JPH01102688A - Pattern recognition system - Google Patents

Abstract

PURPOSE:To perform high-accuracy pattern recognition by performing effective feature extraction for discrimination in consideration of differences of local feature patterns or character patterns in respective layers among categories. CONSTITUTION:This system consists of a feature point selection part 1, a linear integration arithmetic part 2, a deviation corrective arithmetic part 3, a similarity arithmetic part 4, a local feature vector storage part 5, a decision means 6 which decides a maximum value, a pattern memory 10 as a pattern storage means for storing character patterns, and a control part 70. Then feature point selection and decision analysis in consideration of differences in feature of local feature patterns or character patterns among categories is performed to calculate similarity from linear integration, thereby extracting effective features for the discrimination. Consequently, high-accuracy pattern decision making is performed.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a pattern recognition method that provides features effective for identification in fields such as character recognition and speech recognition.

(Prior art) As a method for hierarchically extracting and determining features,
- Divide the unknown pattern described in Publication No. 178578 into multiple partial areas, and calculate the similarity for each using multiple standard partial patterns and simple similarity or composite similarity. There is a method for obtaining a Kabataan similarity matrix, and a well-known method is a berseptron, in which a judgment is made for each pattern by sequentially changing the weighting coefficients of connections from the local area of the pattern.

(Problems to be Solved by the Invention) In the conventional techniques as described above, since there is a correlation between each local feature pattern or each character pattern category,
Although simple similarity or composite similarity is less affected by noise due to local matching, it is difficult to say that it is a small number of features that are effective for identification.

Further, with such conventional techniques, there is a problem in that differences between categories of each local feature pattern or each character pattern are difficult to show in a way that is effective for identification.

The present invention takes this problem into consideration and provides a pattern recognition technique that extracts and determines features effective for identification.

(Means for Solving the Problems) According to the present invention, a pattern storage means for storing unknown patterns, a feature point position storage unit for storing feature point selection positions of local feature patterns in each of a plurality of layers, A feature vector array storage unit that stores the resulting similarity, a feature point selection unit that selects feature points according to the feature point selection position in the local area pattern, and a feature vector array of each layer using a discriminant analysis method. A feature extraction unit consisting of a linear integration calculation unit, a deviation correction calculation unit, and a similarity calculation unit for generating a predetermined similarity is compared with the similarity output from the similarity calculation unit of the top layer feature extraction unit. The output of the pattern storage means is used as a feature vector array in the lowest layer, and in each layer,
Using a discriminant analysis method, the similarity between the feature vectors in which feature points are selected according to the feature point selection positions in the feature vector arrays of all lower layers and the feature vectors of predetermined local feature patterns in each layer is calculated between the feature vector arrays and the feature vectors of the predetermined local feature patterns in each layer. A pattern recognition method is obtained which is characterized by the following.

(Operation) The present invention selects feature points taking into account feature differences between local feature patterns or character pattern categories, and calculates similarity from linear integration using discriminant analysis. This makes it possible to extract features that are effective for identification and ultimately make highly accurate judgments.

The present invention is a pattern recognition method that extracts and judges features hierarchically from the local area of the pattern to the entire pattern while performing intermediate identification using feature point selection and discriminant analysis in local regions. Features are extracted by selecting feature points in a local region according to a variance ratio that takes into account intra-category variance and inter-category variance, and calculating similarity from linear integration using discriminant analysis.

FIG. 2 is an example of extracting and determining features hierarchically while selecting feature points from a local region according to the present invention. The unknown pattern 100 is a characteristic pattern 2 corresponding to multiple edges.
00,201. . . . performs identification processing and expresses the result as a degree of similarity of 300.301. Based on the similarity, feature patterns 400, 401 . . . . Performs identification processing and calculates the results with similarity of 500, 501. ...
Expressed as

Similar processing is repeated, and in the final layer, the similarity is determined to be 2100°, 2101°, etc. by identifying feature patterns 2000, 2001°, etc. corresponding to the character category.
・The character category with the highest degree of similarity compared to the degree of similarity in the final layer is the recognition result. For example, when calculating similarities 210'0, 2101, etc. in the diagram,
5000 feature point selection positions that are not determined in advance by the variance ratio
．． 5001. This means performing identification processing using the features of... and expressing it in terms of similarity.

The processing in each layer is to create a feature vector group X from the selection of feature points in the local region of the lower layer, give a linear exchange matrix A to X,
It consists of processing to perform deviation correction and measure similarity with each local feature vector, and this A can be obtained from the data set (φC) of all local feature categories of the layer. The data set (φC) can include, for example, patterns in which each local feature pattern is shifted in a plurality of directions, or in which the line width is changed if it includes line segments. The local feature vector is a standard vector for each local feature category on the feature axis newly transformed by the linear transformation matrix A.

The process for forming the Sth layer is as follows. However, in the following formula, the subscript S is omitted. 1st,・
..., point (i
, j), feature points are selected based on the variance ratio so that the variance is small among the plurality of feature points within the category to which a certain local area pattern belongs, and the variance between them is large between them and other categories. The feature vector XIJ obtained by feature point selection is linearly integrated using the discriminant analysis method using equation (1),
Find the newly transformed feature vector.

'J+j= A −X+j (
11 On the other hand, the deviation correction vector e obtained in advance from the set of local feature vectors in the S-th layer is used to perform deviation correction according to Equation 21.

? l+j=3/IJ e
(21) From the feature vector hlJ subjected to deviation correction, each local feature vector (gk I k = 1..., M similarity with Tl is calculated by equation (3) as an example. Here, the local feature The vector has been subjected to deviation correction using the deviation correction vector e.

Therefore, the feature vector group in the S-th layer is expressed by equation (4).

F+ ・,・=(ftjt +fth 2 +flJ3
1..., ftjt) (4 squares,
i' and j' are point positions on the feature vector array in the S-th layer. Discriminant analysis methods are well known, and are described, for example, in Yanai and Takeuchi, Projection Matrix, -Passive Matrix, Singular Value Decomposition (1983), 'P 161-163, University of Tokyo Press. A simple outline will be explained using FIG. 3. Average value 3003.3004 for each category in the coordinate system 3000 on a two-dimensional plane as shown in the third figure.
Assume that clusters 3001 and 3002 are given. Apply discriminant analysis to this. 2 clusters 3001
．． 3002 is normalized so that the variance on the vertical and horizontal two axes is 1, the clusters 3101 and 31 in FIG. 3(b) are
It costs like 02. Here, the coordinate system 3100 is translated in parallel to the coordinate system 3200 using the average value of all data. Furthermore, in this coordinate system 3200, a principal axis 3300 with the maximum scattering is determined. This axis 3300 becomes a new feature axis effective for identification.

(Example) Hereinafter, the present invention will be explained in detail with reference to Examples.

FIG. 1 is a block diagram showing an embodiment of the pattern recognition method of the present invention. However, character pattern recognition in the case of a four-layer configuration will be taken as an example.

In FIG. 1, 1 is a feature point selection unit, 2 is a linear integration calculation unit, 3 is a deviation correction calculation unit, 4 is a similarity calculation unit, 5 is a local feature vector storage unit, and 6 is a determination unit for determining the maximum value. means,
10 is a pattern memory of a pattern storage means for storing character patterns; 20, 21, and 22 are feature vector array memories of the 2nd, 3, and 4th layers, respectively; 1100, 1101, . . .
is the local feature vector memory in the first layer, 1200.
1201, . . . are local feature vector memories in the second layer, 1300, 1301, .
Linear integration matrix memory in the 3rd layer, 40.41.42 is the deviation correction vector memory in the 1.2.3 layer, 50
．． 51, 52, 53, 54, 55, 56 are multiplexers, and 260, 61, 62 are feature point position memories in the 1.2.3 layer. 70 is a control section.

According to a command from the control circuit 70, the stored feature vector group is stored in the pattern memory 10 of the pattern storage means, for example, 1
A local region such as 000 is sent to the feature point selection section 1 through the multiplexer 52. The feature point selection unit 1 reads the feature point selection positions of the first layer stored in the feature point position memory 60 through the multiplexer 56, selects a plurality of feature points in the local area according to the feature selection positions, and linearly It is sent to the integrated calculation unit 1. The feature point selection position is, for example, the number of layers "l+a2," as shown in FIG. 4.
and the relative position (bl, c1) from the position (i, j),
(represented by b2.C2>,...) The linear integration calculation unit 2 stores the linear integration matrix of the first layer in the linear integration matrix memory 3.
0 through multiplexer 54, (1)
Linear integration is performed using Eq. The linearly integrated result is sent to the deviation correction calculation unit 3, and the deviation correction vector of the first layer is read out from the deviation correction vector memory 40 through the multiplexer 55 (deviation correction is performed using Equation 2). After the deviation correction is completed, the similarity calculation unit 4 reads the local feature vectors from the local feature vector memories 1100, 1101, etc. one after another through the multiplexer 52, and the similarity is calculated according to equation (3). The obtained similarity is written into the feature vector array memory 20.This series of processing is
This is performed within a predetermined range of the pattern memory 10 of the pattern storage means.

When the first layer processing is completed, a group of feature vectors stored in a local area such as 1001 is transferred from the feature vector array memory 20 to the feature point selection unit 1 through the multiplexer 52 according to a command from the control circuit 70. Sent. The feature point selection unit 1 selects the second
The feature point selection position of the layer is read out through the multiplexer 56, and a plurality of feature points in the local area are selected according to the feature selection position and sent to the linear integration calculation unit 1.

The linear integration calculation section 1 reads out the second layer linear integration matrix from the linear integration matrix memory 31 through the multiplexer 54, and performs linear integration using equation (1). The linearly integrated result is sent to the deviation correction calculation unit 2, and the deviation correction vector of the second layer is read out from the deviation correction vector memory 41 through the multiplexer 55 (deviation correction is performed using Equation 21). After the deviation correction is completed, the similarity calculation unit 3 stores the local feature vectors in the local feature vector memories 1200 and 1201.
, . . . are read out one after another through the multiplexer 52, and the similarity is calculated according to equation (3). The obtained similarity is written into the feature vector array memory 21. This series of processing is performed within a predetermined range of the feature vector array memory 20.

When the second layer processing is completed, the stored feature vector group is sent from the feature vector array memory 21 to the feature point selection section 1 through the multiplexer 52 in response to a command from the control circuit 70. The feature point selection unit 1 stores the feature point position memory 6.
The feature point selection position of the third layer stored in 2 is read out through the multiplexer 56, and a plurality of feature points in the local area are selected according to the feature selection position and all are sent to the linear integration calculation unit 2. The linear integration calculation section 2 reads out the linear integration matrix of the third layer from the linear integration matrix memory 32 through the multiplexer 54, and performs linear integration using equation (11).The linear integration result is sent to the deviation correction calculation section 3. , is read out from the third layer deviation correction vector 42 through the multiplexer 55 (deviation correction is performed using equation 21. After the deviation correction is completed, the similarity calculation unit 4 reads the local feature vector 1300 . 1301, . . . are read out one after another through the multiplexer 52, and the similarity is determined according to equation (3).The obtained similarity is written into the feature vector array memory 22.

In response to a command from the control circuit 70, the determining means 6 reads the degree of similarity from the feature vector array memory 22, identifies a predetermined category having the maximum value, and outputs the category.

The linear integration calculation unit 2 requires multiplication and addition, the deviation correction calculation unit 3 requires subtraction, and the similarity calculation unit 4 requires addition, multiplication, division, and square root calculation, but these are performed using known methods. Since this can be realized by a circuit, the explanation will be omitted.

Note that the present invention is not limited only to the above embodiments. For example, the size, number, etc. of local feature patterns in each layer may be determined according to specifications, and the selection positions of features may also be set as necessary. Furthermore, although the similarity is determined using a normalized inner product, it can also be achieved using Euclidean distance, City Block distance, or the like. Furthermore, although each layer is a single layer, it can also be realized by performing multiple regions in parallel. In short, the present invention can be modified and implemented in various ways without departing from the gist thereof.

(Effects) As described above, according to the present invention, features are extracted and determined hierarchically from the local area to the whole, while selecting feature points based on the variance ratio in the local area and performing intermediate discrimination using discriminant analysis. By doing so, it is possible to absorb the effects of noise, line width fluctuations, and positional movement in local regions, and extract features effective for identification that take into account local feature patterns in each layer or differences between categories of character patterns. Therefore, highly accurate pattern recognition can be realized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the pattern recognition method of the present invention, FIG. 2 is a diagram for explaining the invention in detail, FIG. 3 is a diagram for explaining discriminant analysis, and FIG. The figure is a diagram for explaining feature point selection positions. In the figure, 1 is a feature point selection unit, 2 is a linear integration calculation unit, 3 is a deviation correction calculation unit, and 4
Reference numeral 5 indicates a similarity calculating section, 5 a local feature vector storage section, 6 a determining means, 10 a pattern storage means, and 70 a control section. However, since the wiring from the control section would complicate the diagram, it is omitted, and the input and output are shown with arrows.

Claims (1)

[Claims] 1. A pattern storage unit that stores unknown patterns, a feature point position storage unit that stores feature point selection positions of local feature patterns in each of the plurality of layers, and stores similarities that are identification results. a feature vector array storage unit, a feature point selection unit that selects feature points according to the feature point selection position in the local area pattern, and a linear integration calculation unit that generates a feature vector array of each layer using a discriminant analysis method. , a feature extracting means comprising a deviation correction calculating section and a similarity calculating section, and a determining means for determining a character into a predetermined category by comparing the degrees of similarity output from the similarity calculating section of the top layer feature extracting means. The output of the pattern storage means is taken as the feature vector array of the lowest layer, and in each layer, feature vectors are selected according to the feature point selection positions in the feature vector arrays of all the lower layers, and a discriminant analysis method is used. A pattern recognition method characterized in that the degree of similarity between a predetermined local feature pattern of each layer and a feature vector is set as the feature vector array. 2. The pattern recognition method according to claim 1, wherein the degree of similarity is determined by a normalized inner product of the feature vector group and the local feature vector.