JP2011022621A - Pattern matching device and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce a matching error occurrence rate, even when a standard pattern is compressed. <P>SOLUTION: A pattern matching device 10 compresses the standard pattern by unifying elements similar to an adjoining feature element, into one feature element, in a pattern compression section 12, for each feature element of B1, B2 and B3 for composing the standard pattern. In a compression information creating section 13, a sequence of a compression ratio for each feature element of a compressed compression standard pattern, is created as compression information. The compressed compression standard pattern is stored in a storing section 14, by relating it to the compression information created for the compression standard pattern. A distance between the restored compressed standard pattern and an input pattern is calculated by a recurrence formula of dynamic programming, which has the compression information created for the compression standard pattern, as a variable. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a pattern matching apparatus and method used in the pattern recognition field for determining which input pattern of speech, characters, graphics, etc. is one of standard patterns assumed in advance.

  In the conventional pattern recognition field, dynamic programming (DP) is widely used as a pattern matching method (see, for example, Patent Document 1).

The principle of dynamic programming will be described using the speech pattern recognition field as an example.
The voice pattern A is expressed as the following equation (1).

In the formula (1), i {i = 1, 2,..., I} indicates time, and a indicates a feature element of the voice pattern A at time i.

Therefore, a standard pattern B expressed in a time series of feature elements similar to the voice pattern A is prepared for each of various words. This standard pattern B is expressed as the following equation (2).

In equation (2), j {j = 1, 2,..., J} indicates time, and b indicates a feature element at time j of the voice pattern B.

  In the pattern matching apparatus, the input speech pattern A is compared with the standard pattern B of various words, and the distance between the patterns is obtained. Then, the standard pattern with the minimum distance is output as a recognition result.

  In an actual voice pattern, time axis distortion occurs due to fluctuations in the speech rate. Dynamic programming is extremely effective as a technique for matching this time-axis distortion.

In dynamic programming, the recurrence formula shown in the following formula (3) is repeatedly calculated.

In the equation (3), d (i, j) is an inter-element distance between the feature element a _i of the voice pattern A and the feature element b _j of the standard pattern B. g (i, j) is an inter-element cumulative distance between the voice pattern A and the standard pattern B.

The pattern matching apparatus repeatedly calculates the recurrence formula (3). Then, as shown in FIG. 9, a path (DP-pass) that optimally associates a _i and b _j in the A and B planes is obtained.

  The uppermost expression of the recurrence formula (3) is the relationship between an arbitrary point (i, j) on the A and B planes shown in FIG. Is stipulated. The middle formula of the recurrence formula (3) defines the relationship between the point (i, j) and the point (i-1, j-1) adjacent to the lower left side. The lowermost expression of the recurrence expression (3) defines the relationship between the point (i, j) and the point (i-1, j) adjacent to the left side.

  In some cases, the tilt restriction may be added by omitting the uppermost expression from the recurrence formula (3). In this case, a relationship as shown in FIG. 11 is defined for an arbitrary point (i, j). The tilt limit is provided to limit extreme expansion and contraction.

  In general, a large number of standard patterns necessary for pattern matching are prepared. For this reason, it is considered to improve the efficiency of the area for storing the standard pattern by compressing the standard pattern. There are the following methods for compressing the standard pattern. That is, for each feature element constituting the standard pattern, a difference from an adjacent feature element is taken. If the difference is smaller than a preset threshold value, it is determined that adjacent feature elements are approximate. If there is an approximation, they are integrated into one feature element. Thus, the standard pattern is compressed.

JP 50-96104 A

  However, when the standard pattern is compressed by the method described above, time information useful for pattern recognition is lost. That is, since adjacent feature elements are integrated into one, a part of the time information that each element has disappears. If part of the time information is lost from the standard pattern, a matching error generally tends to occur. In particular, when many feature patterns that are partially approximated due to noise are recognized or when tilt restriction is provided, a matching error is noticeable.

  The present invention has been made based on such circumstances, and an object of the present invention is to provide a pattern matching apparatus and method that can reduce the occurrence rate of matching errors even when a standard pattern is compressed. It is what.

  The pattern matching apparatus of the present invention compresses a standard pattern by integrating, for each feature element constituting the standard pattern, one that approximates an adjacent feature element into one feature element. In addition, a series of compression ratios for each feature element of the compressed compression standard pattern is generated as compression information. Then, the compressed standard pattern is stored in association with the compression information generated for the compressed standard pattern. Then, the inter-pattern distance between the stored compressed standard pattern and the input pattern is calculated by a recursive formula of dynamic programming having as a variable the compression information generated for the compressed standard pattern.

  According to the present invention in which such measures are taken, it is possible to provide a pattern matching apparatus and method that can reduce the occurrence rate of matching errors even when a standard pattern is compressed.

The block diagram which shows the structure of the pattern matching apparatus which is embodiment of this invention. The figure for demonstrating the standard pattern compression process performed with the pattern matching apparatus. The schematic diagram which shows the data structure memorize | stored in the memory | storage part of the pattern matching apparatus. The block diagram which shows the principal part structure of the pattern matching process part in 1st Embodiment. The block diagram which shows the principal part structure of the pattern matching process part in 2nd Embodiment. The figure for demonstrating the 1st recurrence formula calculated in the pattern matching process part of 2nd Embodiment. The figure for demonstrating the 2nd recurrence formula calculated in the pattern matching process part of 2nd Embodiment. The figure for demonstrating the 3rd recurrence formula calculated in the pattern matching process part of 2nd Embodiment. The figure for demonstrating the conventional dynamic programming. The figure for demonstrating the recurrence formula calculated by the conventional dynamic programming. The figure for demonstrating the other example of the recurrence formula calculated by the conventional dynamic programming.

  Embodiments in which the present invention is applied to the field of speech pattern recognition will be described below with reference to the drawings. First, a first embodiment will be described with reference to FIGS.

  FIG. 1 is a block diagram showing a main configuration of a pattern matching apparatus 10 according to the present embodiment. The pattern matching device 10 includes a voice analysis unit 11, a pattern compression unit 12, a compression information generation unit 13, a storage unit 14, and a pattern matching processing unit 15.

  A voice signal M is input to the voice analysis unit 11. The sound signal M is an electric signal converted from sound input through a microphone. The voice analysis unit 11 analyzes the input voice signal M. Then, the feature element a of the input voice is extracted in time series to generate the voice pattern A represented by the above equation (1). The voice pattern A is output to the pattern matching processing unit 15.

  The pattern compression unit 12 receives a plurality of standard patterns B1, B2, B3,. Each standard pattern B1, B2, B3,... Is prepared in advance for each type of word. These standard patterns B1, B2, B3,... Are not compressed. The pattern compression unit 12 compresses each input standard pattern B1, B2, B3,.

The compression method will be described with reference to FIG. In the figure, b ₁ , b ₂ , b ₃ ,..., B ₉ indicate characteristic elements of an arbitrary standard pattern Bx. In the standard pattern Bx, the feature elements b ₁ , b ₂ , b ₃ ,..., B ₉ are arranged in time series in that order.

The pattern compressing unit 12 obtains distances between adjacent feature elements for the feature elements b ₁ , b ₂ , b ₃ ,..., B ₉ constituting the standard pattern Bx. When the distance is equal to or less than a preset threshold value, the plurality of adjacent feature elements are replaced with one average feature element b ′ _k (k = 1, 2, 3,...). By this processing, a plurality of feature elements are locally compressed into average feature elements. By applying this process to the entire standard pattern Bx, a compressed standard pattern B′x is generated.

In the case of the example in FIG. 2, the distance between adjacent feature elements b ₂ and b ₃ is equal to or less than a threshold value, and these feature elements b ₂ and b ₃ are average feature elements b ′ ₂ {b ′ ₂ = (b ₂ + b ₃ ) / 2}. Further, the distance between adjacent feature elements b ₄ , b ₅ and b ₆ is less than or equal to the threshold value, and these feature elements b ₄ , b ₅ and b ₆ are average feature elements b ′ ₃ {b ′ ₃ = (b ₄ + b ₅ + b ₆ ) / 3}. Further, the distance between the adjacent feature elements b ₇ and b ₈ is equal to or smaller than the threshold value, and these feature elements b ₇ and b ₈ become the average feature element b ′ ₄ {b ′ ₄ = (b ₇ + b ₈ ) / 2}. Has been replaced. Thus, the standard pattern Bx (= b ₁ , b ₂ , b ₃ ,..., B ₉ ) is the compressed standard pattern B′x (= b ′ ₁ , b ′ ₂ , b ′ ₃ , b ′ ₄ , b ′ ₅ ). Is compressed.

The compression information generation unit 13 includes a compression ratio n1, for each of the characteristic elements b ′ ₁ , b ′ ₂ , b ′ ₃ , b ′ ₄ , b ′ _{5 of} the compression standard pattern B′x compressed by the pattern compression unit 12. n2, n3, n4. An n5 series is generated as compressed information Nx. In the present embodiment, as shown in FIG. 2, each of the characteristic elements b ′ ₁ , b ′ ₂ , b ′ ₃ , b ′ ₄ , b ′ ₅ of the compressed standard pattern B′x is converted into the original standard pattern Bx. Of the number of feature elements representing the compression ratios n ₁ , n ₂ , n ₃ , n ₄ . n ₅ to be defined.

That is, since the feature element b ′ ₁ of the compressed standard pattern B′x represents only the feature element b ₁ of the standard pattern Bx, the compression ratio n1 is “1”. Similarly, since the characteristic element b ′ ₂ represents the characteristic elements b ₂ and b ₃ , the compression ratio n2 is “2”. Since the characteristic element b ′ ₃ represents the characteristic elements b ₄ , b _5, and b ₆ , the compression ratio n 3 is “3”. Since the characteristic element b ′ ₄ represents the characteristic elements b ₇ and b ₈ , the compression ratio n4 is “2”. Since the characteristic element b ′ ₅ represents only the characteristic element b ₉ , the compression ratio n5 is “1”.

  Thus, the compression information for the compression standard pattern B′x is Nx “1, 2, 3, 1, 2”.

  As shown in FIG. 3, the storage unit 14 associates the compression standard pattern B′x compressed by the pattern compression unit 12 with the compression information Nx generated by the compression information generation unit 13 for the compression standard pattern. Remember.

  As shown in FIG. 4, the pattern matching processing unit 15 includes a restoration unit 21, a smoothing processing unit 22, and a distance calculation unit 23.

Restoration unit 21 restores the compressed standard pattern B'x stored in the storage unit 14, and extends on the basis of the compression information Nx stored in association with the compression standard pattern B'x, the reference pattern B _x1 To do. For example, in the case of the example of FIG. 2, the compression information Nx (n ₁ , n ₂ , n ₃ ) related to the compression standard pattern B′x (b ′ ₁ , b ′ ₂ , b ′ ₃ , b ′ ₄ , b ′ ₅ ). , N ₄ .n ₅ ) is “1, 2, 3, 1, 2”.

Therefore, the characteristic element b ′ ₁ remains as it is. The feature element b ′ ₂ becomes _two adjacent feature elements b ′ ₂ and b ′ ₂ . The feature element b ′ ₃ becomes _three adjacent feature elements b ′ ₃ , b ′ ₃ , and b ′ ₃ . The feature element b ′ ₄ becomes two adjacent feature elements b ′ ₄ and b ′ ₄ . The characteristic element b ′ ₅ remains as it is. Thus, the compressed standard pattern B′x is the standard pattern B _x1 (b ′ ₁ , b ′ ₂ , b ′ ₂ , b ′ ₃ , b ′ ₃ , b ′ ₃ , b ′ ₄ , b ′ ₄ , b ′ ₅ ) Is restored.

The smoothing processing unit 22 performs a smoothing process using a low-pass filter on the standard pattern B _x1 restored by the restoration unit 21. In the restored standard pattern B _x1 , noise due to compression is generated. This kind of noise can be removed by performing the smoothing process.

The distance calculation unit 23 calculates the inter-pattern distance Gx between the speech pattern A input via the speech analysis unit 11 and the standard pattern B _x1 smoothed by the smoothing processing unit 22, using a well-known dynamic plan Calculated by the method. For example, the inter-pattern distance Gx is calculated by repetitive calculation of the recurrence formula shown in the formula (3).

The pattern matching processing unit 15 repeatedly executes the processing in the restoration unit 21, the smoothing processing unit 22, and the distance calculation unit 23 for all the compressed standard patterns B′x. Then, the standard pattern Bx ₁ that minimizes the calculated inter-pattern distance Gx is obtained and output as the recognition result G of the voice pattern A.

  As described above, in the pattern matching apparatus 10 of the present embodiment, when the standard pattern Bx is compressed to generate the compressed standard pattern B′x, the compression ratio for each feature element of the compressed standard pattern B′x is changed. A series is generated as compressed information Nx. The compression standard pattern B′x is stored in the storage unit 14 in association with the compression information Nx generated for the compression standard pattern B′x.

  When the voice signal M is input, the voice analysis unit 11 extracts the feature element a of the input voice in time series, and the voice pattern A is generated. The voice pattern A is output to the pattern matching processing unit 15. The pattern matching processing unit 15 performs the following pattern matching processing on all the compressed standard patterns B′x stored in the storage unit 14.

First, an arbitrary compressed standard pattern B′x and related compression information Nx are read from the storage unit 14. Then, the compressed standard pattern B′x is expanded based on the compression information Nx, and restored to the standard pattern B _x1 . Next, a smoothing process is performed on the restored standard pattern B _x1 . Thereafter, the inter-pattern distance Gx between the input speech pattern A and the smoothed standard pattern B _x1 is calculated by dynamic programming.

Thus, the pattern matching processing unit 15 calculates the inter-pattern distance Gx with the sound pattern A for each compressed standard pattern B′x. Then, the standard pattern Bx ₁ that minimizes the inter-pattern distance Gx is output as the recognition result G of the voice pattern A.

Thus, the compressed standard pattern B′x is not used as it is for the calculation of the inter-pattern distance with the speech pattern A, but is used after being restored to the standard pattern B _x1 with the compression information Nx. Therefore, since the time information lost due to the compression of the standard pattern Bx is taken into account, the occurrence rate of matching errors is reduced.

  Next, a second embodiment will be described. The second embodiment is different from the first embodiment only in the pattern matching processing unit 15. The principal part structure of the pattern matching process part 15 in 2nd Embodiment is shown with the block diagram of FIG.

  The pattern matching processing unit 15 includes a recurrence formula setting unit 31 and a distance calculation unit 32. The distance calculation unit 32 calculates the inter-pattern distance Gx between the voice pattern A and the compressed standard pattern B′x by repetitive calculation of the recurrence formula set in the recurrence formula setting unit 31. At this time, the recurrence formula is calculated using the compression information Nx stored in association with the compression standard pattern B′x.

The recurrence formula shown in the following formula (4) is set in the recurrence formula setting unit 31.

In the equation (4), d (i, j) is an inter-element distance between the feature element a _i of the voice pattern A and the feature element b ′ _j of the compressed standard pattern B′x. g (i, j) is an inter-element cumulative distance between the voice pattern A and the compressed standard pattern B′x. n _j−1 is an element (compression ratio) of the compression information Nx.

  The uppermost expression of the recurrence formula (4) is the relationship between an arbitrary point (i, j) on the A and B planes shown in FIG. Is stipulated.

The lowest equation of the recurrence equation (4) is when the compression information element n _j−1 is “1”. In this case, the relationship between the point (i, j-1) and the point (i-1, j-1) adjacent to the lower left is defined.

The middle formula of the recurrence formula (4) is when the compressed information element n _j−1 is larger than “1”. In this case, with respect to the point (i, j), the point (i−1, j−1) adjacent to the lower left diagonally further follows the compression information element n _j−1 corresponding to the feature element b ′ _j−1. It defines the relationship with the point (in _j-1 , j-1) where the restricted path length is extended.

By repeatedly performing the calculation of the recurrence formula (4), the feature element b ′ _j−1 of the compressed standard pattern B′x can correspond to the n _j−1 elements a _i of the input pattern A. Imposed. Therefore, since the time information lost due to the compression of the standard pattern Bx is taken into account, the occurrence rate of matching errors is reduced.

The recurrence formula that provides the same effect is not limited to the above formula (4). For example, as shown in FIG. 7, the recurrence formula of the following formula (5) for extending the restricted path length with respect to the feature element b ′ _j may be set in the recurrence formula setting unit 31.

Alternatively, as shown in FIG. 8, the recurrence formula of the following formula (6) for extending the restricted path length for both the feature elements b ′ _j and b ′ _j−1 is the recurrence formula setting unit 31. May be set.

  In each of the above embodiments, when compressing the standard pattern Bx, an average of a plurality of approximate feature elements is taken. However, one of the approximate feature elements may be selected as a representative. . It is also possible to use a codebook clustering technique.

In the above embodiment, the element (compression ratio) n _j of the compression information nX is used as it is, but the inter-pattern distance may be calculated using a predetermined magnification, for example, a value of 0.8. Good.

  In the above-described embodiment, the case where the present invention is applied to the voice pattern recognition field has been described. However, the present invention can be similarly applied to the pattern recognition field such as characters and figures. Of course, various modifications can be made without departing from the scope of the present invention.

  DESCRIPTION OF SYMBOLS 10 ... Pattern matching apparatus, 11 ... Speech analysis part, 12 ... Pattern compression part, 13 ... Compression information generation part, 14 ... Memory | storage part, 15 ... Pattern matching process part

Claims (5)

In the pattern matching device that calculates the inter-pattern distance between the standard pattern and the input pattern indicated by the time series of each feature element, and outputs the inter-pattern distance as a recognition result,
Pattern compressing means for compressing the standard pattern by integrating, for each feature element constituting the standard pattern, one that approximates an adjacent feature element into one feature element;
Compression information generating means for generating, as compressed information, a series of compression ratios for each feature element of the compressed standard pattern compressed by the pattern compressing means;
Storage means for storing the compression standard pattern compressed by the pattern compression means in association with the compression information generated by the compression information generation means for the compression standard pattern;
The distance between patterns of the compression standard pattern stored by the storage unit and the input pattern is a dynamic programming method having as a variable the compression information generated by the compression information generation unit for the compression standard pattern. A distance calculating means for calculating by a recurrence formula;
A pattern matching apparatus comprising: The recurrence formula used in the distance calculation means is
The inter-element distance between the feature element a _{i of the} input pattern and the feature element b _j of the compressed standard pattern is d (i, j), and the cumulative inter-element distance between the input pattern and the compressed standard pattern is g (i , J), and compression information that is a sequence of the compression ratio n is n _j ,

The pattern matching apparatus according to claim 1, wherein The recurrence formula used in the distance calculation means is
The inter-element distance between the feature element a _{i of the} input pattern and the feature element b _j of the compressed standard pattern is d (i, j), and the cumulative inter-element distance between the input pattern and the compressed standard pattern is g (i , J), and compression information that is a sequence of the compression ratio n is n _j ,

The pattern matching apparatus according to claim 1, wherein The recurrence formula used in the distance calculation means is
The inter-element distance between the feature element a _{i of the} input pattern and the feature element b _j of the compressed standard pattern is d (i, j), and the cumulative inter-element distance between the input pattern and the compressed standard pattern is g (i , J), and compression information that is a sequence of the compression ratio n is n _j ,

The pattern matching apparatus according to claim 1, wherein A pattern matching method in a pattern matching device that calculates a distance between patterns of a standard pattern and an input pattern that are each shown in time series of feature elements, and outputs the distance between the patterns as a recognition result,
For each feature element constituting the standard pattern, a compression step of compressing the standard pattern by integrating one feature element that approximates an adjacent feature element;
A generation step of generating, as compression information, a series of compression ratios for each feature element of the compression standard pattern compressed by the pattern compression unit;
Recursion of dynamic programming having as a variable the inter-pattern distance between the compressed standard pattern compressed in the compression step and the input pattern, with the compression information generated by the generation step for the compressed standard pattern as a variable A calculation step calculated by an equation;
A pattern matching method comprising:

Images