JPH0916765A - Matching processing method for two-dimensional image - Google Patents

Abstract

PURPOSE: To shorten recognizing time by suppressing the increase of the number of templates without necessity to extract a point corresponding to the start point of template from an input pattern. CONSTITUTION: A step S1 is an image input part for inputting a two-dimensional image and concerning the input image (input pattern) inputted by this image input part, at a step S2, contour extraction, fine line and vectorizing are performed. A vector feature amount A is found here. A step S3 is the template expressed by N pieces of unit vectors and between a vector feature amount B of this template and the vector feature amount A provided by the step S2, similarity is found on the matching matrix of a step S4. Afterwards, the maximum value of similarity is defined as the optimum similarity between the input pattern and the template at a step S5 and the pattern of the template, for which this optimum similarity is maximum, is defined as the recognized result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a two-dimensional image matching processing method used for image recognition processing.

[0002]

2. Description of the Related Art A pattern matching method is used in image recognition processing and voice recognition processing. This pattern matching method is the most basic pattern recognition method that superimposes a prepared template (called a standard pattern or reference pattern) on an input pattern and identifies the input pattern depending on whether they match. is there. D is one of the pattern matching methods.
There is P matching, and this DP matching is widely adopted for the above-mentioned image recognition, voice recognition processing and the like.

Next, a case where this DP matching method is applied to recognize a two-dimensional image will be described. The steps of the recognition method in this case are as follows.

(1) First, contour extraction, thinning, and vectorization are performed on a two-dimensional image (input image) to be recognized, and the vector characteristic amount A is obtained as follows.

A = {a ₁ , a ₂ , ..., a _I } a _i = (x
_is , y _is , x _ie , y _ie ) = (start point and end point of vector) (2) Next, the recognition template B _t (t = 1, 2, ... T) registered in the computer is also The contour is vectorized as in the formula.

B _t = {b _t1 , b _t2 , ..., B _tJt } b _tj
= (X _is , y _is , x _tje , y _tje ) = (start point of vector,
End point) (3) After that, the similarity h (a _i , b _tj ) between the one-component vector a _i of the vector feature A of the input image and the one-component vector b _tj of the template t is obtained by the following formula.

[0007]

[Equation 1]

However, h (a ₁ , b _t1 ) = d (a ₁ ,
b _t1 ).

(4) Finally, h (a _I , b _tJt ) is _calculated for all the templates, and the two-dimensional image is recognized with the template pattern that minimizes this as the recognition result.

[0010]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
The matching method has the following problems.

(1) h (a ₁ , b _t1 ) = d (a ₁ , b _t1 ).
In other words, since the input pattern to be recognized and the starting point of the template are fixed, it is necessary to extract the point corresponding to the starting point of the template from the input pattern in order to obtain the correct similarity. An algorithm is needed.

(2) Similarity between feature quantities (direction vector)
Since it is calculated by the norm of, it becomes sensitive to the difference in length. Therefore, in order to recognize figures of the same shape but different sizes (similar figures, etc.), templates of respective sizes must be created. For this reason, if the number of templates increases, the time required for recognition (recognition time) also increases, which is inefficient.

(3) Even in the case of a figure having the same shape and the same size, it cannot be accurately recognized if the figure is rotated. Therefore, it is necessary to correct the rotation in the preprocessing stage for recognition using a special algorithm, or to create a template (a template in which a figure is rotated with respect to one figure) in consideration of each rotation. Therefore, whichever method is adopted, the recognition time is increased.

The present invention has been made in view of the above circumstances, and it is not necessary to extract the point corresponding to the starting point of the template from the input pattern, and the recognition time is shortened by suppressing the increase in the number of templates. Moreover, it is an object of the present invention to provide a two-dimensional image matching processing method capable of smoothly processing the rotation of a figure.

[0015]

SUMMARY OF THE INVENTION The present invention provides
In order to achieve the above-mentioned object, the first aspect of the invention is to obtain a first vector feature amount forming an input pattern by performing contour extraction, thinning, and vectorization of an input pattern input from an image input unit. , The first vector feature amount and the second vector feature amount of the template that is previously configured with N unit vectors are obtained by obtaining the similarity between the one-component vector of both feature amounts on the matching matrix. The maximum value of the similarity is set as the optimum similarity between the input pattern and the template, and the pattern of the template having the maximum optimum similarity is set as the recognition result.

A second aspect of the present invention is characterized in that an input pattern input from an image input unit and a template are expressed on a complex plane, information necessary for recognition is extracted, and then DP matching is performed based on the information. To do.

According to a third aspect of the present invention, the input pattern is subjected to contour extraction, thinning and vectorization to obtain a first complex correlation coefficient from a complex vector forming the input pattern, and then the first complex phase relationship is obtained. Number and the second complex correlation coefficient of the template composed of N complex unit vectors, the degree of similarity with the one-component vector of both complex correlation coefficients is obtained on the matching matrix, and the maximum degree of similarity obtained The value is the optimum similarity between the input pattern and the template, and the pattern of the template having the maximum optimum similarity is used as the recognition result.

A fourth aspect of the present invention is characterized in that the degree of similarity is a partial similarity between the inner product of the template and the input pattern in the complex vector space on the matching matrix.

[0019]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a flowchart showing the configuration of an embodiment of the present invention. In FIG. 1, step S1 is an image input section for inputting a two-dimensional image, and the input image (input pattern) input by this image input section is step S2. In, contour extraction, thinning, and vectorization are performed, and the vector feature amount A is obtained as described later. Step S3 is a template represented by N unit vectors, the details of which will be described later, and the vector feature quantity B of this template and the vector feature quantity A obtained in step S2 have a similarity on the matching matrix of step S4. Desired. Then, in step S5, the maximum value of the similarity is set as the optimum similarity between the input pattern and the template, and the pattern of the template having the maximum optimum similarity is set as the recognition result.

FIG. 2 is an explanatory diagram of the unit vectors forming the template described above. In FIG. 2, the template is represented by N unit vectors (vector feature amount B _{t of the} template). In general, N may be any number, but FIG. 2 shows an example of N = 12. Further, FIG. 3 shows an example of a template of a semicircle, a rectangle and a circle represented by the 12 unit vectors shown in FIG. In FIG. 3, the center of the drawing is the basic vector display for the template, the right side of the drawing is the vector feature example, and this vector feature example is the template expression actually used.
When the template is expressed by the unit vector, the same template can be used for similar shapes. Next, the vector feature quantity B _t is expressed by the following equation.

Bt = {b _t1 , b _t2 , ... B _tJt } b _tj
= (X _tjs , y _tjs , x _tje , y _tje ) = (start point, end point of vector).

Next, the matching method in step S3 will be described. After the contour extraction, the thinning, and the vectorization are performed on the input image, the vector feature amount A forming the input image is obtained as follows.

A = {a ₁ , a ₂ , ... a _I } a _i =
(X _is , y _is , x _ie , y _ie ) = (vector start point, end point).

Here, the vector feature quantity B of the template
FIG. 4 shows a matching matrix in which _t is the vertical axis and the vector feature amount A of the input image (input pattern) is the horizontal axis. In FIG. 4, the vector feature amount A of the input image (a)
The one-component vector a _{i of the above} and the one-component vector btj of the template have the similarity h (a _j , b _tj ) obtained by the following equation. In addition, in FIG. 4, (b) shows an example of vector features, and (c) shows a matching matrix.

[0025]

(Equation 2)

However, h (a _j , b _tj ) = d (a _j ,
b _tj ). This indicates that the start point of the vector feature amount of the input image (start point of the matrix) does not have to match the corresponding point of the template. Also, (a _i ,
b _ti ) represents the inner product of a _i and b _ti . That is, the vector a _i
The angle formed by and b _ti becomes the partial similarity.

FIG. 5 is an explanatory view showing an example of matching, h
The maximum value of (a _I , b _tj ) (j = 1, 2, ... Jt) is set to the optimum similarity St (t = t) between the input pattern and the template t.
1, 2, ..., T), which is the maximum template pattern as the recognition result.

Next, another embodiment of the present invention will be described. This embodiment is a two-dimensional image matching method using complex information. The basis of this method is to represent a template and an input pattern on a complex plane, and after extracting information necessary for recognition,
Based on the information, DP matching is performed to obtain a recognition result. The symbol j is used as the imaginary unit.

First, the flow chart of the embodiment will be described with reference to FIG. In FIG. 6, step S11 is an image input unit for inputting a two-dimensional image. The input image input by this image input unit is subjected to contour extraction, thinning and vectorization in step S12 in the same manner as in the above-mentioned embodiment, and the input image A complex correlation coefficient (feature amount) is obtained from the constructed complex vector as described later. Next, the degree of similarity between the complex correlation coefficient of the template represented by the N complex unit vectors in step S13 and the complex correlation coefficient obtained in step S12 is obtained on the matching matrix in step S14. Then, in the same manner as in the above-described embodiment, the template pattern having the maximum optimum similarity is set as the recognition result in step S15.

Next, the template of the above embodiment will be described with reference to FIG. FIG. 7 shows a template figure by N complex unit vectors (unit vectors on the complex plane) (vector feature amount B _{t of} template figure).
As a result, a complex unit vector of N = 12 is shown as an example. In general, N may be any number and is given by the following equation.

[0031]

(Equation 3)

FIG. 8 shows a semicircle, a rectangle and a circle represented by the 12 complex unit vectors described above. FIG.
Is the same as FIG. 3, but only the vector feature example is different.
It should be noted that this complex unit vector expression is invariant to the similar shape. Next, for these figures expressed as vectors for the complex, the complex correlation coefficient R _i (i = 1, 2,
.. N).

[0033]

(Equation 4)

Here, the line segment drawn above Q (i + 1) represents a conjugate complex number. A set obtained by adding the obtained sequence of correlation coefficients and a null (NULL) vector (zero vector 0 + j · 0) is used as a template (feature vector sequence B _t ) of the graphic (graphic number t).

B _t = {R _t1 , R _t2 , ..., R _tJt } =
{B _t1 , b _t2 , ..., B _tJt } This template is rotation-invariant for the following reasons. Rotation of the angle θ on the complex plane is equivalent to multiplying the complex vector by e ^P (P = jθ). That is, the rotated complex vector Q ′ (i) is e ^P · Q (i) (P = j
θ). Here, the term relating to rotation is canceled out from the following equation, and therefore the complex correlation coefficient becomes invariant with respect to rotation.

[0036]

(Equation 5)

Next, in the matching method, the contour extraction, thinning, and vectorization are applied to the input image as in the above embodiment, and the complex correlation coefficient (feature amount) A = from the complex vector forming the input image. {A ₁ , a ₂ , ..., a _I } are obtained.
Then, consider the matrix shown in FIG. 9 in which the vertical axis represents the feature amount B _{t of the} template and the horizontal axis represents the feature amount A of the input image. In FIG. 9, 1 of the feature amount A of the input image (a)
Component vector (complex vector) a _i and template t
The similarity h (a _i , b _ti ) between the 1-component vector (complex vector) b _{ti of} Note that, in FIG. 9, (b) shows an example of vector features, and (c) shows a matching matrix.

[0038]

(Equation 6)

However, h (a ₁ , b _tj ) = d (a ₁ ,
b _tj ). This indicates that the starting point of the vector feature of the input image (starting point of matching) does not have to match the corresponding point of the template. This means that no special algorithm for searching the starting point is needed. In the above formula, the molecule inside Re [] is a
∠A is the inner product of _i and b _{ti in} the complex vector space
Represents the phase of

FIG. 10 is an explanatory diagram showing an example of matching. In FIG. 10, the maximum value of h (a _I , b _tj ) (j = 1, 2, ... Jt) is set to the optimum similarity between the input pattern and the template t. St (t = 1, 2, ..., T), and the maximum template pattern is the recognition result.

[0041]

As described above, according to the present invention,
The following effects can be obtained.

(1) There is no need for a special function such as finding the point corresponding to the starting point of the template from the input pattern, and (2) once a template for a certain figure is created, the similar figure of that figure can be obtained. There is no need to create a template again. (3) By using a complex correlation coefficient as a feature quantity, matching that is invariable with respect to the rotation of the figure can be performed. (4) From the effects of (2) and (3) above The recognition time can be shortened by suppressing the increase in the number of templates.

[Brief description of the drawings]

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

FIG. 2 is an explanatory diagram of a unit vector forming a template.

FIG. 3 is an explanatory diagram of a template example.

FIG. 4 is an explanatory diagram of an example of a matching matrix.

FIG. 5 is an explanatory diagram of a matching example.

FIG. 6 is a flowchart showing another embodiment of the present invention.

FIG. 7 is an explanatory diagram of a complex unit vector forming a template figure.

FIG. 8 is an explanatory diagram of a template example.

FIG. 9 is an explanatory diagram of an example of a matching matrix.

FIG. 10 is an explanatory diagram of a matching example.

[Explanation of symbols]

S1 ... Image input step S2 ... Step of obtaining vector feature amount A for input image S3 ... Step of obtaining template represented by N unit vectors S4 ... Step of obtaining similarity S5 ... Step of making template pattern a recognition result

Claims (4)

[Claims] 1. A first vector feature amount forming an input pattern is obtained by performing contour extraction, thinning, and vectorization of an input pattern input from an image input unit, and then the first vector feature amount, The similarity between the second vector feature quantity of the template composed of N unit vectors in advance and the one-component vector of both feature quantities is calculated on the matching matrix, and the maximum value of the calculated similarity is the input pattern. And a template, and the pattern of the template having the maximum optimum similarity is used as the recognition result. 2. A two-dimensional structure characterized by expressing an input pattern input from an image input unit and a template on a complex plane, extracting information necessary for recognition, and then performing DP matching based on the information. Image matching processing method. 3. The contour extraction, thinning, and vectorization of the input pattern to obtain a first complex correlation coefficient from a complex vector forming the input pattern, and then the first complex correlation coefficient, The second complex correlation coefficient of the template composed of N complex unit vectors and the similarity with the one-component vector of both complex correlation coefficients are obtained on the matching matrix, and the maximum value of the obtained similarity is input. A two-dimensional image matching processing method, characterized in that an optimum similarity between a pattern and a template is set, and a pattern of a template having the maximum optimum similarity is set as a recognition result. 4. The two-dimensional image matching processing method, wherein the similarity is a partial similarity of an inner product of a template and an input pattern in a complex vector space on a matching matrix.