JPH04354049A - Learning system for pattern recognition - Google Patents

Abstract

PURPOSE:To improve an identification rate at the time of classifying an unknown set by a standard vector obtained only by a known set by obtaining the sum of the limiting elements of a similarity as the evaluation factor of the standard vector, and controlling the learning factor based on an evaluated result. CONSTITUTION:An identifying processing 1 is operated to a set A in order to search the identification rate. Whether or not the identification rate is more satisfactory than the previous time is searched by a standard vector evaluating processing 2 by using that identification rate as the evaluation factor, and when it is converged, the similarity is searched. The value of the above mentioned evaluation factor is searched based on this similarity, and the re-evaluation value of the standard vector is searched by a standard vector re-evaluating processing 3. Then, when a difference between this re-evaluation value and the previous evaluation value is less than a threshold, the processing is completed, and when it is not less than the threshold, the resetting of the learning factor is operated by a learning factor resetting processing 4. Then, when the evaluated result of the above mentioned standard vector is more satisfactory than the previous evaluated result, the first learning of the standard vector is operated based on the set learning factor, and the second learning of the above mentioned standard vector is operated based on the reset factor.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to pattern recognition, and more particularly to a method for learning standard patterns of classification for a set of patterns for which classification patterns are known.

0002

2. Description of the Related Art Conventionally, this type of learning method consists of a classification process 11, a standard vector evaluation process 12, and a learning process 13, as shown in FIG. On the other hand, the identification rate Rc is mainly used as the evaluation function when calculating the number shown in Equation 8 below, and R
The configuration was such that learning such that c+1 was equal to or greater than Rc was effective learning.

0003

[Math. 8]

0004

[Problem to be solved by the invention] Since the standard pattern according to the conventional method is configured to use only the recognition rate Rc as an evaluation function, it is difficult to obtain the standard pattern in the interval τ where learning is performed. Maximize the identification rate RT obtained when classifying the unknown set T using the optimal standard pattern for the unknown set different from the set used in There is a drawback that it is not possible to obtain a standard pattern that can be used.

[0005]

[Math. 9]

[0006]

[Means for Solving the Problems] The learning method of the present invention uses the classification rate and the similarity rpq (p is the serial number of the vector, q is the set number) in each set as a function for evaluating the learning results.
The similarity rpt (t is the set number to which the vector belongs) in adjacent elements of the value of and the integrated value of the ratio p of rpt', and when the evaluation result is worse than the previous value, A pattern recognition learning method having a feedback loop for changing the learning coefficient C(R) is obtained.

Further, according to the present invention, as a function for evaluating learning results, the classification rate and the similarity of the values of similarity rpq (p is the serial number of the vector, q is the set number) in adjacent elements in each set are evaluated. degrees rpt (t is the set number to which the vector belongs)
and the integrated value for p of the ratio of rpt', and the identification R
, a process to evaluate the standard vector using the discrimination rate R, a process to re-evaluate the standard vector based on the similarity, and a process to perform the re-evaluation when the re-evaluation result is worse than the previous value. , a process of resetting the learning coefficient C(R), and a learning process of performing first learning using the learning coefficient in standard vector evaluation processing and second learning using the reset learning coefficient. A learning method for recognition is obtained.

Furthermore, according to the present invention, there is a set 1 of similar patterns having finite elements, a set 2 having a different similarity from the set 1, and sets 3, 4, etc. that are classified similarly.
In the set A that includes all of these sets 1, 2, etc., the vector representing the characteristics of each element is expressed as in Formula 1, and the standard vector of the sets 1, 2, etc. is expressed as in Equation 2, and the set k of vectors in Equation 1 (k=1,
2,...) is expressed as Equation 3, the set of classified vectors expressed in Equation 1 is a pattern that makes Equation 3 the maximum or minimum value of k. -In the recognition method of In order to improve [%], Formula 5
In a learning method where Equation 2 is newly computed as A learning method for pattern recognition is obtained, which is characterized by having a feedback loop and a process for resetting learning coefficients in the feedback loop.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a diagram sequentially showing the processing of an embodiment of the present invention.

[0011] In process 1, a classification process is performed on the set A, and the calculation according to Equation 3 is performed to obtain the classification rate R [%].

In process 2, the standard vector shown in equation 1 is evaluated using the identification rate R [%] obtained in process 1 as the evaluation function shown in equation 7, and it is determined whether the identification rate R is better than the previous time. If we calculate the degree of similarity and converge, we calculate the similarity of Equation 3.

In process 3, the value of the evaluation function shown by formula 7 is determined based on the similarity of formula 3 obtained in process 2, and the re-evaluation value of the standard spectrum of formula 1 is determined.

In calculating this evaluation function, if the element of the mother set is ai and the degree of similarity to it is ri, then the evaluation function will be in the form of integrating ri with respect to i. However, it is assumed that ri is equal to or larger than an appropriate number C.

[0015] If the difference between this re-evaluation value and the previous evaluation value is smaller than a preset threshold, that is, if they are converged, it is assumed that sufficient learning has been achieved and the entire process is terminated.

In process 4, if the difference between the evaluation values in process 3 is not smaller than the threshold, that is, if convergence does not occur, the learning coefficient is reset, and when convergence occurs, what is done? Neither.

FIG. 3 is a diagram showing the relationship between the ratio of the learning coefficient to the initial value and the recognition rate, and the coefficient is controlled according to this curve.

In process 5, if the evaluation result of the standard spectrum in process 2 is better than the previous evaluation result,
Based on the set learning coefficient, the first
A second learning is performed on the standard vector based on the coefficients reset in process 5, and the process returns to process 1.

FIG. 4 is an image diagram showing a comparison of the degree of separation in the first learning and the degree of separation in the second learning.
It can be seen that the effect of learning becomes greater.

[0020]

[Effects of the Invention] As explained above, the present invention uses a sum of similarities for limited elements as an evaluation function of standard vectors, and a method of controlling learning coefficients based on the evaluation results. This has the effect of improving the identification rate when classifying unknown sets using standard vectors obtained only from known sets.

[Brief explanation of drawings]

FIG. 1 is a diagram showing a processing flow of a pattern recognition learning method according to the present invention.

FIG. 2 is a diagram showing a processing flow of a conventional learning method.

FIG. 3 is a diagram showing the relationship between coefficients for resetting learning coefficients and recognition rate in the present invention.

FIG. 4 is a diagram showing the difference between the degree of separation obtained by first learning and the degree of separation obtained by second learning in the present invention.

[Explanation of symbols]

1 Classification processing 2 Standard vector evaluation processing 3 Standard vector re-evaluation processing 4 Learning coefficient resetting process 5 Learning processing 6. Classification aspect

Claims (3)

[Claims] Claim 1: As a function for evaluating the learning results, the classification rate and the similarity rp of the values of similarity rpq (p is the serial number of the vector, q is the set number) of adjacent elements in each set are calculated.
t (t is the set number to which the vector belongs) and the integrated value for p of the ratio of rpt', and changes the learning coefficient C(R) when the evaluation result is worse than the previous value. −
A learning method for pattern recognition that has a double loop. [Claim 2] As a function for evaluating learning results, the classification rate and the similarity rp of the values of similarity rpq (p is the serial number of the vector, q is the set number) of adjacent elements in each set are calculated.
t (t is the set number to which the vector belongs) and the integrated value for p of the ratio of rpt', a process of calculating the identification R, and a process of evaluating the standard vector using the identification rate R. and the process of re-evaluating the standard vector based on the similarity.
When the re-evaluation result is worse than the previous value, the learning coefficient C (
R) A learning method for pattern recognition which includes a process of resetting , a first learning process using a learning coefficient in standard vector evaluation process, and a second learning process using the reset learning coefficient. Claim 3: A set 1 of similar patterns having finite elements, a set 2 having a different similarity from this set 1, sets 3, 4, etc. classified similarly, and these In the set A that includes all sets 1, 2, ..., the vector representing the characteristics of each element is expressed as in the following formula 1,
The standard vectors of sets 1, 2, etc. are expressed as in Equation 2 below, and the set k of vectors in Equation 1 (k=1, 2,
) is calculated using the following formula 3.
Then, in a pattern recognition method where the classified set of vectors expressed by Formula 1 is the value of k whose maximum or minimum is Expression 3, the following Expression 4 is used. For the set α to which the set to which the vector belongs is known, in order to improve the identification rate R [%] which represents the proportion of correctly identified within the set α classified by the result of the above calculation, the following formula 5 is newly created. Perform calculations on formula 2,
In a learning method that obtains the following Equation 6, the evaluation function of Equation 5 is the recognition rate R and the following Equation 7, and a feedback loop to the learning process based on the evaluation result of the second evaluation function, and this A learning method for pattern recognition characterized by comprising a process of resetting learning coefficients in a feedback loop. [Math. 1] [Math. 2] [Math. 3] [Math. 4] [Math. 5] [Math. 6] [Math. 7]