JPH08263655A - Pattern recognition method - Google Patents

Abstract

PURPOSE: To precisely recognize patterns when there is plural patterns similar in forms as reference patterns. CONSTITUTION: The difference of similarity between a first candidate having the highest similarity of the similarity obtained among a picture PN being a recognition object and all reference patterns B-F and a second candidate in the recognition processing of a first stage. When the difference is large, the pattern of the reference picture of the first candidate is set to be a resolution. When the difference of similarity is small, inter-category distribution between the first candidate extracted from the set of the patterns which are previously decided and which are easily mistaken and the reference pictures of all the patterns B, C and D is obtained in the recognition processing of a second stage. The new pictures B'-D' of the reference picture and the pictures being the recognition objects are generated by data on an axis whose value is large. Then, the pattern of the reference picture which is obtained in the new picture and which has the highest similarity is set to be the resolution. Thus, the pattern can highly precisely be recognized since it is recognized again in the second stage when it is uncertain in the first stage.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method from a multivalued image.

[0002]

2. Description of the Related Art A conventional pattern recognition method is, for example, a normalized correlation method. This is, for example, in FIG.
When comparing the recognition target pattern P composed of a multi-valued image as shown in (b) with the reference pattern S, the brightness value data of each dot in the image of each pattern is applied to the equation (1) to recognize the recognition target pattern. The similarity is obtained by obtaining the normalized correlation between the reference pattern and the reference pattern.

[Equation 1] Here, I (i, j): luminance value of point (i, j) of the recognition target image, T (i, j): luminance value of point (i, j) of the reference image,

[Outside 1] (M, n): sizes of the reference image and the recognition target image. An image representing the recognition target pattern will be referred to as a recognition target image, and an image representing the reference pattern will be referred to as a reference image.

The value of the normalized correlation obtained by the above equation (1) has a value of -1 to +1. This value becomes larger as the degree of similarity becomes higher, and becomes the maximum value 1 when all the normalized dot values are the same. That is, since the image is normalized, the image can be recognized without being affected by the change in the overall brightness of the image. Generally, as shown in (c) of FIG. 9, when the overlapping portion of the reference pattern and the pattern on the recognition target pattern is large, the similarity is determined to be high. The normalized correlation method is
In this way, it is used to find a pattern having a shape similar to the reference pattern.

[0004]

However, in such a conventional pattern recognition method, when the same pattern as the recognition target pattern is obtained from a plurality of reference patterns, the plurality of reference patterns to be compared are compared with each other. If patterns with similar shapes are included, the discrimination judgment between the patterns with similar shapes is uncertain, and the similarity to the wrong reference image is higher due to the influence of illumination when capturing the recognition target pattern. There is a problem that it often happens. This is because the degrees of similarity with a plurality of reference images having patterns with similar shapes are very close to each other.

Therefore, in view of the above conventional problems, the present illumination aims to provide a pattern recognition method capable of accurately recognizing a plurality of patterns having similar shapes as reference patterns. To do.

[0006]

[Means for Solving the Problem] As a result of calculating the degrees of similarity with all the reference patterns in this way, if there is another pattern having a degree of similarity very close to that pattern, it is Since there are other patterns with similar shapes, there is a possibility that they are erroneously recognized. Therefore, if the data is re-recognized using the data that remarkably expresses each feature between the similar patterns, the recognition is surely performed.

Therefore, the present invention according to claim 1 is a pattern recognition method for comparing a recognition target pattern of a multi-valued image with a plurality of reference patterns, and the degree of similarity between the recognition target image and all the reference images. Of the first candidate and the second candidate in descending order of similarity, and when the difference in similarity between the first candidate and the second candidate is greater than or equal to a predetermined threshold value, When the difference between the similarities is less than the predetermined threshold value, the first-stage recognition processing in which the pattern of the reference image of one candidate is used as a solution, and the reference images of the first candidate and the second candidate and the recognition target image First, using a new image
The second step of recognition processing is performed by obtaining the similarity between the candidate and the second candidate and using the pattern of the reference image having the highest similarity among them as the solution.

The above-described second-stage recognition processing includes the first step of obtaining A = inter-category variance / total variance for each feature axis of the reference images of the first candidate and the second candidate, and the first candidate and the first candidate. 2 The second step of creating a new image composed of data on the axis having a large A value of the reference image of the candidate and the recognition target image, and the similarity between the newly created reference image and the recognition target image And a third step of determining the degree.

Further, the invention described in claim 3 obtains the similarity between the recognition target image and all the reference images, and the learning process for defining a set of patterns that are likely to be mistaken in a plurality of reference patterns. The reference images of the first candidate and the second candidate are obtained in descending order of similarity, and the first candidate and the second candidate are obtained.
When the difference between the similarities of the candidates is equal to or larger than a predetermined threshold value, the recognition process of the first step in which the pattern of the reference image of the first candidate is used as a solution, and the difference between the similarity degrees are less than the predetermined threshold value When
Similarity of the reference image of the pattern defined to be easily mistaken for the first candidate pattern using the reference image of the pattern defined to be easily mistaken for the first candidate pattern and the new image of the recognition target image in the set The pattern recognition method comprises a second step of recognition processing in which the degree of similarity is determined and the pattern of the reference image having the highest degree of similarity is used as the solution.

In the above-described second-stage recognition processing, A = inter-category variance / total variance for each feature axis of all reference images of a pattern defined to be easily mistaken for the first candidate pattern is calculated. A step, a second step of creating a new image composed of all the reference images according to the definition and data on the axis of the recognition target image having a large value of A, and the newly created reference image It is preferable to include a third step of obtaining the similarity with the recognition target image.

[0011]

According to the pattern recognition method of claim 1, among the similarities obtained between the recognition target image and all the reference images in the first-stage recognition processing, the first candidate having the highest similarity is selected. The difference between the second highest similarities of the second candidates is obtained, and when the difference is equal to or larger than the threshold value, the pattern of the reference image of the first candidate is determined as a solution with high certainty of recognition. When the difference between the similarities is less than the threshold value, the second stage recognition process is performed, a new image is created for each candidate and the recognition target, and the new image is used to generate the first candidate and the second candidate. The similarity is calculated, and the pattern of the reference image having the highest similarity is set as the solution. As a result, even if there are a plurality of reference patterns having similar shapes, they can be recognized with high accuracy.

In the pattern recognition method according to a third aspect of the present invention, in the first-stage recognition processing, a set of patterns that are likely to be mistaken among a plurality of reference patterns is defined as preprocessing. Next, of the similarities obtained between the recognition target image and all the reference images, the difference between the first candidate having the highest similarity and the second candidate having the next highest similarity is calculated, and the difference is calculated. If it is equal to or higher than the threshold value, the pattern of the first candidate reference image is set as a solution because the certainty of recognition is high.

When the difference between the similarities is less than the threshold value, the second-stage recognition processing is performed, and the reference image of the pattern which is easily confused with the pattern of the first candidate is extracted from the set determined in the preprocessing, A new image is created for these reference image and recognition target image, the similarity of the reference image of each pattern that is easily mistaken for the pattern of the first candidate is calculated using this new image, and the highest similarity among them is obtained. Let the pattern of the reference image be the solution. Since the patterns that are likely to be mistaken in a plurality of reference patterns are set as a set by learning in advance, all the reference patterns of the set can be re-recognized in the second-stage recognition processing.
The certainty of recognition is high when there are a plurality of similar-shaped reference patterns.

In each of the second-stage recognition processes described above, A = inter-category variance / total variance for each feature axis of the narrowed-down reference image is obtained, and the A of the reference image and the recognition target image is obtained.
By constructing the new image with the data on the axis having a large value of, the different characteristic portions of each pattern are conspicuously displayed even between patterns that are apt to be mistaken, and the identification is easy.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a flow chart showing the outline of the first embodiment of the present invention. In this embodiment, first, as preprocessing, in step S21, a table defining a set of patterns having shapes that are easily mistaken in the reference pattern is created. Hereinafter, this table is referred to as "similar pattern table" Z. And after this,
The recognition process starts from step S22. The recognition process is
It consists of a first stage recognition process and a second stage recognition process.

The process of the first stage comprises steps S22 and S23 for obtaining the first candidate pattern. Step S2
In No. 2, the similarity between the recognition target image and all the reference images is calculated using the normalized correlation method to perform matching. Among them, the pattern with the highest similarity and the pattern with the next highest similarity are calculated. Are determined as the first candidate and the second candidate, respectively. In step S23, the difference between the similarity with the first candidate and the similarity with the second candidate is obtained, and the certainty of the recognition result is determined based on the difference in the similarity. When it is determined that the certainty of the pattern of the first candidate is correct, the reference pattern of the first candidate is set as the solution of the recognition target image.

On the other hand, if the certainty that the first candidate pattern is the correct answer is low, the process proceeds to step S24 as the second stage recognition process. In step S24, the previous step S
Based on the "similar pattern table" Z created in step 21, a new reference image and recognition target image are created by converting and extracting the features of only the first candidate pattern and the reference pattern that is easily mistaken. Re-recognize using. Then, as a result of this recognition processing, the reference pattern having the highest similarity is set as the solution of the recognition target image.

The details of each of the above processes will be described below. FIG. 2 is a flow chart showing the procedure for creating the “similar pattern table” Z in step S21. Here, first, in step S211, for each of the M types of reference patterns, N known learning sample images SG each representing the pattern are prepared. It is assumed that these learning sample images SG include some noises and shadows and are not completely the same within each type.

Next, learning is performed using this sample image SG. Prepare a "voting table" ZH for learning. The “voting table” ZH has M × M columns corresponding to the types of the reference pattern in each of the vertical and horizontal directions, and the value v of each column
all (x, y) is set to 0 at initialization.

Here, the case where a reference pattern which is easily mistaken for the pattern B is defined is taken as an example. First, step S2
At 12, the N images representing the pattern B are used as recognition targets, and the similarity between one of them and the M types of reference images is calculated. The normalized correlation equation of the equation (1) is used to calculate the similarity. Then, in step S213, it is checked whether or not the result of this calculation is the correct answer.

If the result of this operation is a correct answer, that is, if the pattern with the highest degree of similarity among the M pieces is the reference pattern B, the degree of similarity with the degree of similarity with the reference pattern B is determined in step S214. Search for other reference patterns with a difference less than a predetermined threshold in between.

If there is a reference pattern X satisfying the condition, in step S216, the point on the "voting table" ZH is at the position of B in the vertical direction and X in the horizontal direction is the reference pattern X, which is an appropriate constant defined in advance. Add K. That is, when the reference pattern X satisfying the expression (2) exists, the operation of applying the expression (3) to the “voting table” ZH is performed. (Similarity with correct reference pattern B)-(Similarity with reference pattern X) <Threshold value (2) value (B, X) = value (B, X) + K (3)

When there are a plurality of reference patterns X whose difference in similarity is less than the threshold value, the values of the positions of the reference patterns on the "voting table" ZH are similarly set in all the reference patterns. Add the score K. For example, when the threshold value is 0.2, the similarity with the reference pattern B is 0.8,
When the similarity with the reference patterns D, E, F, O, and R is 0.4, 0.7, 0.5, 0.2, and 0.65, respectively, the difference in similarity is 0.8-. E with 0.7 = 0.l <0.2 and R with 0.8-0.65 = 0.15 <0.2
Assuming that the degree of similarity is high, points are added to the positions of F and R where B is vertical and X is horizontal in the "voting table" ZH.

On the other hand, if the check result in step 213 is not correct, that is, if the pattern with the highest degree of similarity is not the reference pattern B, the reference pattern with the highest degree of similarity is extracted in step S215. Then, the process proceeds to step 216, and the point is added to the position Y of the reference pattern in which the vertical on the "voting table" ZH is B and the horizontal is the highest in similarity. For example, the similarity with the reference pattern B is 0.6, the similarity with the reference patterns D, E, F, O, and R is 0.4, 0.7, 0.5, 0.2, respectively.
When it was 0.65, the vertical direction of the "voting table" ZH was B.
Then, the score is added to the position of E having the highest degree of similarity in the horizontal direction.

The same processing is performed on N images, and a reference pattern having a score equal to or higher than a predetermined threshold value on the "voting table" ZH after the end is set as "a pattern which is easily mistaken for pattern B". Define. The above operation is repeated for all M types of reference patterns, and a "similar pattern table" Z is created in step S217. That is, the "pattern that is easily mistaken for the pattern B" defined in step S216 defines the row B of the "similar pattern table" Z, and each row is similarly defined for the other patterns.

Next, FIG. 3 shows details of the recognition process in the first stage. First, in step S221, the similarities between the recognition target image indicating the recognition target pattern and all the M reference images indicating the reference pattern are obtained using the normalized correlation method according to the above-described equation (1). Then, step S22
2, the pattern with the highest similarity (first candidate) and the pattern with the second highest similarity (second
Candidates).

In step S223, the difference between the similarity of the first candidate and the similarity of the second candidate is obtained, and it is checked whether or not the difference is less than a predetermined threshold value set in advance. When the difference in similarity is equal to or more than the threshold value, the reliability of the recognition result is high, and the pattern of the first candidate is set as the solution at that time. On the other hand, when the difference between the similarities is less than the threshold value, the reliability of the recognition result is low, and the process proceeds to the second step for re-recognition. Step S above
221 and S222 correspond to step S22 in FIG. 1, and step S223 corresponds to step S23.

FIG. 4 shows the second step in the previous step S24.
The details of the process of steps are shown. First, in step S241, a pattern that is likely to be mistaken for the first candidate pattern is extracted from the "similar pattern table" Z created in the learning process.
Hereinafter, this set of patterns is referred to as a “similar pattern set”. For example, when the first candidate pattern is B, the “similar pattern set” is defined as {B, E, R} based on the “similar pattern table” Z shown in FIG. Then, in step S242, from the sample image SG, the above B, E, R
Each of the N images belonging to each category is selected.

Next, in step S243, the degree of similarity with the selected pattern belonging to the "similar pattern set" is obtained. Here, first, a value A represented by the following equation is obtained for each feature axis between the reference images representing the reference patterns in the “similar pattern set”. A = inter-category variance / total variance (4) where inter-category variance = luminance value of dots at the same position in an image representing the number of similar patterns created from the average value of N images representing similar patterns Variance is the variance of the brightness values of the dots at the same position in N images for all patterns in the “similar pattern set”. FIG. 5 illustrates data used in the above calculation when the similar pattern sets are B, E, and R. In equation (5), similar pattern sets are B, E, and R.
It is a method of obtaining A of the point (i, j) at the time. Using this equation, A is calculated for all dots on the reference image.

[Equation 2]

Returning to FIG. 4, after this, step S244.
Then, since the data on the feature axis having a large value of “A” is the data that remarkably represents the features of the patterns in this set, the data on the feature axis having a value of “A” larger than the predetermined threshold value. A new reference image and a recognition target image composed only of the data of are created.

FIG. 6 is a diagram showing a method of creating this new image. Here, it is assumed that the size (horizontal × vertical) of the reference image is 4 × 3 and a new image composed of six pieces of data is created. First, "A" of each dot is obtained from the above-mentioned equation (2), and the positions of the top six places where the value of "A" is large are obtained. In Figure 6, the top 6 positions are (0,0),
(3,0), (1,1), (2,1), (3,1),
In (3, 2), the luminance value data at these positions is used to convert the horizontal 6 image. This conversion process is performed on all the reference images and recognition target images in the similar pattern, and a new image composed of the brightness at the six positions is created.

Returning to FIG. 4, in step 245, the similarity between the recognition target pattern and the reference pattern is obtained by the normalized correlation method of the equation (1) using this newly created image. Then, as a result of this processing, the reference pattern having the highest similarity is set as the solution of the recognition target image.

Next, the operation of the above recognition method will be described. Each size (horizontal x vertical) 3 as shown in Figure 7
A recognition process for searching for a recognition target pattern represented by a recognition target image PN in the lower right part of the figure from five reference patterns of patterns B, C, D, E, and F of × 5 will be described as an example. The pattern of the recognition target image PN has a low degree of similarity to the patterns E and F because it has a large difference in shape, but the patterns B, C, and D have similar shapes, and therefore, the degree of similarity with them is all. Get higher That is, when the similarity with all five patterns is calculated in the recognition process of the first stage, the difference between the similarity of the first candidate and the similarity of the second candidate becomes small. Therefore, the recognition process of the second stage is performed.

In the second-stage recognition processing, three patterns B, C, and D having a high degree of similarity in the first-stage processing are used as reference images. When the inter-category variance of each dot of these three images is obtained, the values at the positions of points (4, 0) and points (2, 1) showing different values for each pattern become large. Therefore, by using the data of these two points where the value of “A” is large, the patterns B, C and D are the patterns B ′, C ′,
Converted to D '. Patterns B ', C', and D'are pixels that remarkably represent different features between patterns B, C, and D. Therefore, this pattern B ', C',
When the similarity with the new recognition target image P'created with the dot value at the same position as D'is obtained, an extremely high similarity is obtained with the pattern C ', and a highly reliable recognition result is obtained. To be

As described above, in the present embodiment, when a pattern similar to the recognition target pattern is obtained from the plurality of reference patterns, first, candidates are obtained from the degree of similarity with all the reference images, and the matching result is obtained. In the case of uncertainty, a new image is created and re-recognized with respect to a reference image having a similar pattern, so that it is possible to accurately recognize even when there are a plurality of reference patterns having similar shapes. .

In particular, for a new image for re-recognition, based on a set of similar patterns defined in advance,
Since only the reference images of similar patterns included in the set are used to obtain the inter-category variance / total variance for each dot and to create an image composed of only the dots at positions with large values, When recognizing between patterns having similar shapes, it is possible to perform recognition using image data in which different portions between them are conspicuous, and it is possible to reliably perform recognition between similar patterns.

FIG. 8 is a flow chart showing the configuration of the second embodiment of the present invention. The recognition process includes the first stage process and the second stage process as in the previous embodiment. The processing of the first stage is performed by steps S51 and S5 for obtaining the first candidate pattern.
It consists of two. First, in step S51, the similarity between the recognition target image and all the reference images is calculated using the normalized correlation method, and the pattern with the highest similarity among them and the pattern with the second highest similarity are obtained. , And the first and second candidates, respectively.

In step S52, the difference between the similarity with the first candidate and the similarity with the second candidate is obtained. Then, if the difference between the similarities is equal to or larger than a predetermined threshold value set in advance, the reference pattern of the first candidate is set as the solution of the recognition target image. Steps S51 and S52 are the same as steps S22 and S23 of FIG. 1 in the previous embodiment.

On the other hand, when the difference between the similarities is less than the above threshold value, the process proceeds to step S53 as the second stage recognition process. In step S53, the two reference patterns that are the first candidate and the second candidate are re-recognized. That is, the first candidate and the second candidate are used instead of the "similar pattern set" defined in the previous embodiment. The re-recognition process is the same as that of the previous embodiment, that is, considering the inter-category variance for each dot, new reference images and recognition target images of the first candidate and the second candidate are created, and this new image is generated. The similarity between the two is obtained, and the one with the higher similarity is used as the recognition result.

Since this embodiment is configured as described above, when the reference pattern does not include a similar pattern or when the number of similar patterns is less than 3, the "similar pattern table" may not be created. Even if it can be created, it is also effective when it is not possible to easily draw out the features in the “similar pattern set” to create a new reference image. As in the previous embodiment, a plurality of reference patterns having similar shapes can be obtained. Even if there is, it can be recognized accurately.

[0041]

As described above, according to the present invention, in a pattern recognition method for comparing a recognition target pattern of a multivalued image with a plurality of reference patterns, the degree of similarity obtained between the recognition target image and all the reference images. First among the most similar
If the difference between the second candidate and the second highest similarity is found, and the difference is equal to or more than a threshold value, the recognition process of the first stage using the pattern of the reference image of the first candidate as the solution and the difference in the similarity. When is less than the threshold value, a new image is created for each candidate and the recognition target, the similarity between the first candidate and the second candidate is calculated using this new image, and the reference image with the highest similarity is calculated. Since the second step of recognition processing using a pattern as a solution is adopted, it is possible to accurately recognize even when there are a plurality of reference patterns having similar shapes.

In addition, before the first-stage recognition processing, a learning process for determining a set of easily misleading patterns in a plurality of reference patterns is performed as a pre-processing, and a new image in the second-stage recognition processing is , The first candidate pattern extracted from the set defined in the pre-processing and the reference image and the recognition target image of a pattern that is easily confused with each other are created, and the similarity among the similarities obtained using this new image is the highest. By using the pattern of the reference image with a high degree as a solution, all the patterns that are likely to be mistaken can be re-recognized, so that there is higher certainty in recognition when there are multiple reference patterns with similar shapes. Is obtained.

Furthermore, a new image in the second-stage recognition processing is obtained by A = inter-category variance / total variance for each feature axis of the reference images of the first candidate and the second candidate, or of the pattern that is easily mistaken. By configuring the reference image and the image to be recognized with the data on the axis where the value of A is large, different characteristic portions of each pattern are prominently displayed, and the identification is facilitated.

[Brief description of drawings]

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

FIG. 2 is a flowchart showing how to create a “similar pattern table”.

FIG. 3 is a diagram showing details of a first-stage recognition process.

FIG. 4 is a diagram showing details of second-stage recognition processing.

FIG. 5 is an explanatory diagram showing pattern categories when obtaining the inter-category variance of reference images.

FIG. 6 is a diagram showing a method of creating a new image.

FIG. 7 is an explanatory diagram showing an example of recognition by a new image.

FIG. 8 is a flowchart showing the configuration of the second exemplary embodiment.

FIG. 9 is a diagram showing an example of images of a recognition target pattern and a reference pattern.

[Explanation of symbols]

 P recognition target pattern S reference pattern SG sample image Z "similar pattern table" ZH "voting table"

Claims (4)

[Claims] 1. A pattern recognition method for comparing a recognition target pattern of a multi-valued image with a plurality of reference patterns, wherein the similarity between the recognition target image and all reference images is calculated, and the highest similarity among them is obtained. First, reference images of the first candidate and the second candidate are obtained in order, and when the difference between the similarities of the first candidate and the second candidate is equal to or more than a predetermined threshold value, the pattern of the reference image of the first candidate is set as a solution. When the difference between the similarities is less than the predetermined threshold value in the first-stage recognition process, the first candidate and the second candidate reference images and the new image of the recognition target image are used to detect the first candidate and the second candidate. A pattern recognition method, comprising: a second-stage recognition process in which a similarity of a second candidate is obtained, and a pattern of a reference image having the highest similarity among them is used as a solution. 2. The recognition process in the second step comprises a first step of obtaining A = inter-category variance / total variance for each feature axis of the reference images of the first candidate and the second candidate, and the first step. Second step of creating a new image composed of the reference image of the candidate and the second candidate and the data on the axis of the recognition target image where the value of A is large, and the newly created reference image and recognition target image And a third step of obtaining the degree of similarity with the pattern recognition method. 3. A pattern recognition method for comparing a recognition target pattern of a multi-valued image with a plurality of reference patterns, wherein a learning process for defining a set of easily erroneous patterns in the plurality of reference patterns and all the recognition target images Of the reference image of the first candidate and the second candidate in order from the highest similarity among them, and the difference in similarity between the first candidate and the second candidate is predetermined. When the threshold value is equal to or more than the threshold value, the first-stage recognition processing in which the pattern of the reference image of the first candidate is used as a solution, and when the difference between the similarities is less than the predetermined threshold value, the first candidate of the set is selected. Using the reference image of the pattern defined to be easily mistaken for the pattern and the new image of the recognition target image, the similarity of the reference image of the pattern defined to be easily mistaken for the pattern of the first candidate is obtained, and Most Also includes a second-stage recognition process in which a pattern of a reference image having a high degree of similarity is used as a solution. 4. The second step of recognition processing is to obtain A = inter-category variance / total variance for each feature axis of all reference images of a pattern defined to be easily mistaken for the first candidate pattern. 1 step, the second step of creating a new image composed of all the reference images according to the definition and the data on the axis having the large value of A of the recognition target image, and the newly created reference A pattern recognition method comprising a third step of obtaining a similarity between an image and an image to be recognized.