JPH05197811A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To reduce the capacity of a memory which stores the parameters by cutting off the redundancy of the feature value. CONSTITUTION:A main component axis calculating device 1 calculates an average vector and a unit vector to be set in the main component axial direction from the feature value vector of the sample data and stored these obtained vectors in a memory 2. The feature value converting devices 3a and 3b convert the feature value vectors of the sample data and the identifying object data into the compressed feature value vectors respectively. A template producing device 4 produces the template and template auxiliary data from the output of the device 3a and stores these data in the memory 2. A data selecting device 5 decides the normalcy the abnormality of the identifying object data based on the output of the device 3b and the data stored in the memory 2. Then a matching degree calculating device 6 and a class deciding device 7 output the identifying results based on the feature value vector converted from the normal data and the data stored in the memory 2.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition device applied to an image and signal identification device.

[0002]

2. Description of the Related Art In a conventional pattern recognition apparatus, as parameters used for recognition, an average vector (template) for each class and eigenvalues and eigenvectors (template auxiliary data) of a covariance matrix regarding the average are calculated. , Was stored in the memory provided in the device.

[0003]

However, in the above-mentioned conventional method, for example, when the feature quantity is f, the number of classes is s, and the maximum number of corrections is t, the number of each parameter is the average vector f * s eigenvector f *. s * t eigenvalue s * t

[0004] Therefore, in the memory of the pattern recognition device, the total number of parameters (f * s) is stored.
A capacity for holding + f * s * t + s * t) parameters is required. The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a pattern recognition device capable of reducing the capacity of a memory for storing necessary parameters.

[0005]

The pattern recognition apparatus according to the present invention is provided with a sufficient number of sample data whose feature quantities and correct classes are known in a data group represented by a plurality of feature quantity vectors. At this time, a storage means for storing data generated in processing and an average vector for the plurality of sample data are calculated, a unit vector in the axial direction of the principal component is obtained using principal component analysis, and a unit vector is determined in advance from the unit vector. Number of unit vectors in the principal component axis direction are selected, and the unit data in the principal component axis direction selected by the principal component axis calculation unit are used to extract the sample data and the identification target data. A feature amount conversion device for converting a feature amount vector, a feature amount vector of sample data obtained by the feature amount conversion device, and a correctness of the sample data A template creating device that creates a template and template auxiliary data for each class from a class, an average vector of the sample data obtained by the principal component axis calculation device, a unit vector in the principal component axis direction, and the feature amount conversion device. A data selection device that determines whether the identification target data is abnormal data or normal data from the converted feature amount vector of the identification target data and the feature amount vector of the identification target data before conversion, and a normal state by this data selection device. A matching degree calculation device that obtains a matching degree with each class using the template and template auxiliary data for the identification target data determined to be data, and the identification target data from the matching degree obtained by this matching degree calculation device Class judgment to determine the class to which Characterized by comprising a device.

[0006]

The feature amount includes a feature amount which is always a constant value such as a pixel value at a corner of a binary image, and a redundant feature amount such as a feature amount having a correlation such as vehicle height and vehicle weight. Has been. If the feature quantity to be handled can be compressed by discarding such redundancy of the feature quantity, the number of parameters used for recognition will also decrease. At this time, when the number of parameters increased for feature amount compression is smaller than the number of decreased parameters used for recognition, the capacity of the memory holding the parameters required in the pattern recognition apparatus can be reduced.

When a sufficient number of sample data whose feature quantities and correct classes are known in a data group represented by a plurality of feature quantity vectors are given, the principal component axis computing device calculates an average vector ( (Used by the data selection device) is calculated, and the unit vector in the principal component axis direction, which becomes the feature amount conversion parameter, is obtained by the following procedure.

First, principal component analysis is performed on the entire sample data to find a unit vector in the principal component axis direction. Next, in the principal component analysis on the data represented by the f-dimensional feature amount vector, f ordered principal component axes are obtained, but the feature amount space is selected by selecting the upper k (k <f) of them. Performs compression of the number of dimensions. The unit vectors of the top k principal component axis directions are used as the feature amount conversion parameters.

In the feature quantity conversion device, the f-dimensional feature quantity vector of each sample data is converted using the conversion parameter (unit vector in the principal component axis direction) obtained by the above principal component axis calculation device, and a new A k-dimensional feature vector is obtained. The new feature amount vector is a value obtained by projecting the feature amount vector before conversion in the principal component axis direction. The conversion formula is

[0010]

[Equation 1] Becomes

The template creating device calculates the template and the template auxiliary data from the new feature amount vector obtained by the feature amount converting device. Assuming that the number of classes is s and the maximum number of corrections is t, the new feature amount vector is k-dimensional, so the parameters are the numbers shown below. Average vector k * s eigenvector k * s * t eigenvalue s * t The parameters that need to be retained are: Feature conversion parameters (unit vector in the principal component axis direction) f * k All sample data before feature conversion Mean vector of
Since f 1 and the parameters obtained by the template creating apparatus are combined, a memory capacity for storing the parameters of k * s + k * s * t + s * t + f * k + f is also required. Here, the difference from the number of parameters that needs to be held in the conventional method is shown below.

[0012]

[Equation 2]

From the above comparison, (f−k) * s * (t +
If 1) is larger than f * (k + 1), memory can be saved. When the compression rate k / f is good for data with a large number of feature vector dimensions f, and the number of classes is s even with the same compression rate
The higher the number, the higher the saving effect.

When the feature amount vector of the identification target data is given, the feature amount vector of the identification target data is converted by the feature amount conversion device using the conversion parameter stored in the memory in advance, and a new k Obtain the dimensional feature vector.

However, if the identification target data has a large component in the truncated principal component axis direction, that is, data that is out of the distribution of the sample data, that component disappears in the converted feature vector, It may lead to misidentification. In order to prevent this erroneous identification in advance, the data selection device detects data that may be erroneously identified as abnormal data by the following procedure.

First, the feature vector before conversion of the identification target data, the average vector of the sample data obtained by the principal component axis calculating device and the unit vector in the principal component axis direction, and the new vector obtained by the feature converting device. The component D shown in the following equation is calculated using the k-dimensional feature vector. This D is the size of the feature amount component that disappears due to the conversion. Therefore, when D is equal to or larger than a predetermined threshold value, the data is regarded as abnormal data.

[0017]

[Equation 3]

When the identification target data is not abnormal in the above method, the matching degree calculation device calculates the matching degree with each class. Then, the class determination device determines that the class having the highest matching degree obtained by the matching degree calculation device is the class to which the identification target data belongs.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

The structure of the pattern recognition apparatus according to the present invention is shown in FIG. The principal component axis calculation device 1 calculates an average vector of all sample data when a sufficient number of sample data whose feature value is known is given in a data group represented by a plurality of feature vectors. Further, a unit vector in the principal component axis direction is calculated, the higher one is selected, and these are stored in the memory 2. Feature amount conversion device 3a
Uses the unit vector in the principal component axis direction stored in the memory 2 to convert the feature quantity vector of all sample data to generate a compressed feature quantity vector, and to generate the new feature quantity vector and the sample The correct answer class of the data is sent to the template creating device 4. The template creation device 4 creates templates of each class which are the basis of the matching degree calculation and template auxiliary data necessary for the matching degree calculation process, and stores these in the memory 2.

When the feature quantity vector of the identification target data is given, the feature quantity transforming device 3b having the same structure as the feature quantity transforming device 3a uses the unit vector in the principal component axis direction stored in the memory 2. The feature amount vector of the identification target data is converted by using the data, and the compressed feature amount vector is sent to the data selection device 5. The data selection device 5 calculates the size of the feature amount component that disappears due to the feature amount conversion of the identification target data, and if the value is larger than a predetermined value, outputs the identification target data as abnormal data. .. The data selection device 5 regards normal data other than abnormal data, and when the data is normal data, the feature quantity conversion device 3b.
The compressed feature quantity vector of the identification target data sent from the above is output to the matching degree calculation device 6.

The matching degree calculation device 6 includes the memory 2
Using the template and template auxiliary data stored in, the identification target data determined to be normal data by the data selection device 5 is calculated for the degree of matching with each class and output to the class determination device 7. The class determination device 7 determines that the class having the highest matching degree calculated by the matching degree calculation device 6 is the class to which the identification target data belongs. Next, the operation of the embodiment will be described.

When a sufficient number of sample data whose feature value is known is given to a data group represented by a plurality of feature amount vectors, the principal component axis calculation device 1 calculates the average vector of all sample data. Then, the unit vector in the principal component axis direction is calculated, and the higher one is selected. The average vector thus generated and the unit vector in the principal component axis direction are sent to the memory 2 and stored therein.

Further, the feature quantity vector of the sample data is converted into a new vector, that is, a compressed feature quantity vector in the feature quantity conversion device 3a. At this time, the feature quantity conversion device 3a applies the unit vector in the principal component axis direction stored in the memory 2. The compressed feature vector is sent to the template creating device 4 together with the correct answer class code. As a result, the template creation device 4 generates the template of each class which is the basis of the matching degree calculation and the template auxiliary data necessary for the process of the matching degree calculation. These templates and template auxiliary data are stored in the memory 2.

When the feature quantity vector of the identification target data is given, the feature quantity vector of the identification target data is converted into the feature quantity vector compressed by the feature quantity conversion device 3b. At this time, the unit vector in the principal component axis direction stored in the memory 2 is applied. But,
If the identification target data has a large component in the principal component axis direction that is truncated as shown in FIG. 2, that is, data that is out of the distribution of the sample data, that component disappears in the converted feature amount vector. This may result in misidentification.

Therefore, the generated compressed feature amount vector of the identification object data is sent to the data selecting device 5 and the size of the feature amount component of the identification object data which disappears due to the conversion is calculated. .. When the calculated value is larger than the predetermined value, the identification target data is output from the data selection device 5 as abnormal data. If the calculated value is smaller than a predetermined value, the identification target data is determined to be normal data,
The compressed feature quantity vector is output to the matching degree calculation device 6.

The matching degree calculation device 6 calculates the degree of matching with each class for the identification target data which is judged to be normal data. Class determination device 7 from this calculation result
Determines that the class with the highest degree of matching is the class to which the identification target data belongs and outputs it as the identification result.

Next, a method of determining the number of principal component axes selected in the principal component axis calculating apparatus 1 will be specifically described by using an example using data represented by 128-dimensional feature vector.

There are about 1000 pieces of sample data whose correct answer classes are known in advance. After changing the number k of principal component axes selected by the principal component axis calculation device 1 from the top 10 to 128 and creating parameters (template, template auxiliary data) used for recognition from the sample data, the sample data Identification is performed by using the. Since the correct answer class is known in advance from the sample data, the discrimination rate is calculated from the discrimination result, and k is determined by looking at the discrimination rate.

FIG. 3 is a graph showing the relationship between k and the discrimination rate (vertical axis:
The identification rate is shown by the horizontal axis: k). In FIG. 3, k ≧ 30
When, the identification rate is high and stable. Therefore, in this example, k = 30 is set. Since the number of classes s = 43 and the maximum number of corrections t = 10, the number of parameters stored in this example is as follows. (Number of dimension of feature quantity before conversion f = 128) k * s + k * s * t + s * t + f * k + f = 18588 When the data of this example is recognized by the conventional method, the number of parameters is as follows. f * s + f * s * t + s * t = 60974 The reduction number of parameters used for recognition is (f−k) * s * (t + 1) = 46354, and the number of parameters increased for compression is f * (k + 1) = 3968. Yes, the number of parameters used for recognition decreases> the number of parameters increases for compression. Therefore, in this example, the memory capacity is about 30% (18588) of the conventional one.
/ 60974).

[0031]

As described above in detail, according to the present invention,
In any template matching, the feature of data can be described with a smaller amount of feature by cutting off the redundancy of the feature amount by using the principal component analysis. Therefore, the capacity of the memory for storing the necessary parameters can be reduced. Further, the data selection device can prevent erroneous identification due to the redundancy of the feature quantity being discarded.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a pattern recognition device according to an embodiment of the present invention.

FIG. 2 is a graph showing an example of data determined to be abnormal data by the data identification apparatus according to the embodiment.

FIG. 3 is a graph showing the relationship between the number of selected principal component axes and the identification rate in the example.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Principal component axis calculation device, 2 ... Memory, 3a, 3b ... Feature amount conversion device, 4 ... Template creation device, 5 ... Data selection device, 6 ... Matching degree calculation device, 7 ... Class determination device.

Claims (1)

[Claims] 1. When a sufficient number of sample data whose feature quantity and correct class are known in a data group represented by a plurality of feature quantity vectors are given, a storage means for storing data generated in processing. , Calculate the average vector for the above multiple sample data, find the unit vector in the principal component axis direction by using the principal component analysis, and select a predetermined number of unit vectors in the principal component axis direction from them. A principal component axis calculation device, and a feature amount conversion device that transforms the feature amount vectors of the sample data and the identification target data using the unit vector in the principal component axis direction selected by the principal component axis calculation device, From the feature quantity vector of the sample data obtained by the feature quantity conversion device and the correct class of the sample data, the template and template of each class A template creating device for creating auxiliary data, an average vector of the sample data obtained by the principal component axis calculating device, a unit vector in the principal component axis direction, and the identification target data converted by the feature amount converting device. A data sorting device that determines whether the identification target data is abnormal data or normal data from the feature amount vector of the above and the feature amount vector before conversion of the identification target data, and an identification target that is determined to be normal data by the data sorting device. A matching degree calculation device that obtains the degree of matching with each class using the template and template auxiliary data for the data, and a class that determines the class to which the identification target data belongs from the matching degree obtained by this matching degree calculation device Pattern recognition characterized by comprising a judgment device apparatus.