JPH0737095A - Line figure element learning recognition method - Google Patents

Abstract

PURPOSE:To automatically recognize the entire graphic from the element pieces of the graphic regardless of the position fluctuation, rotation, magnification and reduction of the graphic by converting graphic information plotted in a drawing in to information handlable with a computer. CONSTITUTION:The information for which curvature information for accurately expressing a graphic shape is quantized by an adaptive vector quantizer 20, is supplied as input information to a learning vector quantizer 60. Then, by imparting the category information of the graphic as teacher information, the graphic is learned. Also, by imparting the curvature information of the new graphic to the learning vector quantizer 60 which has completed learning as the input information, the vector quantizer 60 is made to perform the categorizing (classification processing/recognition processing) of the new graphic. Thus, the entire graphic can be recognized from the graphic element pieces not being influenced by the change of a size, a position and the rotation just by showing the graphic to a system (to make it learn) and thus, graphic recognition programming like a conventional technique is unnecessitated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a line figure segment learning and recognition method, and in particular, in a drawing information system for designing, maintaining, managing, etc., information on a drawing, The present invention relates to a line pattern element learning recognition method in a pattern recognition method that automatically recognizes graphic information in a drawing as a procedure of converting it into information that can be handled by a computer, that is, acquiring an initial graphic from the drawing.

[0002]

2. Description of the Related Art As a method for obtaining graphic information in a drawing by a computer, a typical conventional method for converting the information into manageable information is to stick the drawing on a large tablet and use a cursor to draw a line segment that composes the graphic. There is a method of manually extracting each of the above. This method requires a lot of time, so
As another method, a method of automatically recognizing a figure in a drawing has been developed in order to automatically acquire figure information.
A typical conventional method for automatically recognizing a figure in this drawing is a software (a figure recognizing program) for classifying the figure type after extracting characteristic information of the target figure for each type of figure. It is created and the figure in the drawing is recognized by using the figure recognition program.

On the other hand, in recent years, techniques related to learning functions such as neural networks and learning vector quantization have been developed, and development of techniques to be used for recognition of character patterns and voice signals is also in progress.

The learning vector quantizer automatically learns the correspondence between the input value and the category approximately statistically by giving the input information and the category information corresponding thereto as teacher information, and theoretically Is in the limit,
It has been shown that the statistical error is minimal.

[0005]

However, in the conventional method by the above-mentioned figure recognition program (procedural programming), it is usually necessary to program for each figure type to be handled, and as the number of figure types to handle increases. Also, the more complicated the structure of the figure becomes, the more programming work is required.
The development of the software generally requires a great deal of development cost and time, which is a big problem. Therefore, this conventional method has a problem that the shape of a graphic that can be handled is limited to a simple shape, and it is difficult to apply it to automatic acquisition of general graphic information of a drawing.

[0006] Further, some methods have been developed for recognizing figures by automatically learning using a neural network or the like. With this method, the figure learned and the figure to be recognized are It cannot be applied when the size is different (enlarged / reduced),
If the figure to be recognized is rotated, it cannot be applied, or cannot be recognized at other rotation angles because it is assumed to be rotated in 90 ° units. It may not be applicable in the case of only a fragment corresponding to a part. Therefore, the applicable cases are limited to a very small part.

Therefore, in order to practically and effectively use the technology relating to the learning function, how to provide information having a high learning possibility is a problem. That is, how to improve the quality (learnability) of the input information (learning information) to the learning vector quantizer is an unsolved problem and must be solved for each individual case.

The present invention has been made in view of the above points,
The above-mentioned conventional problem is solved, and automatic recognition of a figure is realized without creating software for figure recognition corresponding to each figure shape, regardless of the position change, rotation, enlargement or reduction of the figure. An object of the present invention is to provide a line figure element learning recognition method capable of automatically recognizing an entire figure from an element.

[0009]

FIG. 1 is a diagram (No. 1) for explaining the principle of the present invention.

According to the present invention, the core line of a graphic is obtained from the line graphic drawn on the paper, the category information of the graphic is given, the core line is extracted as a pixel row, and as the information expressing the shape structure of the graphic, The curvature information expressing the degree of bending of each pixel in the pixel row of the figure is obtained as the information expressing the shape of the figure, the inflection point is extracted from the curvature information as the feature point, and the distance information between the feature points is calculated. , The normalized curvature information calculation processing of the figure that normalizes the curvature information of the section delimited by the feature points to become the input information of the adaptive vector quantizer (step 100), and applies the normalized curvature information Vector quantizer learning processing is performed by applying the adaptive vector quantizer to the adaptive vector quantizer (step 101). If there are a plurality of figures, normalized curvature information calculation processing and adaptive vector quantum chemistry are performed. The process is repeated for a plurality of learning target figures (step 102), and the normalized curvature information of the section delimited by the feature points is quantized by using the adaptively learned adaptive vector quantizer. Curvature information quantization processing is performed (step 103), and quantized curvature information, distance information between feature points, and category information are given to a learning vector quantizer to perform learning vector quantizer learning processing ( Step 104), when there are a plurality of figures, the normalized curvature information quantization processing and the learning vector quantizer learning processing are repeated for the plurality of learning figures (step 105), and recognition other than the learning figures is performed. For shapes,
The normalized curvature information calculation process is performed (step 106), the normalized curvature information of the recognition figure is quantized by the adaptive vector quantizer (step 107), and the quantized curvature information and the distance between the feature points are quantized. Information is input to the learning vector quantizer, recognition processing is performed, and recognition result information of the recognition graphic is obtained from the output information of the quantizer (step 108).

FIG. 2 is a diagram (part 2) for explaining the principle of the present invention.

According to the present invention, the core line of a graphic is obtained from the line graphic drawn on the paper, the category information of the graphic is given, the core line is extracted as a pixel row, and as the information expressing the shape structure of the graphic, The curvature information expressing the degree of bending of each pixel in the pixel row of the figure is obtained as the information expressing the shape of the figure, the inflection point is extracted from the curvature information as the feature point, and the distance information between the feature points is calculated. , Performing a normalized curvature information calculation process of a figure that normalizes the curvature information of the section delimited by the feature points to become the input information of the adaptive vector quantizer (step 200), and applies the normalized curvature information. The adaptive vector quantizer learning process for giving the adaptive vector quantizer an adaptive learning is performed (step 201). When there are a plurality of learning figures, the normalized curvature information calculation process and the adaptive vector are performed. amount Normalization curvature information quantization for quantizing the normalized curvature information of the distance information of the feature points using the adaptive vector quantizer. The processing is performed (step 203), and based on the curvature information quantized by the normalized curvature information calculation processing and the adaptive vector quantizer learning processing and the distance information between the feature points, The distance information of each section within a certain range of the section is converted into a relative value based on the distance of the section of interest, and the quantized curvature information and category information are combined into a learning vector quantizer. A learning vector quantizer learning process for giving a learning vector quantizer a figure is performed (step 204). When there are a plurality of figures, the normalized curvature information quantization process and the learning vector quantum are performed. The learning process is performed (step 205), the normalization curvature information calculation process similar to step 200 is performed on the new recognition figure other than the learning figure (step 206), and the recognition figure is normalized in step 206. The curvature information is quantized by an adaptive vector quantizer (step 207), and the quantized curvature information and the section length information are converted into relative values, which are then input to the learning vector quantizer. The recognition result information of the recognition graphic is obtained from the output information of the digitizer (step 208).

[0013]

The present invention utilizes the degree of bending (curvature information) of a figure in the process of human beings recognizing and understanding the figure. Also, paying attention to the fact that the figure is captured by the characteristics of the partial shape (a sharp bend, etc.) and its combination,
Further, in recognition / understanding of a figure in early childhood, the focus is on gradually understanding the figure while learning the shape of the given figure many times.

Here, the adaptive vector quantization technique is used for grasping the characteristics of the partial shape of the former, and the learning vector quantization technique which has been rapidly advanced in recent years is used for the learning function of the latter. Thus, focusing on the fact that the curvature information of the figure accurately represents the feature of the shape of the figure, the curvature information is extracted. Furthermore, the adaptive vector quantization with the curvature information as an input collects the curve parts having similar characteristics, and the information of the relative length with the adjacent curve parts is used as the input information to the learning vector quantization. . Thus, the figure is learned, and the learning vector quantizer of the learning result is used to recognize the figure to be recognized.

As a result, by simply presenting (learning) the figure to the system, it becomes possible to recognize the figure element from the figure element without being affected by changes in size, position and rotation. No such figure recognition programming is required.

[0016]

First, the principles of the adaptive vector quantizer and learning vector quantizer used in the present invention will be described.

FIG. 3 is a block diagram showing the principle of the adaptive vector quantizer.

The adaptive vector quantization method has recently attracted attention for its capability, and various methods / methods have been developed as existing technologies. The content does not conflict with the scope of the claims of the present invention and is treated as a black box.

The adaptive vector quantizer 20 treated as this black box has input information 10 and output information 40.
There is. The learning process is performed by providing the adaptive vector quantizer 20 with a predetermined fixed amount of the input information 20 for learning.

The input information 20 referred to here is an arbitrary (but predetermined) numerical value vector.

Inside the adaptive vector quantizer 20,
Reference vector 30 having the same dimension as any number of input information
Is held, the input information 10 is sequentially compared with the reference vector 30 at the time of being given, and the reference vector having the closest distance to the given input information 10 is selected. The value of the selected reference vector is modified by a value obtained by multiplying the difference by a coefficient ε smaller than 1 so that the difference with the input information in each dimension of the vector becomes small. The coefficient ε monotonically decreases as the learning progresses, and the input information 10 for learning is presented for the preset number of times of learning, and is controlled to become 0.0 when the learning ends. That is, as the learning progresses, each reference vector 30 becomes a value representative of the distribution of the input information 10 around the vector value, and when the learning ends, the reference vector inside the adaptive vector quantizer 20. Group 30 is fixed and does not move thereafter.

The quantization process is performed using the reference vector group 30 for which learning has been completed. That is, the input information 10 is compared with each reference vector 30 as in learning, and the reference vector having the shortest distance is selected. And
The adaptive vector quantizer 20 outputs the index given to the selected reference vector 30 as output information.

In the adaptive vector quantizer utilized in the present invention, this index is
It is a unique code (number). In the present invention, curvature information is given to this adaptive vector quantizer to perform quantization. By this quantization, the same index is obtained for the portions having similar shapes, and different indexes are obtained for the portions having different shapes.

FIG. 4 is a block diagram showing the principle of the learning vector quantizer.

Regarding the learning vector quantizer, various variations have been developed as existing techniques. The content thereof is a black box which does not conflict with the claims of the present invention. The learning vector quantizer 60 provided as the black box has the output information 80 together with the input information 50 as in the adaptive vector quantizer 20 described above.

The difference from the adaptive vector quantizer 20 is that the category information for the input information 50 is the teacher signal 90.
Is to give as. The learning process is performed by supplying the learning vector quantizer 60 with the input information 50 and the teacher information 90.

The learning of the learning vector quantizer 60 is basically the same as the learning of the adaptive vector quantizer 20 described above. The difference is that each reference vector is labeled, and that label represents a category as an output. The output information 40 of the adaptive vector quantizer 20 is the index of the reference vector, but the output information 80 of the learning vector quantizer 60 becomes the above-mentioned label (category).

In the learning process, the reference vector 70 having the closest distance to the presented input information 50 for learning is selected in the same manner, but the label assigned to the selected reference vector is the same. The reference vector 70 depends on whether or not it matches the correct category given as the teacher information 90.
How to update is different. When the label matches the correct category, the difference is multiplied by a coefficient ε smaller than 1 so that the difference with the input information 50 in each dimension of the vector becomes small as in the adaptive vector quantizer 20. Only the value is modified. When the label does not match the correct category, the difference is corrected by a value obtained by multiplying the difference by the coefficient ε so that the difference with the input information 50 is increased. The coefficient ε is the adaptive vector quantizer 2
Similar to that of 0, it is controlled so that it decreases monotonically as the learning progresses and becomes 0.0 at the same time as the end of the learning. That is, as the learning progresses, each reference vector moves to the darker part of the distribution of the category of the label, and after learning ends, it does not change.

Learning vector quantizer 60 for which learning has been completed
When new input information 50 is given to, the most similar reference vector 70 among the reference vectors 70 acquired from the already learned input information 50 has (category information) as the output information 80 of the learning vector quantizer 60. You can get it.

FIG. 5 shows the outline of the present invention.

In the present invention, as the input information 50 to the learning vector quantizer 60, the curvature information which accurately expresses the figure shape which is the main object of the present invention is quantized by the adaptive vector quantizer 20 described above. Give things. Then, the figure is learned by giving the category information of the figure as teacher information. The learning vector quantizer 60, which has finished learning, is given the curvature information of a new figure as input information to cause the new figure to be categorized (classification processing / recognition processing).

In general, the input information, the output information, and the teacher information are given as arbitrary numbers.

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 6 shows an outline of the operation of one embodiment of the present invention. The process of the present invention includes a learning process A and a recognition process B.

First, in the learning process A, the learning graphic 100
Is input, the normalized curvature information calculation processing a of the learning graphic is performed, the normalized curvature information is input to the adaptive vector quantizer, and the adaptive vector quantizer learning processing b that performs learning is performed. To do. Next, the adaptive vector quantizer reference vector group 110 is input to the adaptive vector quantizer, and the quantization process c is performed. Further, the learning vector quantizer reference vector group 120 is input to the learning vector quantizer, and the learning process d is performed.

In the recognition process B, the recognition graphic 13 is used instead of the learning graphic 100 used in the learning process A-a.
0 is used, the normalized curvature information calculation process e of the recognized figure is performed, and when the normalized curvature information is obtained, the learning process A
The learned adaptive vector quantizer reference vector group 140 generated by the process b) is input, and the quantization process f by the adaptive vector quantizer is performed. Finally, the learned learning vector quantizer reference vector group 150 generated by the processing d in the learning process A is input, and the recognition processing g by the learning vector quantizer is performed.

Here, the learning process A will be described in detail.

FIG. 7 shows the operation up to the learning process in one embodiment of the present invention.

Graphic core line extraction processing The line graphic drawn on the paper can be read as digital information from an existing scanner device into an electronic computer (hereinafter simply referred to as a computer), and the line graphic read into the computer. The information is subjected to the core-forming processing (thinning processing) as a conventional technique to obtain the core-forming of the line figure (center line of one pixel width) (step 1).

Category information addition process Category information of the target figure (learning figure) (name of figure:
The classification name) is computerized by inputting it from a keyboard or the like connected to the computer (step 2).

Core Line Pixel Column Extraction Process The above core line information is connection information of pixels having a width of one pixel,
By tracking the connection of the connected pixel columns, the core line pixel column information can be obtained (step 3). The pixel column information is a column of coordinate values of each pixel. Hereinafter, the pixel column is represented by F, and each pixel is represented by d _i . That is, the pixel row F
F = | d _i | (where n is the number of pixels in the pixel row, 1 ≦
i ≦ n, {} represents a set of d _i . The same shall apply hereinafter).
Further, the coordinate value of each pixel is represented by (x _i , y _i ), and the pixel array F is represented by F = {(x _i , y _i )}.

Curvature Information Calculation Processing The curvature information of each pixel in the pixel array expresses the degree of bending of the line figure and can be regarded as accurately expressing the geometrical shape information of the line figure. . The curvature information is calculated for each pixel as follows (step 4).

[0043] That is, represents the curvature of the pixel d _i in C _i, obtained as follows. As shown in FIG. 8A, two elements (d _(ik) and d _{(i + k)} ) of pixels that are k away from the pixel d _i are set, and the pixel d _(ik) and the pixel d ₍ Let L _{k be} the direct line segment connecting _{i + k)} . Next, a perpendicular is created from the pixel d _i to the straight line L _k , the foot of the perpendicular (intersection point of the perpendicular and the straight line L _k ) is set to q _k , and the distance between the pixel d _i and the point q _k is set to B _k . . Further, the value of the k 1, 2, 3, ... and when going by sequentially increasing, B _k ≦ E for the parameters E whose distance B _k is B _k is preset for parameter E to be set in advance
Under the condition, the maximum value of k is obtained, and the maximum value is set to L. Then, as shown in FIG. 8B, a vector having a pixel d _(iL) as a start point and a pixel d _i as an end point is set as V ₋ , and a pixel d
A vector having _i as a starting point pixel d _{(i + L)} as an ending point is V ₊ .

The angle formed by the two vectors (V ₋ , V ₊ ) thus created is θ _i (the angle is expressed in radians, −π ≦ θ _i ≦ π: π represents the circular constant). And Here, as shown in FIG. 8C, pixels d _(iL) , d _i , d
_It is approximated as an arc passing through the three points _{(i + L)} . The resulting arc having a radius as R _i, the reciprocal 1 / R _i of the radius R _i
And I shall express the degree of bending in the pixel d _i, the curvature information C _i in the pixel _{d i, C i = 1 /} R i
And

The curvature information (referred to as F _C ) of the pixel row F created as described above is represented by F _C = {C _i }.

In the description of the present embodiment, for simplification of the description, it is assumed that the target line figure (cord line) is a figure that draws a loop.

That is, the first pixel ((x ₁ ,
y ₁ ): the starting pixel d _i ) and the last pixel d _i ((x _n ,
y _n ): the end pixel) are adjacent, and
It is possible to easily obtain two pixels (d _(ik) , d _{(i + k)} ) that are k pixels away from the start point pixel and the end point pixel and are adjacent to each other.

Further, in the above, it has been experimentally confirmed that the preset parameter E may be set to several times the pixel interval length.

The term "curvature" used in the present invention does not mean a mathematically strictly defined curvature, but as described above in order to make it intuitively easy to understand,
The term "curvature information" is used to represent the degree of bending of a figure.

An example of the curvature information F _c of the pixel row F obtained as described above will be shown.

FIG. 9 is a diagram showing an example of the curvature information according to the embodiment of the present invention as a curve (line) graph. In the figure, the curvature information F _c of the pixel row F is shown.

Feature Point Extraction Process Next, from the curvature information F _c = {C _i } of the pixel row (core line) obtained as described above, the feature points of the target graphic are extracted, and each feature point is extracted. The distance (the number of pixels) between them is calculated (step 5). A specific example of this is shown below.

As in the example of FIG. 9, the curvature information is represented by a curved (target) graph having irregularities. In this method, points in the graph that cross the line of 0 are extracted as feature points. That is, the inflection point of the unevenness of the line figure on the two-dimensional plane is calculated as the characteristic point. As for the method of calculating the feature points, as can be easily understood from the example of FIG. 9, the points having the curvature of 0.0 may be extracted. The distance between these feature points is
It can be expressed by the number of pixels between feature points.

Curvature information normalization processing Since the curvature information of each section delimited by the feature points obtained by the above procedure is information that varies depending on the enlargement / reduction of the figure, this information is used for enlarging / reducing the figure. Normalize to be irrelevant (step 6).

Since the curvature information is the reciprocal of the radius of the arc as described above, it has an inversely proportional relationship with the enlargement / reduction of the figure. Therefore, by multiplying the curvature information of each section by the length of that section, it is normalized regardless of the enlargement / reduction of the figure as the shape of the section is indicated with respect to the enlargement / reduction of the figure. Curvature information is calculated. One of the main points of the present invention is to use the curvature information that is normalized so that it is irrelevant to enlargement / reduction of the figure as the shape of the section by dividing this figure by its characteristic points. .

Adaptive Vector Quantizer Learning Processing Next, the normalized curvature information obtained for each section divided by the feature points as described above is used as learning input information for the adaptive vector quantizer. Then, the adaptive vector quantizer performs the learning process as described in the conventional technique (step 7).

At this time, curvature information (consisting of curvature columns of several pixels in each section as described above) is input to the number of input vector dimensions of the adaptive vector quantizer before being given as learning input information. To fit.

Specifically, the pixel string in the section is equally divided by the number obtained by subtracting 1 from the number of input vector dimensions of the adaptive vector quantizer, and the average of the curvatures in the neighborhood is taken at each division point. Give things as input information. Since one figure usually has a plurality of inflection points, it is divided into a plurality of sections. The curvature information of the plurality of sections is collectively given to the adaptive vector quantizer as learning input information.

Quantization Processing by Adaptive Vector Quantizer for Plural Figures Similarly, when there are plural learning figures, inflection points are similarly obtained for all the learning figures, and the learning points are divided. Obtains curvature information that has been normalized for all intervals, adapts the curvature information to the input vector dimensionality of the adaptive vector quantizer, and collectively learns the curvature information in the adaptive vector quantizer. It is given as input information for use and learning processing is performed (step 8).

Quantization processing by the adaptive vector quantizer The adaptive vector quantizer adaptively acquires the reference vector group by the above procedure. As a result, the adaptive vector quantizer enables the sections from the inflection points having similar curve shapes to the inflection points to be grouped together, and the same code (index) can be attached to the similar sections. Become. To this end, the learning input information calculated in step 8 described above is given as recognition input information, the quantization processing shown in the conventional technique is performed, and the result is quantized to obtain a code. (Index) is obtained (step 9). Hereinafter, the code (index) of the section obtained by this adaptive vector quantizer will be referred to as index information.

Learning Vector Quantizer Learning Process Index information of each section delimited by the inflection point of the learning figure obtained in step 9 and the length of each section obtained in step 5 ( Number of pixels) information as input information,
Further, the category information given in step 2 is given to the learning vector quantizer as the teaching information as described below, and the learning process is performed (step 10).

Assuming that one section A of one learning graphic is the section of interest at the moment and La is the length of the section A, it is given to the learning vector quantizer at one time as one piece of data. It is the input information of the section within the length range obtained by multiplying the length La of each of the front and rear sections A by a predetermined number (C). The input information of the section A of interest is the number of columns (B) in which one column is defined by the number of input elements of the reference vector of the adaptive vector quantizer T.
Column) is prepared, and only the value of the element corresponding to the index of the section A in each column is set to 1.0, and the other values are set to 0.0. In the sections before and after the section of interest, the number of rows according to the relative length based on the length of the section A of interest, before and after the element array of the current row B, respectively.
It is represented by arranging the input element arrays representing the index information of the section. This is repeated for adjacent sections until B × C input element arrays are obtained for each of the front and rear. As a result, the number of input elements required for the learning vector quantizer becomes T × (2C + 1) × B. In this way, it is one of the main points of the present invention to provide the learning vector quantizer with the section index information before and after the section of interest and the information of the relative length represented by the number of columns of input elements. Is.

As a result, learning / recognition that is not affected by the enlargement / reduction of figures becomes possible. The above learning process is performed for each section divided by the inflection points of the figure to be learned.

Learning Vector Quantizer Learning Process for Plural Figures When there are a plurality of learning figures, the learning process is performed for each learning figure using the procedure of step 9 and step 10 ( Step 11) The learning vector quantizer does not immediately end learning only after it is presented with the input information and the teacher information once.
It is necessary to repeat learning processing a specified number of times until learning is completed.

With the above processing procedure, the learning process of a plurality of learning figures is completed, and the learning vector quantizer has learned a plurality of figures.

The recognition processing of the recognition graphic created separately by using this learning vector quantizer will be described below with reference to FIG.

FIG. 10 shows a recognition process according to an embodiment of the present invention.

Normalized curvature information calculation processing of recognized figure First, normalized curvature information and section length information are calculated for each section delimited by the inflection point of the recognition figure (step 12). . The procedure for obtaining this is the same as the procedure from step 1 to step 6 of the learning processing described above.

Quantization processing by adaptive vector quantizer Learning of curvature information of each section of the recognition figure obtained by the normalized curvature information calculation processing of the recognition figure in step 12 obtained by the procedure of step 8. Quantization is performed using the adaptive vector quantizer that has already been performed, and index information is obtained (step 13).

Recognition Processing by Learning Vector Quantizer The index information obtained in the procedure of step 13 and the information of the length of each section obtained in the process of step 12 are learned in step 11. It is input as input information to the already-learned learning vector quantizer, the recognition processing is performed, and the processing result (output information of the learning vector quantizer) is obtained (step 14).

Since the category information is given as the teacher information during the learning process of the learning vector quantizer, the output information at the time of the recognition process can be obtained as the category information of the given recognition graphic. That is, it is possible to obtain information about which figure is most similar to the figure presented and learned in advance, and the figure can be recognized.

Further, since the curvature information is used as the basis of the processing, the learning / reduction is not affected by the translation / rotation of the figure.
Recognition processing is performed. In addition, the shape of the curve is normalized by the length, and since only the relative length between adjacent sections is used, learning is not affected by the scaling of the figure. -The recognition process is performed. If some of the recognized figures cannot be observed due to crossing with other figures, etc., when calculating the distance between the input information and the reference vector, leave the distance for the unobservable section empty. The condition is calculated as 0.0. By doing so, the distance from the reference vector group can be calculated from the similarity using only the observed part, and as a result, the entire figure can be recognized from the figure element.

One of the main objects of the present invention is to obtain the output from the learning vector quantizer so that the unknown part does not affect the result.

In the above description, the learning target figure and the recognition target figure are input as a figure drawn on the paper from the scanner device or the like, but the mouse or the like is used to directly input the figure on the display device connected to the computer. Even in the graphic information input to the calculator while drawing, the core line of the graphic (center line)
Can be easily obtained as the pixel row described above, and
Other methods may be used within the scope of the present invention, as long as the core formation of the target graphic is obtained as the above-described pixel row.

[0075]

As described above, according to the present invention, the curvature information that accurately expresses the shape of a figure is adaptive vector quantizer and learning vector quantum so as not to be affected by the enlargement / reduction of the figure. Since it is a method of recognizing a separately created figure by using the adaptive vector quantizer and the learning vector quantizer that have been trained by using a digitizer, Automatic recognition of figures can be realized without creating figure recognition software (program) corresponding to, and unlike the method that has other learning functions, the figure can be recognized regardless of position change, rotation, enlargement and reduction. Since the whole figure can be automatically recognized from the elemental pieces, its practical value is extremely high.

[Brief description of drawings]

FIG. 1 is a diagram (No. 1) for explaining the principle of the present invention.

FIG. 2 is a diagram (part 2) for explaining the principle of the present invention.

FIG. 3 is a principle configuration diagram of an adaptive vector quantizer.

FIG. 4 is a principle configuration diagram of a learning vector quantizer.

FIG. 5 is a diagram showing an outline of the present invention.

FIG. 6 is a diagram showing an outline of the operation of one embodiment of the present invention.

FIG. 7 shows an operation up to a learning process according to an embodiment of the present invention.

FIG. 8 is a diagram for explaining curvature information.

FIG. 9 is a diagram showing an example of curvature information according to an embodiment of the present invention as a curved line graph.

FIG. 10 is a diagram showing a recognition process according to an embodiment of the present invention.

[Explanation of symbols]

 10, 50 Input information 20 Adaptive vector quantizer 30, 70 Reference vector group 40 Output information (index information) 60 Learning vector quantizer 80 Output information 90 Teacher information 100 Learning figure 110 Adaptive vector quantizer reference vector Group 120 Learning vector quantizer reference vector group 130 Recognition figure 140 Learned adaptive vector quantizer reference vector group 150 Learned learning vector quantizer reference vector group a Learning curve normalization information calculation processing b Adaptation Vector quantizer learning process c Quantizer process by adaptive vector quantizer d Learning vector quantizer learning process e Normalized curvature information calculation process of recognition pattern f Adaptive vector quantizer process g Learning vector quantizer Recognition processing by

Claims (2)

[Claims] 1. A core line of a graphic is obtained from a line graphic drawn on a paper surface, category information of the graphic is given, the core line is extracted as a pixel row, and the information is used to represent the shape structure of the graphic. , The curvature information expressing the degree of bending of each pixel of the pixel row of the figure is obtained as information expressing the shape of the figure, the inflection point is extracted from the curvature information as a feature point, and the distance between the feature points is calculated. Information is calculated, and normalized curvature information calculation processing of a figure for normalizing curvature information of a section delimited by the feature points to be input information of the adaptive vector quantizer is performed. To an adaptive vector quantizer to perform adaptive vector quantizer learning processing for adaptive learning, and when there are a plurality of figures, the normalized curvature information calculation processing and the adaptive vector quantization Multiple learning targets Form number of iterations, using the adaptively learned adaptive vector quantizer to perform normalized curvature information quantization processing for quantizing the normalized curvature information of the section delimited by the feature points, The quantized curvature information, the distance information between the feature points, and the category information are given to a learning vector quantizer to perform learning vector quantizer learning processing. When there are a plurality of figures, The normalized curvature information quantization processing and the learning vector quantizer learning processing are repeated for a plurality of learning figures, and the normalized curvature information calculation processing is performed on recognition figures other than the learning figures, The normalized curvature information of the recognition figure is quantized by the adaptive vector quantizer, and the quantized curvature information and distance information between the feature points are input to a learning vector quantizer for recognition processing. And the quantization Line figure segment learning and recognition method characterized in that the output information obtaining recognition result information of the recognition graphic. 2. A core line of a graphic is obtained from a line graphic drawn on a paper surface, category information of the graphic is added, the core line is extracted as a pixel row, and the information is used to represent the shape structure of the graphic. , The curvature information expressing the degree of bending of each pixel of the pixel row of the figure is obtained as information expressing the shape of the figure, the inflection point is extracted from the curvature information as a feature point, and the distance between the feature points is calculated. Information is calculated, and normalized curvature information calculation processing of a figure for normalizing curvature information of a section delimited by the feature points to be input information of the adaptive vector quantizer is performed. To the adaptive vector quantizer to perform adaptive vector quantizer learning processing for adaptive learning, and when there are a plurality of learning figures, the normalized curvature information calculation processing and the adaptive vector Quantizer learning process It is repeated for the number of figures to be learned, and the normalized curvature information quantization processing for quantizing the normalized curvature information of the distance information of the feature points is performed using the adaptive vector quantizer, and the normalized curvature information calculation processing is performed. Based on the curvature information quantized by the normalization and the adaptive vector quantizer learning process and the information on the distance between the feature points, the distance of each section within a certain range of the section of interest Information is converted into a relative value based on the distance of the section of interest, and the quantized curvature information and the category information are combined and given to the learning vector quantizer to be applied to the learning vector quantizer. Learning vector quantizer learning processing for learning the learning figure is performed. When there are a plurality of figures, the normalized curvature information quantization processing and the learning vector quantization learning processing are performed, and For recognition The normalized curvature information calculation process is performed on the shape, the normalized curvature information of the recognition figure is quantized by the adaptive vector quantizer, and the quantized curvature information and the length of the section are quantized. A line-figure element learning recognition method, wherein information is converted into relative values and then input to the learning vector quantizer, and recognition result information of a recognition graphic is obtained from output information of the learning vector quantizer.