JPH04295893A - Pattern recognition system - Google Patents

Abstract

PURPOSE:To offer the pattern recognition system having high discriminating ability. CONSTITUTION:Voice data is sampled by a sampling means 1, an LPC cepstrum coefficient is calculated by an LPC cepstrum analyzing means 2, vowel data is segmented by a vowel data segmenting means 3, this segmented data is set as learning data, this learning data is recognized by using standard patterns prepared by plural pieces each at every category by a standard pattern learning means 4, the degree of an erroneous recognition at that time is constituted as a function of this standard pattern, an optimal pattern for minimizing this function is derived and stored in a storage means 5, the derived optimal standard pattern and input data are collated by a collating means 6, and based on its result, a category corresponding to the standard pattern being the nearest to the input data is derived as a result of recognition by a deciding means 7.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to a pattern recognition method for recognizing data using standard patterns, and in particular,
This invention relates to a pattern recognition method that identifies data by preparing a plurality of standard patterns for each category.

0002

SUMMARY OF THE INVENTION The present invention relates to a standard pattern construction method for pattern recognition in which data is identified using a plurality of standard patterns prepared for each category. By configuring the degree of misrecognition when recognizing using the standard pattern as a function of the standard pattern, and then finding the optimal standard pattern that minimizes this function, we can use the conventional standard pattern learning method. This is a pattern recognition method that provides higher identification performance than when recognition is performed by configuring standard patterns.

0003

2. Description of the Related Art As a standard pattern construction method for a pattern recognition system in which a plurality of standard patterns are prepared for each category and data is identified, for example, T. Koho
nen“Self Organization a
LVQ (Learning Vecto
r Quantization), and its improved version, LVQ2, is widely used. LVQ2 is
Considering the t-th learning step in a method of repeatedly modifying the standard pattern using learning data, first select one learning data (represented by x), and then select the standard pattern (mi (t ) belongs to a different category from x, and the second closest standard pattern to x (denoted by mj (t)) belongs to the same category as x, then x, mi (t), mj (t ) satisfies the condition that x falls within the "window" determined by mi (t) and mj (t): (1) mi Modify to move away from x;mi
(t+1)=mi(t)−α(t)(x−mi(t)
) (2) Modify mj to bring it closer to x; mj (t
+1)=mj (t)+α(t)(x-mj(t)) (3) Standard patterns other than mi and mj are not modified; mk (t+1)=mk (t), k≠i,
jIn the above formula, α(t) has an initial value of 0≦α(0)≦
It is a monotonically decreasing function of t that satisfies 1 and is set to 0 at the final stage of the iteration. It is known that in LVQ2, if the number of repetitions of such learning and α(t) are appropriately designed, a set of standard patterns can be obtained that reduces the number of false recognitions for the learning data.

0004

[Problems to be Solved by the Invention] It is widely known that LVQ2 is effective in configuring standard patterns, but LVQ2
Q2 has the problem that ``Since the purpose is only to reduce the number of false recognitions for learning data, it is difficult to obtain high discrimination ability for unknown data, especially when the amount of learning data is small.''

[0005] Therefore, the purpose of the present invention is to expand the concept of the number of misrecognitions, define a function that expresses the degree of misrecognition, and then minimize this function with respect to learning data, thereby increasing accuracy even from a small amount of learning data. The object of the present invention is to provide a pattern recognition method that has a higher ability to identify unknown data than conventional methods, within the framework of composing an optimal standard pattern and identifying data using this optimal standard pattern.

[0006]

[Means for Solving the Problems] In order to achieve the above object, the present invention provides a pattern recognition method in which data is recognized using a plurality of standard patterns prepared for each category. The method is characterized in that the degree of misrecognition when recognizing using a pattern is configured as a function of the standard pattern, an optimal standard pattern that minimizes the function is found, and data is recognized using the optimal standard pattern. do.

[0007]

[Operation] According to the present invention, the degree of misrecognition when the entire given learning data is recognized using a standard pattern is constructed as a function of the standard pattern, and an optimal standard pattern that minimizes this function is determined. By performing data recognition using this optimal standard pattern, high identification performance can be obtained.

[0008]

EXAMPLES The present invention will be explained in detail below based on examples.

First, the notation is defined as follows.

Rp: p-dimensional Euclidean space K: number of categories of patterns to be recognized Ni: number of learning data for category i (i=1,
..., K) Mi: number of standard patterns for category i (i=1
,...,K) xij: learning data (xij ∈Rp) (i=1,
..., K, j=1, ..., Ni) tir: r-th standard pattern for category i (tir ∈Rp) (i=1, ..., K, r=1, ..., Mi) E: set of standard patterns Functions L, m that express the degree of misrecognition when all the learning data {xIJ} are identified using {tir}: Number of repetitions The present invention creates a set of standard patterns {t
ir}. One of the methods for solving combinatorial optimization problems as a method to minimize E
A simulated annealing method
Flowcharts of an embodiment of the present invention in which the processing method (nealing) is used are shown in FIGS. 1 to 3. The execution contents of each block shown in the figure are as follows.

B2: Using methods such as clustering,
After creating a predetermined number of standard patterns for each category from the learning data set {xij}, the resulting set of standard patterns {tir} is output. In this example, an initial standard pattern was created using the LBG clustering algorithm.

B4: Using the set {tir} as a standard pattern, calculate a function E representing the degree of misrecognition when the learning data set {xij} is recognized. Although E may be defined by the number of pieces of learning data that are erroneously recognized, in this example, E was calculated as follows. First, some learning data xi
The following two distances d1 and d2 are defined for j.

[0013]

[Math 1]

[0014]

[Equation 2] Here, d represents the Euclidean distance between two patterns. At this time, λ

[0015]

Defined by [Equation 3]. Obviously, 0≦λ≦1. For this λ and the parameter κ (0<κ<0.5), the function μ
Define (xij, {tir}). Specifically, λ<
If 0.5-κ, μ=0, 0.5-κ≦λ<0.5
μ=(λ+κ−0.5)/2κ for +κ. Let μ=1 for 0.5+κ≦λ. In the case of identification using a standard pattern, if λ<0.5, it is correctly identified, and if λ≧0.5, it is incorrectly identified. Therefore, λ≧0
．． A function μ that is 1 for 5 and 0 for other λ
′, μ′ expresses whether or not it is a misrecognition using 0 and 1. μ in the above equation is obtained by introducing the parameter κ into μ′, and is a function representing the degree of misrecognition. function E

[0016]

Defined by [Equation 4]. Obviously E is a function of the set of standard patterns {tir}, and as standard patterns {ti
r} represents the degree of misrecognition when the entire learning data is identified. In this example, κ was set to 0.1.

B6: Assign 0 to variable n.

B8: Assign 0 to variable i.

B10: Compare the two degrees of misrecognition E and E*.

B12: A block executed when E<E* in B10, which is a set of standard patterns {tir}
Let all elements of set {tir*} be elements, and let the value of E be E
Perform the operation of assigning to *.

B14: "Temperature" Tn, which is a block to be executed when E<E* is not satisfied in B10 and after B12, and is one of the parameters in the annealing method.
Set.

B16: Select one element xij from the learning data set {xij} by some method. In this example, one learning data xij is randomly selected from the set {xij}.

B18: Select one element tur from the standard pattern set {tij} by some method. In this example, one standard pattern tur is randomly selected from the set {tij}.

B20: Perturb the standard pattern set {tir} and output a new standard pattern set {tir'}. The perturbation is given by the following equation.

[0025] tur'=tur+ν・θ・(xij−tur), ti
k'=tik; i≠μ, k≠r. Here, θ is a uniform random number on the interval [-1.0, 1.0], and ν is a parameter satisfying 0≦ν≦1, which is 0.0 in this example.
It was set to 2.

B4: In the same manner as described above, a function E' representing the degree of misrecognition when the learning data set {xij} is recognized is calculated using the set {tir'} as a standard pattern.

B22: Generate a random number ω uniformly distributed between the interval [0, 1).

B24: E'-E is negative or ω generated in B22 is exp(-(E'-E)/Tn
).

B26: This block is executed when the determination in B24 is Yes, and is a block that is executed when the determination in B24 is Yes, and is a block that executes when the determination in B24 is Yes.
'} are set as elements of the set {tir}, and the value of E' is assigned to E.

B28: When the determination in B24 is No, this block is executed next after B26, and the value of i is set to 1.
increase only.

B30: Determine whether i is greater than a predetermined value m. If the determination is No, proceed to B10.

B32: This block is executed when the determination in B30 is Yes, and increases the value of n by 1.

B34: Determine whether n is larger than a preset value L. If the determination is No, proceed to B8.

B36: This block is executed when the determination in B34 is Yes, and is the same as B10.

B12: Same as B12 above B38: Set {
tir*} is the final set of standard patterns.

In the embodiment shown above, the case where the annealing method is used has been described, but for example, if the judgment condition for B24 is set only to E'-E<0, this corresponds to the case where the steepest descent method is used. Therefore, it goes without saying that the present invention is applicable not only to cases where the annealing method is used, but also to cases where other optimization methods are used.

As mentioned above, standard patterns with high recognition accuracy can be constructed from a small amount of training data, so if the problem is pattern recognition using standard patterns, this method can be used to improve recognition ability. Able to adapt.

Next, we will discuss the case where the present invention is applied to a device that recognizes vowels using standard patterns in speech recognition. FIG. 4 is a block diagram showing the functional configuration of a pattern recognition device that implements the present invention. Details are as follows.

1: Sampling means for sampling input voice information.

2: LPC cepstral coefficients are calculated for each short time interval based on sampled audio data.
PC cepstrum analysis tool.

3: Vowel data extraction means for extracting LPC cepstral coefficients at the center of vowels from the voice data from means 2.

4: A standard pattern learning means that uses the data from the means 3 as learning data and constitutes a standard pattern, which stores a CPU and a program for causing the CPU to execute each block shown in B2 to B38 above. ROM
and a RAM having a work area.

5: Storage means storing the standard pattern obtained by means 4.

6: Set of standard patterns of means 5 and means 3
Collation means that collates the input speech vowel data from and calculates the distance between the input data and the standard pattern.

7: Determining means for determining the vowel category corresponding to the standard pattern closest to the input based on the calculation result from the means 6, and outputting the result as a recognition result.

The results of vowel recognition using the above device will be described. 6 vowels /a/, /i including pellicles as vowels
/, /u/, /e/, /o/, /N/ were assumed, and a case where a plurality of standard patterns were prepared for each of these six categories and vowel recognition was performed was studied. However, the number of standard patterns for each category was the same. In the study, the speech data uttered by the speaker MAU in the ATR database in units of phrases was sampled at a sampling frequency of 15 kHz (Means 1), and after performing 18th-order LPC cepstral analysis (Means 2), Vowel data was created by averaging three frames centered on the vowel. The vowel center was determined based on the vowel center label given in the same database. Create 350 pieces of data for each category, divide them into 50 pieces (7 parts), and create Dataset A.
~Created G. Further, data sets 1 and 2 were created by dividing each category into 175 pieces (divided into 2) (means 3).

As a case in which the number of learning data is small, experiments were conducted by means 4, 5, 6, and 7 using data sets A to G. First, a standard pattern was constructed using data set A, and data other than data set A (unknown data) was recognized. Furthermore, similar recognition was made when learning was performed using data other than data set A. These seven results were averaged to obtain the final recognition result. In addition, an experiment similar to the above was conducted using Data Sets 1 and 2 in a case where there was a relatively large amount of learning data. In addition, when using the present invention, the number of repetitions was set as L=5000 and m=10.

A similar recognition experiment was also conducted using the LVQ2 algorithm, which is a typical conventional standard pattern learning algorithm. First, 10 patterns were created by randomly rearranging all the learning data, learning data were selected in order from among these, and the standard pattern was corrected using LVQ2. In that case, α(0)=0.015
2 in the window that determines the conditions for standard pattern correction.
A 0% window was used. These are LVQ2 proposer T. Koh
This is the value recommended by onen. The initial standard pattern was given by the LBG clustering algorithm as in the present invention.

The results of the experiment are shown in tables in FIGS. 5 and 6. Figure 5
Fig. 6 shows a case where the number of learning data is relatively large, and Fig. 6 shows a case where the number of learning data is small. The vertical axes in FIGS. 5 and 6 show the false recognition rate, and the horizontal axes show the number of standard patterns for each category. For reference, FIGS. 5 and 6 also show the results obtained when the standard pattern obtained by the LBG algorithm (that is, the initial standard pattern of the present invention and LVQ2) was used. As shown in FIG. 5, when the number of learning data is relatively large, both the present invention and LVQ2 use LBG
The false recognition rate is lower than that of the algorithm, indicating that learning is performed well. Moreover, the present invention has better results. When the number of learning data is small (FIG. 6), the results of LVQ2 are not much different from the LBG algorithm, and therefore the effect of LVQ2 learning is hardly visible. On the other hand, when the present invention was used, generally good results were obtained.

These experiments have shown that the present invention is a method that improves the drawback of the conventional method that learning effects are not obtained when the number of learning data is small, and moreover, even when the number of training data is increased to a certain extent, It was confirmed that this method is superior to conventional methods.

[Brief explanation of drawings]

FIG. 1 is a part of a flowchart showing the operation of an embodiment of the present invention.

FIG. 2 is another part of the same flowchart.

FIG. 3 is the remaining part of the same flowchart.

FIG. 4 is a block diagram showing the functional configuration of a pattern recognition device according to an embodiment of the present invention.

FIG. 5 is a diagram showing an example of experimental results in an example of the present invention.

FIG. 6 is a diagram showing another example of the same experimental results.

[Explanation of symbols]

1 Sampling means 2 LPC cepstral analysis means 3 Vowel data extraction means 4 Standard pattern learning means 5. Storage means 6. Verification means 7 Judgment means

Claims (1)

[Claims] Claim 1: In a pattern recognition method that recognizes data using a plurality of standard patterns prepared for each category, the degree of misrecognition when given training data is recognized using standard patterns. A pattern recognition method characterized in that the optimal standard pattern is configured as a function of the standard pattern, the optimal standard pattern that minimizes the function is found, and data is recognized using the optimal standard pattern.