JPH05135178A - Pattern recognition method - Google Patents

Abstract

PURPOSE:To make a reliable judgement even if a given figure is deformed and included in an image by determining weight regarding the pattern obtained by giving the given pattern deformation to be considered, finding the adaptability image and adding the product of the both, and judging whether or not the result is larger than a given reference value. CONSTITUTION:The adaptability to an image A of the pattern B given in specific data form when the reference point of the pattern B is put on coordinates on the image A is calculated (30). The pattern B or A is deformed and the weight between the pattern B' or A' after the deformation and the source pattern B and A is determined (40). While the weight of the similar pattern B' of the pattern B and the similar pattern B' of the pattern B or the weight of the similar image A' of the image A and the similar image A' of the image A is found, the adaptability between the similar pattern B' and image A or similar image A' and pattern B is found and the weighted sum of the adaption by the weight is found (60). When the maximum value obtained while the coordinates on the image A are moved exceeds the reference value, the pattern B is in the figure A.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a figure recognition method, and more particularly to a figure recognition method suitable for recognizing a figure expected to be present in an image.

[0002]

2. Description of the Related Art In a conventional figure recognition method, in order to recognize a figure expected to be present in the image from an image, (1) the figure is considered as a template, and while moving in the image, When the size is different, the size is changed, and when there is a possibility that the shape is deformed, the shape is considered and the degree of coincidence is checked. If the figure is equal to or more than a predetermined criterion, it is assumed that the figure is present.
A method of calculating the feature amount of the image and the expected figure, comparing the feature amounts with each other, and determining whether or not the figure exists (3) A method of performing both (1) and (2) above Is being done.

What is common to both has a calculated value (degree of conformity) based on a criterion for judging the degree of coincidence,
If there is a figure having the optimum value, it is determined that there is an expected figure, and if there is no or a small value, it is determined that there is no expected figure. Further, as a method for dealing with the case where noise is superimposed on the image or when distortion is included, a smoothing operator such as a Gaussian operator is applied to one or both of the image and the expected figure,
After that, there is a method of taking the method of (1) or (2) or (3).

[0004]

In the prior art, since the calculation for judgment is made using only the expected figure, the figure is deformed even though the figure is similar to the original figure. In this case, there is a problem that the degree of conformance does not reach a predetermined value and it is overlooked. On the other hand, the method using the smoothing operator has a problem that the sharpness of the feature amount is lost, the degree of coincidence becomes low, and at the same time, the number of figures having the same degree of coincidence increases.

[0005]

The above object is achieved by calculating the goodness of fit, including possible deformations of expected figures. That is, when calculating the goodness of fit of a figure in a certain position in an image with the expected figure, try to shift the position of the figure to be judged variously in the vicinity, and The possible deformations are variously compared, and the degree of shifting the position of the figure and / or the degree of deformation of the figure becomes smaller, the smaller the degree of shifting the position of the figure and / or the deformation of the figure. With a weighting function that increases as the degree of degree decreases, the fitness degree is calculated in the same way as the conventional method, and whether or not there is an expected figure in the image is calculated based on the weighted average value of the weights with the corresponding weight. Make a decision. Here, the expected deformation of the figure can be obtained even if the figure to be judged of the image A is deformed.
For simplicity of explanation, only the case of expected deformation of the figure will be described.

[0006]

Consider the case where the expected deformation of the figure B in the image A is stochastically performed. Even if the transformation is not done stochastically, if a large number of elements are intertwined and cause the transformation like in the real world,
Since probabilistic approximation is possible, it can be considered in the same way as the following.

If the probability that the figure B in the image A is transformed into a figure B'similar to it is p (B '| B), the figure B'
The probability that an image A containing a is obtained is p (A | B ') p
It can be written as (B '| B) p (B). According to the calculation of the goodness of fit considering the displacement of the figure to be judged on the image and the expected deformation of the figure, this probability is integrated with respect to B ′, and approximately the weighted sum is obtained. Obtained as. That is, the posterior probability that the image A is obtained is calculated under the assumption that the figure B exists, and B is maximized.
Will be seeking. That is, the maximum posterior probability solution is obtained in consideration of position change and deformation.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Since the present invention can be implemented by a general data processing device, that is, a so-called computer comprising an input / output device, a memory and a central processing unit, description of the device as an embodiment is omitted. In the following explanation, the image data is
For example, it is introduced into a data processing device as a serial signal through an image pickup device, and a constant lightness is used as a reference value and the lightness is 0
Alternatively, the data having a brightness of 1 is stored in the memory and processed. Although the figure data may be similarly input, it may be given in advance in the form of a dictionary for character recognition or the like.

An embodiment will be described below with reference to FIGS. 1 and 2. Although FIG. 1 and FIG. 2 are the same in substance, FIG. 1 shows the overall concept as a PAD diagram, whereas FIG. 2 shows a PAD diagram mainly for the form of a computer program. In the present embodiment, two types of lightness 0 and 1 are used.
A value image, where the ground part has a brightness of 0 and the brightness near the center is 1
Consider the figure recognition for judging the shape of the figure and the presence / absence of the figure from an image in which there is a figure part and noise that is inverted at a certain frequency is superimposed.

The image A has n * m pixels arranged in a square lattice. The expected figure B is assumed to have p * q pixels arranged in a square lattice. The coordinate origin of the pixel of the image A is the upper left, the x coordinate is to the right, and the y is to the bottom.
Take the coordinates. That is, the upper left is (0,0), the right side is (1,0), and the right side is (2,0). Below the origin is (0,1) and below that is (0,2). Similarly to the image A, the reference point of the figure B is at the upper left. The coordinates on the figure are taken in the same way as the image.

When the reference point of the figure B is located at (x, y) on the image A, the matching degree between the image A and the figure B is shown.

[0012]

[Equation 1]

Specified by. However, δ (u, v) is a function that becomes 1 when u = v and 0 otherwise. This is a value for all pixels where the image A and the figure B overlap, and δ = 1 if the pixel values (brightness) are the same between the image A and the figure B, and δ = 0 if they are different. Is the sum of This is represented by f (A, B, x, y). The transformation is
Consider that one or a plurality of pixel values at the boundary between the brightness 0 and the brightness 1 in the figure B are inverted. The weight of r inversions is set to exp (-r).

Pixels on the boundary between lightness 0 and lightness 1 in figure B are extracted and their coordinates are arranged in array b (step 10 in FIG. 2).
00). Let the number be l. That is, the coordinates are stored in b (0) to b (l-1). The reference point of the figure B is moved within a possible range on the image A (step 10 in FIG. 1; step 1010 in FIG. 2; hereinafter, for simplification, illustration of numbers and steps are omitted). The degree of conformity between the image A and the figure B in this arrangement is calculated (30; 1030). The figure B is transformed over all the transformation methods to be considered (4
0; 1040). For a method of sequentially selecting r pieces to be reversed in this procedure, see, for example, CAMC, 5 (1962), page 344 (CACM, Vol.
5 (1962) p344) can be used. Of the pixels of figure B, the figure obtained by inverting the pixels of the r selected coordinates is B '. S
The figure B ′ obtained next is referred to as B ′ (s). In this way, the goodness of fit is calculated for all the figures B ′ obtained by deforming the figure B, and the product is multiplied with the weights to obtain the sum (60;
1060). The obtained value (meaning the value as the goodness of fit,
Display with c. ) Together with the position of the reference point (7
0; 1060).

When the above calculation is completed at all possible positions of the reference point, the maximum value among them is taken out (80; 108).
0). If the maximum value is equal to or larger than the predetermined value, it is determined that the graphic B is included in the image A (90; 1090).

Next, step 40-9 of the above embodiment.
An embodiment in which the 0 part is modified is shown in the PAD diagram of FIG. Also in this case, the same problem as in the previous embodiment is considered. However, figure B
Is composed of n * m pixels and has the same size as the image A, and has an occurrence probability equal to the stationary probability of the Markov random field shown by the following equation.

[0017]

[Equation 2]

Below, Markov random fields, probabilities, energy terms, and steps 2040, 205 of FIG.
No. 0, 2052, 2054, 2056, the validity of the method of determining the transition probability is described in IEE, PMI-6 (1984), pages 721 to 741.
Page (IEEE, vol. PAMI-6 (1984), p.
p. 721-741).

At this time, the energy function giving the stationary probability is

[0020]

[Equation 3]

[0021] Also,

[0022]

[Equation 4]

It is abbreviated as

A graphic B ', which is a modified graphic of the graphic B, is randomly generated using a random number or a pseudo random number (203).
0, 2040, 2050). First, the coordinates of points to be changed in the figure B are randomly selected (2030). Then i,
Whether to set the lightness at j to 0 or 1
Randomly determined (2040, 2050, 2052,
2054, 2056). The graphic thus configured is one deformed graphic. The degree of matching between this figure and the image A is obtained in the same manner as in the first embodiment (2060, 2062). The weight corresponding to the figure B'is set to exp (-r) where r is the number of different things in the pixels of the figure B and the figure B '. C is updated using both of these (2060, 206
2). Based on the figure B'obtained here, the following transformation (203
0 to 2062).

The above is repeated a predetermined number of times (2020). c
Since the value of is generally proportional to the number of iterations, the obtained c
Is divided by the number of repetitions (2070). If the value c as a result of the division is equal to or larger than a predetermined value, it is determined that the figure B exists (2090).

[0026]

As described above, according to the present invention, the matching degree between a given figure and a given image is calculated in consideration of the matching degree of the transformed figure that can be considered in the given figure. Then, even when the given image is included in the given image, it is possible to detect with high reliability that the given figure is included. As described above, this modification of the figure can be similarly performed as a modification of the image, and the same result can be obtained.

[Brief description of drawings]

FIG. 1 is a PAD diagram showing the overall concept of an embodiment.

FIG. 2 is a PAD diagram mainly showing an embodiment of a computer program form.

FIG. 3 is a PAD diagram mainly showing a computer program form in which a part of the embodiment is modified.

Claims (1)

[Claims] 1. A graphic recognition method for checking whether or not a graphic B or a graphic similar to the graphic B is included in the image A, in which coordinates on the image A given by data of a predetermined shape are set. Calculating the degree of conformity of the graphic B to the image A when the reference points of the graphic B given by the data of the predetermined shape are superposed, and
Alternatively, the image A is deformed, and the weight is determined such that the higher the degree of similarity between the changed figure B ′ or image A ′ and the original figure B or image A ′ is, the smaller the lower the degree of similarity is. , The weight of the similar pattern B ′ of the pattern B and the similar pattern B ′ of the pattern B or the image A while changing the reference point.
Of the similar image A ′ and the weight of the image A to the similar image A ′, the degree of matching between the similar pattern B ′ and the image A or the degree of matching between the similar image A ′ and the pattern B is calculated, and Finding the weighted sum of the goodness-of-fit by weighting the image A
While moving the coordinates on the point, repeat the above calculation using that point as a reference point to find the sum of the values obtained, and find the maximum value or an approximate maximum value, which is the predetermined reference value. If it exceeds, it is judged that the figure B is in the image A, and the figure recognition method.