have exactly the same configuration as the elastic normalization means 5, and the learning pattern 20 and the second mapping function 26 are input, and the mapping function j=ψ _S (i ), the correction learning pattern j=f ₁ (i)...f ₅ (i) is expanded/contracted and normalized to obtain j=f _S (i)=f _S (ψ^ _S (i)).

Further, j=f^ _S (i) obtained in this way is sent to the average pattern generating means 8 as the expansion/contraction normalization pattern 27.

That is, the second expansion/contraction normalization means 7 sets each of the learning patterns to j=f ₁ (i), j=f ₂ (i), ..., j=f ₅ (i),
Each pattern obtained by the second nonlinear normalization process is j
= f^ ₁ (i), j = f^ ₂ (i), ..., j = f^ ₅ (i), then f _k (i) = f _k (ψ _k (i)). (k=1, 2, 3, 4, 5) Therefore, the standard pattern j=f _S (i) is the nonlinear normalized pattern j=f^ _k
It is determined by the average of (i) as shown in the following formula.

f _S (i)=1/5 ₅ 〓 ^k=1 f^ _k (i) If the pattern j=j(i) to be recognized can be treated as a binary pattern, the average of the nonlinear normalized patterns The pattern obtained by non-linear normalizing the provisional standard pattern using the mapping function j=ψ(i) are similar, but the pattern to be recognized j=j(i) When is treated as a multivalued pattern, it is essential to obtain a standard pattern by averaging the patterns j=f^ _k (i). Although the explanation has been given here assuming that k=5, the value of k is not specifically set to 5 and is arbitrary. The average pattern generation means 8 is constructed using a storage means and arithmetic means such as addition, multiplication, division, etc., and calculates the temporary standard pattern j=g^ _S (i) and the learning pattern for correction j=g _o (i) ( n=1),
..., N) is defined as the mapping function between i'=ψ _o (i) (n=1,
..., N), then the new standard pattern j=〓g _S (i)
is defined as follows.

g〓 _S (i)=1/M+N {M・g^ _S (i)+ _N 〓 ⁿ⁼¹ g _o (ψ^ _o (i))} (Equation 2) This procedure corrects the standard pattern. The resulting new standard pattern is the average of the patterns used to create and modify the standard pattern.

As described above, according to the present invention, an average mapping function is obtained from a plurality of mapping functions obtained from the cumulative number of data, the correction learning pattern, and the existing standard pattern, and the average mapping function is obtained from the existing standard pattern and the A provisional standard pattern is obtained from the average mapping function, each second mapping function obtained from the provisional standard pattern and each correction learning pattern is obtained, and then each expansion/contraction normalization pattern and the provisional By determining a weighted average pattern from the standard pattern and the accumulation and determining it as the standard pattern, a corrected standard pattern can be automatically obtained. By using this corrected standard pattern, the recognition rate can be improved, and since processing can be performed automatically, it is also possible to significantly reduce the number of man-hours required for correcting the standard pattern.

In the above explanation, it has been assumed that the standard pattern and input pattern series have the same length, but if the lengths are different, processing such as normalization of the length can be performed, and the shape of the pattern can be changed by preprocessing. If it changes significantly, it can be treated as a separate subcategory. In addition, stretch matching processing is not limited to DP matching, and even when using DP matching,
The recurrence formula used is not limited to the recurrence formula used in the example.

(Effects of the Invention) As explained above, the present invention has the effect of further increasing the ability to determine the similarity of input patterns by modifying the standard pattern.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a standard pattern correction method in an embodiment of the present invention, FIG. Figures 3a and b are diagrams showing differences in mapping functions due to differences in standard patterns, Figure 4 is a diagram showing mapping function generation means used in this example, and Figure 5 is a diagram showing nonlinear normalization processing using mapping functions. FIG. 6 is a diagram showing an example of the DP matching method, and FIG. 7 is a diagram showing an example of expansion/contraction normalization processing. DESCRIPTION OF SYMBOLS 1...Cumulative number storage means, 2...Standard pattern storage means, 3...Mapping function generation means, 4...Average mapping function generation means, 5...Stretching/contraction normalization means, 6...Second mapping function generation Means, 7...Second expansion/contraction normalization means, 8...Average pattern generation means, 31...Distance calculation section, 32...Recurrence formula calculation section, 33...Cumulative value storage section, 34...Minimum value Selection unit, 35... Mapping locus storage unit.

Claims (1)

[Claims] 1. A standard pattern storage means for storing standard patterns in a standard pattern modification method in which a standard pattern required for pattern recognition using a pattern expressed as a series of feature vectors is modified using a plurality of learning patterns or one learning pattern for modification. , a cumulative learning pattern storage means that stores the number of learning patterns used to create the standard pattern, each of the learning patterns for correction that are inputted sequentially, and the standard pattern; mapping function generating means for performing expansion/contraction matching and outputting respective mapping functions indicating correspondence when optimum matching is achieved; and for generating each of the mapping functions, the number of the correction learning patterns, and the standard pattern. an average mapping function generating means for generating a weighted average mapping function determined by the number of learning patterns used; and stretching and contraction normalizing the standard pattern using the weighted average mapping function as a mapping function for expansion and contraction normalization;
an expansion/contraction normalization means that generates a temporary standard pattern; and a second normalization means that sequentially inputs the learning pattern for correction, performs expansion/contraction matching with the temporary standard pattern, and indicates correspondence when optimal matching is achieved. a second mapping function generating means for outputting a mapping function; and stretching and normalizing each of the correction learning patterns by using each of the second mapping functions as a mapping function for expansion/contraction matching, and generating each of the correction learning patterns. a second elastic normalization means for outputting a correction elastic normalization pattern corresponding to the correction elastic normalization pattern; and a second elastic normalization means for sequentially inputting each of the correction elastic normalization patterns to create the temporary standard pattern and the standard pattern. A weighted average pattern determined by the number of learning patterns used is output as a new standard pattern, and at the same time, the sum of the number of learning patterns used to create the standard pattern and the number of correction learning patterns is calculated as the new standard pattern. 1. A standard pattern correction method for pattern recognition, comprising: average pattern generation means for outputting the number of learning patterns used to create the pattern.