JPH064714A - Pattern recognition dictionary generating device and pattern recognizing device - Google Patents

Abstract

(57) [Abstract] [Purpose] Conventional pattern recognition in which the geometrical deformation of a pattern cannot be absorbed by expressing the features of a pattern as a relative position between characteristic points and a one-dimensional sequence of feature quantities at those points. By eliminating the drawbacks of the method, it is possible to follow up even when the geometric deformation of the input pattern is large, and enable correct recognition. In a device for obtaining a feature amount from an image pattern and generating a dictionary for recognition, a portion for calculating a feature amount at each point of an image to generate a feature map and a portion for storing the feature map And the part that selects the point with the maximum evaluation value in the feature map, and the evaluation value of the feature value of the point cloud around the previously selected point and the selected point in the feature map. , An emphasis control part for emphasizing an evaluation value of a part of a point group other than the suppressed points to control the position of a point to be selected next, a series of generated movement vectors and movement of movement vectors. It is provided with a portion for storing a series of feature quantities of the previous points.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition dictionary generation device and a pattern recognition device, and more particularly to a pattern recognition dictionary generation device and a pattern recognition device for extracting a series of feature point movement vectors for verification.

[0002]

2. Description of the Related Art Conventional pattern recognition methods for image patterns include "Electronic Information and Communication Handbook, First Volume,
p. 1117, 5th department, 1.6 Recognition of character image ", there is a representative method called a pattern matching method. In this method, at each point of the input image, a feature amount representing the feature of the image at that point is extracted, sampled at fixed fixed intervals of the image, and arranged in a line and represented as a feature vector. Alternatively, the image pattern is divided into fixed squares, the average value of the feature amount in each square is obtained, and the obtained values are arranged in a line and expressed as a feature vector.

To generate the dictionary, the feature vector of the corresponding category is used. Further, at the time of recognition, the feature vector of the corresponding category is obtained, and the dictionary is generated based on the feature vector. At the time of recognition, the feature vector is calculated from the input pattern by the same method as the dictionary generation, and the similarity between the dictionary feature vector of the corresponding category and the feature vector obtained from the input pattern is calculated. To do. The category having the feature vector with the highest similarity is used as the recognition result.

[0004]

When recognizing an image pattern, a standard pattern is geometrically deformed as an input pattern, and although there are local features, the local features exist. The position of is deviated from the original position. In the pattern held in the dictionary shown in FIG. 8, the vertices of the figure are overlapped with the sample points 802 and feature extraction is performed. However, in the case where the pattern to be recognized is as shown in FIG. 9, the vertex of the figure is at the point 906, even though it is the same category as the category in FIG. At 902, the feature at that vertex cannot be sampled. Therefore, the pattern shown in FIG. 9 may not be recognized correctly. this is,
Since the features sampled at regular intervals are represented by the absolute coordinates in the pattern, the recognition cannot be performed without being affected by the positional deviation of the features caused by the geometric deformation of the pattern.

However, the point 806 of the standard pattern in FIG.
The bent shape and the T-shaped shape of the point 805 can be detected at the points 906 and 905 in the input pattern shown in FIG. Furthermore, point 90 in FIG.
The relative position of the point 905 from 6 is almost the same as the relative position of the point 806 to the point 805 in FIG. An object of the present invention is not to express the features of a pattern by the absolute position in the pattern and the feature amount at that point, but by the relative position between characteristic points and the feature amount at that point, The pattern recognition dictionary generator and pattern recognition for eliminating the disadvantage of the conventional pattern recognition method that cannot absorb the geometrical deformation of the pattern and enabling the correct recognition by following the large geometrical deformation of the input pattern To provide a device.

[0006]

A pattern recognition dictionary generating apparatus according to a first aspect of the present invention obtains a feature amount from an image pattern,
In a device for generating a dictionary for recognition, a means for calculating a feature amount at each point of an image to generate a feature map, a means for storing the feature map, and a maximum evaluation amount in the feature map. A means for selecting a point having a value, a movement vector generating means for obtaining a movement vector from the previously selected point, a selected point, and a feature amount of a point group around the selected point in the feature map. An emphasis suppression control unit that suppresses the evaluation amount and emphasizes the evaluation amounts of some other points to control the position of the point to be selected next, and a series of generated movement vectors and movement vectors. It is characterized in that it is provided with a means for storing a series of feature quantities of the destination point.

In the pattern recognition dictionary generating apparatus of the second invention, in the first invention, the emphasis suppression control means suppresses or emphasizes the evaluation amount of the feature quantity of the suppressed or emphasized point according to the passage of time thereafter. It is further characterized by having a function of controlling the recovery to the value before being performed.

The pattern recognition dictionary generating apparatus according to the third invention is the pattern recognition dictionary generating apparatus according to the second invention, which detects that the sequence of movement vectors generated by the vector generating unit is circular, and moves. And a cyclic detection means for outputting a characteristic amount associated therewith.

A pattern recognition apparatus according to a fourth invention is the pattern recognition apparatus according to the first,
In a device that performs pattern recognition using a dictionary generated by the pattern recognition dictionary generating device according to the second or third aspect of the invention, a point that is a point in the pattern that is the starting point and that is moved by the movement vector read from the dictionary is constant. Means for detecting a point within the range that is closest to the feature amount read from the dictionary;
A means for accumulating a difference in feature amount between the selected point and the dictionary and determining whether the input pattern can be matched with the category of the dictionary being collated is provided.

A pattern recognition apparatus according to a fifth aspect of the invention is
A device for performing pattern recognition using a dictionary generated by the pattern recognition dictionary generating device according to the second or third aspect of the present invention, means for obtaining a sequence of movement vectors and a sequence of feature amounts when the dictionary is generated, and the generated means. The sequence of movement vector features and the sequence of movement vectors and features of a certain category stored as a dictionary are sequentially collated, and if the accumulated difference is less than a certain value, then the input pattern matches the category. And a determining means.

[0011]

According to the present invention, the feature amount is obtained for each pixel or each grid of the image pattern from the image pattern, the feature map is generated from the feature amount, and the evaluation value at each point of the feature amount is obtained. At position (x, y) in the image pattern,
When the type of feature amount is k, the feature amount at each feature point is f (x, y, k). Furthermore, the evaluation value obtained from the feature amount is g (x, y, kl). Here, kl indicates a set of feature types, which is a subset of the set K of all features.

For example, the evaluation value of the feature can be expressed by the variance of kl, but this is not an essential problem in the present invention.
When expressing the distributed evaluation value,

[0013] It represents an average value in kl, and n represents the number of elements of k belonging to kl.

Although it is possible to easily extend even if K is considered to have a plurality of subsets kl, all the feature evaluation values obtained for each point in each pattern for simplification of explanation. The case of defining from the feature k will be described. Therefore, hereinafter, the feature amount for the feature k (1 or less and K or less) is treated as a vector having K elements, and is set to f (x, y), and the point (x,
The evaluation value of the feature set in y) is g (x, y).

The larger the evaluation value, the more characteristic the points are, and the patterns are sequentially selected from the patterns. The movement vector in the pattern between the selected points is expressed as the relative position between the points. This converts a two-dimensional or more pattern into a one-dimensional series pattern by sequentially selecting from characteristic elements.

In the first aspect of the invention, the evaluation values g (x, y) in the neighborhood of the feature point selected immediately before are suppressed so that no more than one feature point existing in the vicinity is selected. . In addition, by emphasizing the evaluation value g (x, y) of the feature in a slightly wider range than the suppressed place, the feature point to be selected next can be a feature that is close to the feature point without being selected too far. Selected by connecting dots. The case without emphasis is also an example of this method, and in this case, the position of the next selected feature point is not restricted.

The suppression and the emphasis are realized as a filter function showing the degree of the emphasis suppression as shown in FIGS. 6 and 7. Let h (x, y) be the filter function, g ′ (x, y, t) be the evaluation value of the feature at the previous step t, and the feature points selected at that time be (x (t), y (t )), The evaluation value at time t + 1 is g ′ (x, y, t + 1) = g ′ (x, y, t) · h (x−x (t), yy (t)) (equation Calculated in 2).

Further, the maximum value is selected from g '(x, y, t + 1), and the point (x (t + 1), y having the maximum value is selected.
(T + 1)) as the next feature point, and the movement vector (x (t + 1) − between the feature point at time t and the feature point at time t + 1.
x (t), y (t + 1) -y (t)) and (x (t +
1), y (t + 1)) in the feature vector f (x (t +
1), y (t + 1)) are output as a part of the dictionary for recognition.

The recognition dictionary of the target pattern is generated by repeating the selection of the feature point position, the output of the movement vector, and the feature vector at the feature point a certain number of times.

In the second invention, the part for obtaining the initial feature vector f (x, y) and its evaluation value g (x, y) in the pattern recognition dictionary generator of the first invention is the same, The method of selecting the feature points (x (t), y (t)) is also the same, but when the evaluation value g ′ is modified, only the enhancement suppression is not used, but the initial evaluation obtained by (Equation 1) is used. It is characterized in that a recovery term is added using the value g. That is, (Equation 2) is changed to g ′ (x, y, t + 1) = a · g ′ (x, y, t) · h (x−x (t), yy (t)) + b · g (x, y) It is used as (Equation 3). Here, a and b are coefficients for controlling recovery, b is a parameter indicating that g ′ recovers to the initial value g in about (1 / b) steps, and a is a value close to 1 (1 -B) is good. The selection of the feature points and the output of the movement vector and the feature vector at the selected feature points are repeated a fixed number of times to generate the recognition dictionary of the target pattern.

In the third invention, the feature vector f'generated in the pattern recognition dictionary generating apparatus of the second invention is used.
Rather than detecting a cycle from the sequence of (t) and the movement vector (dx (t), dy (t)) and selecting a feature point by repeating a predetermined number of times, when the cycle is detected The selection of the feature points ends.

Assuming that the cyclic period is c, the approximation threshold is δ, and the present is the t'th step, all c
When | f ′ (t′−h) −f ′ (t′−c−h) | <δ (formula 4) holds for the following h, it is assumed that the cycle of the cycle c is detected and the dictionary End the generation process of f '(t'), (dx
(T '), dy (t')), f '(t'-1), (dx
From (t'-1), dy (t'-1)) to f '(t'-c
+1), (dx (t'-c + 1), dy (t'-c +
Up to 1)), a total of c sets of feature vectors and movement vectors are output as a recognition dictionary. After the feature points have been selected a certain number of times, if no tour is detected, the feature vectors and movement vectors at all the feature points are output.

In the above first, second, and third inventions, the dictionary for the i-th representative pattern is such that the category name Ci, the number of pairs of feature vectors and movement vectors ni, and the features from the first to the ni-th feature. The vector f ′ (Ci, j) and the movement vector (dx (Ci, j), dy (Ci, j)). However, j is a feature point number and is 1 or more and ni or less. The feature vector is treated as a vector having K elements.

It is an object of the fourth and fifth inventions to collate the input pattern with the target category Ci. In the fourth invention, first, the feature extraction is performed by the same method as that used in the pattern recognition dictionary generating apparatus of any one of the first, second and third inventions, and the feature map f (x, y) is extracted from the input pattern. Is generated. Next, select the first feature vector f ′ (Ci, l) in the recognition dictionary,
The most similar feature vector is selected from the feature map. Let it be f (xl, yl). The similarity is defined by the Euclidean distance between Euclidean feature vectors and the simple similarity. If the Euclidean distance is used, the distance d is

[0026]

[0028] Represents the norm of A. In the example of the two similarities, the closer the pattern is, the smaller the similarity is, and either one may be used. That is, the first feature vector f ′ in the recognition dictionary
(Ci, l) is selected, and the point (xl, yl) that gives the minimum similarity is selected from the feature map f (x, y), and the similarity is stored as the cumulative similarity D.

Next, the second movement vector (dx (C
i, 2), dy (Ci, 2)) is taken out, and (xl, y
l), the next feature point (x2 ′, y2 ′) = (xl
+ Dx (Ci, 2), yl + dy (Ci, 2)) are calculated. Furthermore, the second feature vector f ′ (Ci, 2)
A similar feature to is searched for around (x2 ′, y2 ′).
The point that gives the minimum similarity is (x2, y2), and the similarity is added to the cumulative similarity D. 3rd, 4th
The search for the feature point that gives the above minimum similarity is repeated up to the n-th and the n-th. If the cumulative similarity exceeds the threshold value on the way, it is determined that the input pattern does not belong to the target category, and if the cumulative similarity up to the nith feature point is equal to or less than the threshold value, the input pattern is input. It is determined that the pattern belongs to the target category with a reliability of 1 / D. When a plurality of feature points similar to the first feature vector f (Ci, l) are obtained from the feature map obtained from the input pattern, the above process is performed on all of them. In the fifth invention, pattern recognition is performed using the recognition dictionary generated by any one of the pattern recognition dictionary generation devices of the first, second, and third inventions. In addition, the feature map f (x, y) is calculated by the feature extraction method used when the recognition dictionary is generated,
Furthermore, a sequence of feature vectors and movement vectors is generated by the feature point generation method, and the sequence of f (k), (dx ′ (k),
dy ′ (k)). Here, k is a natural number from 1 to N, and N is the number of feature points detected from the input pattern.

Next, a method of collating the feature vector and movement vector obtained from the input pattern with the feature vector and movement vector of the Ci category in the dictionary will be described.

The similarity d between the first feature vector f (l) in the input pattern and the feature vector f '(Ci, k) in the dictionary is obtained, and the jth point giving the minimum similarity is found. obtain. The similarity is assigned to the cumulative similarity D.

Next, the second input feature vector f
(2) A movement vector (dx ′ (2), dy ′ (2)) is selected, the similarity with f ′ (Ci, j + 1) in the dictionary is d, and the movement vector (dx (Ci, j + 1), dy (C
i, j + 1)) and the squared error s. When the similarity d is smaller than the threshold ε and smaller than the moving vector error s-screen threshold η, the similarity d is added to the cumulative similarity D.

When the movement vector error s is larger than the threshold value η, the sum of the j + 1th movement vector and the j + 2nd movement vector is set as a new movement vector (dx '(2), dy' ( 2)) and the error s is calculated, and the j + 2nd feature vector f ′ (Ci,
j + 2) and the input feature vector f (2) are found to have a similarity d. The similarity d and the movement vector error s are set to thresholds ε and η
If it is smaller than the threshold value, the cumulative similarity D is added with the similarity d, and the detected number E of similar feature vectors E is added with 1. If this condition is not satisfied, the similarity vector d and the movement vector error s are calculated using the movement vector obtained by adding the movement vectors in the order j + 1, j + 2, j + 3, and the feature vector at the point beyond that, and It is determined whether or not the threshold values are equal to or less than ε and η. If either the similarity or the movement vector error is greater than the threshold value even after adding the J movement vectors, it is determined that the second feature vector of the input pattern does not belong to the target category Ci. To do.

The above procedure is performed for the third to Nth feature vectors and movement vectors of the input pattern,
The similarity is accumulated in D. After performing the above processing on all N feature vectors, the cumulative similarity D is set to N.
Average similarity D ′ divided by and the number of similar feature vector detections E
Is divided by N to obtain a similar feature vector detection rate E ′.

The average similarity D and the similar feature vector detection rate E ′ are compared with the respective threshold values, and the average similarity D ′ is smaller than the threshold R, and the similar feature vector detection rate E ′ is obtained.
Is larger than the threshold value T, it is determined that the input pattern belongs to the category Ci.

[0037]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A pattern recognition dictionary generation device and a pattern recognition device according to the present invention will be described below with reference to the drawings. Figure 1
Shows the configuration of the pattern recognition dictionary generation device of the first invention. The pattern input from the pattern input unit 101 is
The movement vector feature point generation unit 110 generates N movement vectors and feature vectors at N feature points, and stores the movement vector feature point storage unit 106.

The inside of the motion vector feature point generator 110 will be described. The pattern input from the pattern input unit 101 is feature-extracted by the feature map generation unit. The input pattern is an image pattern I (x, y) of m × m pixels. It is a natural number of 1 or more and m or less together with x and y. The operation of the feature map generation unit 102 will be described by taking a case of extracting Gabor features as an example. Gabor features can be extracted by convolving an image with a Gaussian filter and a Gabor filter that is a filter that convolves a two-dimensional sine wave, and features that depend on direction and scale are extracted for each point in the image. The Gabor filter has the following formula: G (x, y, u, p) = exp (iu (cos (p), sin (p)). (X, y)) exp (-u ² | (x, y) | ² / 2v ² ) / q (Equation 7) Here, i is an imaginary unit and q is a normalization coefficient.
v is a coefficient representing the variance of the Gaussian filter, and is set to π here, but this value is not essential. Further, u is a parameter indicating the frequency of the two-dimensional sine wave, and p is a parameter indicating the direction in which the two-dimensional sine wave propagates. In the embodiment, u is two kinds of 16π / m and 32π.
/ M, but this is not an essential problem and may be set according to the scale of the feature to be extracted. Further, a single scale or a plurality of scales of 3 or more may be used, and in that case, the number of dimensions of the feature vector of each point in the feature map changes. Further, in this embodiment, the direction p is represented by eight directions, and eight kinds of directions are taken every 0 to π / 8. This is not an essential problem as well, and features can be extracted in arbitrary directions. The Gabor filter shown above is used to perform convolutional feature extraction on the input image using the following equation.

[0039] The number of dimensions of the feature vector at each point (x ', y') is 16. Further, the variance of f (x ', y', u, p) with respect to u and p is obtained according to (Equation 1), and the point (x ', y')
And an evaluation value g (x, y) of Here, n in (Equation 1)
Is 16, and k representing the type of feature is a set of 16 types of scales and directions represented by (u, p). There is one kind of k1 and it is ignored here.

The feature vector f (x ', y', u, p) generated by the feature map generator 102 and the evaluation value g (x ',
y ′) is transferred to the feature map storage unit 103, and its copy g ′ (x ′, y ′, 0) is stored. Furthermore, g '
(X, y, 0) is transferred to the feature point selection unit 104, and the point (x (1), y (1)) having the maximum value is selected.
The coordinates (x (t), y (t)) (t = 1) of the point and the feature vector f (x (t), y (t), u, p) (t = 1)
Are transferred to the movement vector 105.

The selected feature point (x (t), y
(T) is transferred to the emphasis suppression control unit 107.

The forced suppression control unit 107 applies an emphasis suppression filter according to (Equation 2) with the received feature point (x (t), y (t)) as the center. Here, a filter as shown in FIG. 6 defined by the following formula is used as the enhancement suppression filter, but this is not an essential problem and a filter of any shape may be used. h (x, y) = (1-a + a · exp (A | (x, y) | ² )) (1-b · exp (B | (x, y) | ² )) (Equation 9) where , A = 0.7, b = 1, A = m / 2, B = m /
It was set to 8.

The filtering result g '(x, y, t +
Also from 1), the feature points are selected and the enhancement is suppressed, and the feature vector f (x (t), y (t), u, p) and the feature point coordinates (x (t), y (t)) are sequentially obtained. Movement vector generation unit 10
Transfer to 5.

In the movement vector generation unit 105, two subsequent feature point coordinates (x (t), y (t)) and (x (t-
1), y (t-1)) to the movement vector (dx (t),
dy (t)) is calculated by the following formula. (Dx (t), dy (t)) = (x (t), y (t))-(x (t-1), y (t-1)) (Formula 10) Generated movement vector (dx (T), dy (t)) and the feature vector f (x (t), y at the feature point after that
(T), u, p) are transferred to the movement vector feature point storage unit 106 and sequentially accumulated and stored as a recognition dictionary. The category name Ci of the finally input pattern and the stored number of movement vectors and the number of characteristic points ni are also stored.

Next, an embodiment of the second pattern recognition dictionary generating apparatus will be described with reference to FIG. The pattern input from the pattern input unit 101 is the movement vector feature point generation unit 2
At 10, N motion vectors and feature vectors at N feature points are generated and stored in the motion vector feature point storage unit 106.

The inside of the movement vector feature point generator 210 will be described. The pattern input from the pattern input unit 101 is feature-extracted by the feature map generation unit 102 according to (Equation 8). Furthermore, f (x ', y', u,
The variance of p) with respect to u and p is obtained according to (Equation 1), and is set as the evaluation value g (x, y) of the point (x ′, y ′).

The feature vector f (x ', y', u, p) generated by the feature map generator 102 and the evaluation value g (x ',
y ′) is transferred to the feature map storage unit 103, and its copy g ′ (x ′, y ′, 0) is stored.

Furthermore, g '(x, y, 0) is transferred to the feature point selection unit 104, and the point (x (l),
y (l)) is selected. The coordinates of that point (x (t), y
(T)) (t = 1) and the feature vector f (x (t), y
(T), u, p) (t = 1) is the movement vector generation unit 10
5 is transferred.

In the movement vector feature point generation unit 210, the feature map generation unit 102 and the feature map storage unit 10
3, feature point selection unit 104, movement vector generation unit 105
Performs the same operation as the movement vector feature point generation unit 110.

The emphasis, suppression, and recovery control unit 207 performs emphasis suppression using the filter h (x, y) of (Equation 9), and the initial evaluation value g obtained from the input pattern.
An evaluation value g ′ (x, y, t + 1) for selecting the next feature point is obtained according to (Equation 3) using (x, y). Here, the coefficients a and b of (Equation 3) are 0.9 and 0.1, respectively.
However, it is not limited to this value.

The filtering result g '(x, y, t +
Also from 1), feature point selection and enhancement suppression recovery are performed, and feature vector f (x (t), y (t), u, p) and feature point coordinates (x (t), y (t)) are sequentially obtained. Movement vector generation unit 10
Transfer to 5.

In the movement vector generation unit 105, two subsequent feature point coordinates (x (t), y (t)) and (x (t-
1), y (t-1)) to the movement vector (dx (t),
dy (t)) is calculated according to (Equation 10).

The generated movement vector (dx (t), d
y (t)) and the feature vector f (x
(T), y (t), u, p) are transferred to the movement vector feature point storage unit 106 and sequentially accumulated and stored as a recognition dictionary. The category name Ci of the finally input pattern and the stored movement vector number and feature point number ni are also stored.

Next, an embodiment of the pattern recognition dictionary generating apparatus of the third invention will be described with reference to FIG. Pattern input section 1
In the pattern input from 01, the motion vector feature point generation unit 310 generates a feature vector at the motion vector feature point, detects a cyclic portion from the sequence, and detects ni moving vectors of the detected cyclic portion and ni. The individual feature vectors are stored in the movement vector feature point storage unit 106.

The inside of the motion vector feature point generator 310 will be described. The pattern input from the pattern input unit 101 is feature-extracted by the feature map generation unit 102 according to (Equation 8). Furthermore, f (x ', y', u,
The variance of p) with respect to u and p is obtained according to (Equation 1), and is set as the evaluation value g (x, y) of the point (x ′, y ′).

The feature vector f (x ', y', u, p) generated by the feature map generator 102 and the evaluation value g (x ',
y ′) is transferred to the feature map storage unit 103, and its copy g ′ (x ′, y ′, 0) is stored.

Further, g (x, y, 0) is transferred to the feature point selection unit 104, and the point (x (1), y having the maximum value thereof).
(1)) is selected. The coordinates of that point (x (t), y
(T)) (t = 1) and the feature vector f (x (t), y
(T), u, p) (t = 1) is the movement vector generation unit 10
5 is transferred.

In the movement vector feature point generation unit 310, the feature map generation unit 102 and the feature map storage unit 10
3, feature point selection unit 104, movement vector generation unit 105
Performs the same operation as the motion vector feature point generator 110,
The enhancement suppression recovery control unit 207 performs the same operation as that of the movement vector feature point generation unit 210.

The filtering result g '(x, y, t =
Also from 1), feature point selection and enhancement suppression recovery are performed, and feature vector f (x (t), y (t), u, p) and feature point coordinates (x (t), y (t)) are sequentially obtained. Movement vector generation unit 10
Transfer to 5.

In the movement vector generation unit 105, two subsequent feature point coordinates (x (t), y (t)) and (x (t-
1), y (t-1)) to the movement vector (dx (t),
dy (t)) is calculated according to (Equation 10).

The generated movement vector (dx (t), d
y (t)) and the feature vector f (x
(T), y (t), u, p) are transferred to the cycle detection unit 308, and the cyclic test is performed from the t'th element of the sequence in descending order to the first.

First, assuming that the cycle period is c, the approximation threshold is δ, and the present is the t'th step, when (Equation 4) is satisfied for all h less than or equal to c, Assuming that the cycle c has been detected, the dictionary generation process is terminated, and f ′ (t ′), (dx (t ′), dy
(T ')), f'(t'-1), (dx (t'-1),
dy (t'-1)) to f '(t'-c + 1), (dx
Up to (t'-c + 1), dy (t'-c + 1)), total c
A set of feature vectors and movement vectors is output as a recognition dictionary. After the feature points have been selected a fixed number of times, if no tour is detected, the feature vectors and movement vectors at all the feature points are output. The output movement vector and feature vector are stored in the movement vector feature point storage unit 10
6 is stored together with the category name Ci of the pattern and the stored number of movement vectors and the number of characteristic points ni (= c).

Next, an embodiment of the pattern recognition apparatus of the fourth invention and the first invention will be described with reference to FIG. The pattern recognition device of the fourth invention can recognize the pattern recognition dictionary generated by any one of the pattern recognition dictionary generation devices of the first, second, and third inventions, but here, the recognition dictionary of the first invention is recognized. A description will be given using the recognition dictionary generated by the generation device.
The movement vector and the feature vector generated by the pattern recognition dictionary generating device are stored in the movement vector feature point storage unit 10 together with the category name Ci and the feature score ni of the pattern.
6 is stored. The pattern to be recognized is input from the pattern input unit 101, converted into a feature vector by a feature map generation unit 102 similar to the movement vector feature point generation unit 110 in FIG. 3, and f (x, y,
u, p) are stored in the feature map storage unit 103.

First, matching with the category Ci is performed. The next feature point selection unit 421 first extracts the first feature vector f ′ (Ci, l, u, p) of the category Ci from the movement vector feature point storage unit 106, and the feature point matching unit 42.
Send to 2. Further, the next feature point selection unit 421 receives the feature vector f for each coordinate (x, y) from the feature map storage unit 103.
(X, y, u, p) is taken out and transferred to the feature point matching unit 422.

In the feature matching unit 422, the similarity d between the received dictionary feature vector f'and the input feature vector f
(X, y) is calculated according to (Equation 6). The obtained similarity d (x, y) is transferred to the matching degree determination unit 423, and compared with the minimum similarity D, and if d (x, y) is smaller than the minimum similarity D, the minimum similarity D Be d (x, y), and its coordinates (x, y) are stored. The signal is transmitted from the matching degree determination unit to the matching control unit 422. In the matching control unit 422,
A signal obtained by sequentially counting up (x, y) is sent to the next feature point selection unit 421, the feature vector f of the coordinate (x, y) is read from the feature map storage unit 103, and transferred to the feature point matching unit 422. Let The similarity d (x, y) obtained by the feature point matching unit 422 is sequentially transferred to the matching degree determination unit 423, and the above-described minimum similarity calculation is performed. After the feature vectors of all the coordinates are evaluated, the minimum similarity D of the matching degree determination unit 423 is stored as D (t) with t = 1, and the coordinates giving the minimum similarity are (x (t), y (t). )).

(X (t), y (t)) is the matching control unit 42
4, the matching control 424 transfers the dictionary feature vector f ′ (Ci, t + 1,
u, p) and movement vector (dx (Ci, t + 1), dy
(Ci, t + 1) is transferred to the next feature point selection unit 421. Further, the matching control unit 424 determines that the coordinate value (x (t), y
(T)) is transferred to the next feature point selection unit 421.

The next feature point selection unit 421 has (x '(t + 1), y' (t + 1)) = (x (t), y (t)) + (dx (Ci, t + 1), dy (Ci, t + 1). ) (Equation 11) predicts the coordinates of the next feature point, reads the feature vector of points around the coordinates from the feature map storage unit 103,
It is transferred to the feature point matching unit 422 together with the dictionary feature vector f ′.

The feature point matching unit 422 calculates the similarity between the received dictionary feature f '(Ci, t + 1, u, p) and the feature vector f (x, y, u, p) of the input pattern. However, here, the region is represented by a square of one side (2r + 1) for convenience. That is, x ′ (t + 1) −r ≦ x ≦ x ′ (t + 1) + r (Equation 12) y ′ (t + 1) −r ≦ y ≦ y ′ (t + 1) + r. There is no problem with other peripheral area setting methods.

The degree of similarity is obtained according to (Equation 6) and is transferred to the matching degree judging unit 423 together with its coordinates (x, y). The matching degree determination unit 423 obtains the minimum similarity D and obtains the cumulative similarity D (t + 1) by the same procedure as the minimum similarity determination at t = 1. D (t + 1) = D (t) + D (Equation 13) D (t + 1 together with coordinates giving the minimum similarity is (x (t
+1), y (t + 1)) is transferred to the matching control unit 424 and stored.

In the matching degree judging unit 423, D (t + 1) is compared with a threshold value, and when it is smaller than the threshold value or when t is smaller than ni, t is set to t + 1 and a signal is sent to the matching control unit. The sending, the next feature point selection, and the cumulative similarity selection are repeated.

If D (t) is smaller than the threshold value when t = ni, the input pattern is determined to belong to the category Ci, and Ci is output from the terminal 425. t <n
If D (t + 1) becomes larger than the threshold value at i, the processing is interrupted, the input pattern is determined not to belong to the category Ci, and the signal is output from the terminal 425.

The above operation is performed by the movement vector feature point storage unit 1
This is performed for all categories in 06 and all category names to which the input pattern may belong are output.

Finally, an embodiment of the fifth invention will be described with reference to FIG. The pattern input from the input pattern unit 101 is converted by the movement vector feature point generation unit 510 as a series of feature vectors and movement vectors. The conversion method is the movement vector feature point storage unit 106 in which the recognition dictionary is stored.
The same one as the movement vector feature point generation unit when the is generated is used. Here, the movement vector feature point generation unit 11 used in the pattern recognition dictionary generation device of the first invention is used.
0 will be described, but the generation units 210 and 31 in other cases will be described.
The same applies to 0.

In the movement vector feature point generation unit 510, the feature vector f (t, u, p) and the movement vector (dx
It is assumed that n pieces of (t) and dy (t)) are generated. Therefore, the feature vector f ′ (Ci, t, u, p), which is the recognition dictionary of the Ci-th category, and the movement vector ((dx ′
(Ci, t), dy '(Ci, t)) is collated.

First, assuming that t = 1, the feature vector f (t,
u, p) and the movement vector (dx (t), dy (t)) are transferred to the feature point matching unit 522. Feature point matching unit 522
Is j = 1, and the dictionary feature vector f ′ (Ci, j,
u, p) and movement vector (dx'Ci, j), dy '
(Ci, j)) is read from the movement vector feature point storage unit 106, and (dx, dy) = (dx ′ (Ci, j), dy ′ (Ci, j)) (Equation 14). The similarity between the input feature vector f and the dictionary feature vector f ′ is obtained according to (Equation 6). Further, the error s between the input movement vector and the dictionary movement vector (dx, dy) is calculated according to (Equation 15). s = | (dx (t), dy (t))-(dx, dy) | (Equation 15) The feature point matching unit 522 transfers the similarity d and the error s to the matching determination unit 523. The matching determination unit 523 compares the similarity d and the error s with the respective thresholds ε and η, and when both are smaller than the threshold, the cumulative similarity is calculated as D = D + d, and the similar feature is calculated. After counting up the number of detections L, t is counted up as t + 1, the next input feature vector is taken out from the movement vector feature point generation unit 510, and a similar dictionary vector is searched.

If either the similarity d or the error s obtained by the feature point matching unit 522 is larger than the threshold value, j = j +
After counting up 1 and j, a signal is sent to the matching control unit 524, the j-th movement vector and the feature vector are read from the movement vector feature point storage unit 106, transferred to the feature point collation unit 522, and the feature point collation unit 522. Then, the calculation of (dx, dy) = (dx, dy) + (dx ′ (Ci, j), dy ′ (Ci, j) (Equation 16) is performed. Then, the similarity d and the error s are calculated according to (Equation 6) and (Equation 15), and these are transferred to the matching determination unit 523 and the t-th feature point matching determination is performed. If η and η are smaller than each other, the process proceeds to the next input vector matching, and if any of them is larger than the threshold value, the above feature point matching and matching determination are repeated.

After matching with all feature vectors,
If the similarity below the threshold is not obtained, the input feature vector f (t, u, p) is not present in the recognition dictionary.

After t is matched with all the feature vectors from 1 to n and the movement vector, the matching determination unit 523
An average similarity D ′ obtained by dividing the cumulative similarity D by n and a similar feature vector detection rate E ′ obtained by dividing the number E of detected similar feature vectors by n are obtained. If these are compared with a predetermined threshold value and D'is smaller than the threshold value R and E'is larger than the threshold value T, it is determined that the input pattern belongs to the category Ci, and the category name Ci is assigned to the terminal. Output from 425.

The above operation is performed by the movement vector feature point storage unit 1
All categories in 06 are output, and all categories to which the input pattern may belong are output.

In this embodiment, when there is no combined vector of dictionary movement vectors to be matched with the input movement vector, the input vector is simply rejected. However, the combined vector of a plurality of continuous input movement vectors and the dictionary movement are It can also be compared with vector composition, which can be easily inferred from the fifth invention. In addition, the matching determination unit 5 of the fifth invention
23, the input pattern series and the dictionary pattern series can be matched by using the DP matching method as shown in "Uesaka, Ozeki, Bunichi General Publishing, Pattern Recognition and Learning Algorithm, p.91".

According to the present invention, a pattern of two or more dimensions can be converted into a one-dimensional series of feature vectors according to the feature evaluation value. The evaluation value is determined by the distribution of the features and the size of the feature amount. Selecting a point having a large evaluation value means selecting a point effective for pattern recognition. Conventionally, when feature points are selected as sampling from a grid of evenly spaced two-dimensional patterns,
A point effective for recognition is not always selected. Therefore, it is necessary to set a sufficiently fine grid so that a point effective for recognition can be selected and to select a sufficient number of feature points. In the pattern recognition dictionary generation device of the present invention, feature points that are evaluated to be effective for recognition are sequentially selected to form a dictionary, so that there is no wasteful selection of feature points and effective recognition is realized with a small dictionary storage capacity. it can.

Further, in the recognition dictionary generated by the pattern recognition dictionary generating apparatus of the present invention, the relationship between the selected feature points is realized at the relative positions, so that the pattern is rotated or moved in parallel, or handwritten Chinese characters or the like. It is possible to realize the recognition without being affected by the change in the position of each radical and the distortion of the radical shape that frequently occur.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a pattern recognition dictionary generation device of a first invention.

FIG. 2 is a diagram showing a configuration of a pattern recognition dictionary generation device of a second invention.

FIG. 3 is a diagram showing a configuration of a pattern recognition dictionary generation device of a third invention.

FIG. 4 is a diagram showing a configuration of a pattern recognition device of a fourth invention.

FIG. 5 is a diagram showing a configuration of a pattern recognition device of a fifth invention.

FIG. 6 is an example of a filter for controlling suppression emphasis of an evaluation value when a feature point is selected from the evaluation values.

FIG. 7 is an example of a filter for controlling suppression emphasis of an evaluation value when selecting a feature point from the evaluation values.

FIG. 8 is a diagram showing a relationship between feature points of a standard pattern.

FIG. 9 is a diagram showing a relationship between feature points of a standard pattern.

[Explanation of symbols]

101 pattern input unit 102 feature map generation unit 103 feature map storage unit 104 feature point selection unit 105 movement vector generation unit 106 movement vector feature point storage unit 107 enhancement suppression control unit 110, 210, 310, 510 movement vector feature point generation unit 207 Enhancement suppression recovery control unit 308 Travel detection unit 421 Primary feature point selection unit 422, 522 Feature point matching unit 423, 523 Matching degree determination unit 424, 524 Matching control unit

Claims (5)

[Claims] 1. An apparatus for obtaining a feature amount from an image pattern and generating a dictionary for recognition, a means for calculating the feature amount at each point of an image to generate a feature map, and storing the feature map. Means, a means for selecting a point having the maximum value of the evaluation amount in the feature map, a movement vector generation means for obtaining a movement vector from the previously selected point,
The evaluation amount of the feature amount of the selected point and the point group around the selected point in the feature map is suppressed, the evaluation amount of a part of points other than them is emphasized, and A pattern recognition dictionary generation apparatus comprising: an enhancement suppression control unit that controls a position; and a unit that stores a sequence of generated movement vectors and a sequence of feature amounts of points at a destination of movement in the movement vector. 2. The enhancement suppression control means further has a function of controlling recovery of the evaluation amount of the feature quantity of the suppressed or emphasized point to a value before suppression or emphasis according to the passage of time thereafter. The pattern recognition dictionary generation device according to claim 1, wherein the pattern recognition dictionary generation device comprises. 3. A cyclic detection unit that detects that the sequence of motion vectors generated by the motion vector generation unit is cyclic and outputs the cyclic motion vector sequence and the characteristic amount associated therewith. The pattern recognition dictionary generation device according to claim 2, further comprising: 4. An apparatus for performing pattern recognition using a dictionary generated by the pattern recognition dictionary generating apparatus according to claim 1, wherein the dictionary is read from a point in a pattern serving as a starting point. A means for detecting the point closest to the feature amount read from the dictionary within a certain range of the points moved by the movement vector, and the difference between the feature amounts of the selected point and the dictionary are accumulated and input. A pattern recognition device, comprising means for determining whether or not a pattern matches a category in a collating dictionary. 5. An apparatus for pattern recognition using a dictionary generated by the pattern recognition dictionary generating apparatus according to claim 1, 2, or 3, wherein a sequence of movement vectors and a sequence of feature amounts when the dictionary is generated. And a sequence of the generated movement vector and the feature amount and a sequence of the movement vector and the feature amount of a certain category stored as a dictionary are sequentially collated, and if the accumulated difference is less than a certain value, input A pattern recognition device comprising: a unit that determines that a pattern matches a category.