JPH05197812A - Figure shape learning recognition method - Google Patents

Abstract

(57) [Summary] [Object] An object of the present invention is to provide a figure shape learning recognition method capable of automatically recognizing a figure without individually programming the shapes of various figures. [Structure] The present invention uses neural network technology for recognizing and understanding a figure in early childhood, as well as gradually learning the figure while learning the shape of the given figure many times. Therefore, paying attention to the fact that the curvature information of the figure accurately represents the feature of the shape of the figure, the curvature information is extracted, the curvature information is input as input information to the neural network, and the neural network Learning is performed, and the learning result is used to recognize the figure to be recognized.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a graphic shape learning recognition system, and more particularly, in a drawing information system for designing, maintaining and managing information on a drawing, the graphic information drawn in the drawing can be handled by a computer. The present invention relates to a graphic shape recognition system that automatically recognizes graphic information in a drawing as a means for converting it into information, that is, acquiring initial graphic information from the drawing.

[0002]

2. Description of the Related Art A conventional method of figure shape recognition for converting figure information in a drawing into information that can be handled by a computer is to attach the drawing to a large tablet and use the cursor to form each figure. The line is extracted manually.
This method requires a lot of time, so as another method,
A method of automatically recognizing a figure in a drawing and acquiring figure information has also been developed.

A typical conventional method for automatically recognizing a figure in this drawing is software (a figure recognizing program) for extracting characteristic information of a target figure for each figure and then classifying the figure type. Then, the figure recognition program is used to recognize the figure in the drawing.

On the other hand, in recent years, the neural network technology has progressed, and the technology for utilizing it for recognition of character patterns and voice signals is also being developed. Neural networks automatically learn input information by applying input information and category information corresponding to it as teacher information, and using techniques such as the error back-propagation method that applies general learning methods to multilayer networks. It has been verified that it is effective for automatic learning and recognition processing of simple character patterns whose size, direction, and position are fixed.

However, how to effectively use the neural network technology is to give information having a high learning possibility to the neural network, that is, input information (learning information) to the neural network. How to raise the quality (learnability) of is an unsolved problem and must be solved for each individual case.

[0006]

The above-mentioned conventional method using a figure recognition program, which is procedural programming, usually requires programming for each figure type to be handled. Therefore, as the number of figure types handled increases, The more complicated the structure of a figure becomes, the more enormous programming is required, and the greater the development cost and time required for the software development, which is a big problem. Therefore, in the conventional method, the shape of a graphic that can be handled is limited to a simple shape, and it is difficult to automatically acquire general graphic information of a drawing. The present invention has been made in view of the above points, and it is an object of the present invention to solve the above problems and provide a figure shape learning recognition method capable of automatically recognizing a figure without individually programming the shapes of various figures. To aim.

[0007]

FIG. 1 illustrates the principle of the present invention. The graphic shape recognition method of the present invention comprises a graphic core line extracting means 1 for obtaining a center line of a graphic from a line graphic, a category information adding means 2 for adding category information of the line graphic,
A core line pixel array extraction unit 3 for extracting the center line of the graphic obtained by the graphic core line extraction unit 1 as a pixel array, and a curve for obtaining curvature information for expressing the shape structure of the graphic for each pixel of the pixel array. The information calculating means 4, the characteristic point extracting means 5 for extracting, as characteristic points, the points at which the curvature information calculated by the curve information calculating means 4 changes, and the curvature information in the vicinity of the characteristic points calculated by the characteristic point extracting means 5. Normalizing means 6 for generating input information of the neural network 21 by normalizing
The normalized curvature information obtained by the normalizing means 6 and the category information of the target graphic obtained by the category information adding means 2 are input to the neural network 21, and the neural network 21 is made to learn the target graphic. From the learning means 7 and the figure core line extraction means 1 to the target figure learning means 7
A plurality of figure learning control means 8 for making the neural network 21 learn a plurality of target figures by using the above, and a recognition figure 1 created by the neural network 21.
2 by using each of the figure core line extracting means 1 except the category information adding means 2 to the normalizing means 6
Normalized curvature calculating means 9 for obtaining the normalized curvature information of
And the curvature information obtained by the normalized curvature calculation means 9 is input to the neural network 21, and the recognition result information of the recognition graphic 12 is obtained from the output information from the neural network 21.
Have and.

[0008]

The present invention utilizes the bending degree (curvature information) of a figure in the process of human beings recognizing and understanding the figure, and the figure given to the figure recognition and understanding in early childhood. Focusing on understanding the shape gradually while learning the shape many times, the latter learning function tries to solve the conventional problem by using the neural network technology that has advanced rapidly in recent years. It is a thing.

In particular, paying attention to the fact that the curvature information of the figure accurately expresses the feature of the shape of the figure, the curvature information is extracted, and the curvature information is input as input information to the neural network, The figure is learned, and the learning result is used to recognize the figure to be recognized. As a result, since it becomes possible to recognize the figure only by learning various figures, the conventional figure recognition programming is not necessary.

[0010]

First, the principle of the neural network used in the present invention will be described. FIG. 2 shows the principle structure of the neural network. Neural network 32
Various configurations have been developed as existing technology. The content does not conflict with the scope of the claims of the present invention and is a black box. The neural network 32 given as the black box has input information 31 and output information 33, and category information for the input information 31 is given as teacher information 34. The learning process is performed by giving the input information 31 and the teacher information 34 to the neural network 32.
The difference between the output information 33 and the teacher information 34 is processed as small as possible. As a method of making this difference (error) as small as possible, an error back propagation method or the like is generally used. By repeatedly learning a plurality of input information 31 and teacher information 34 a plurality of times,
The difference (error) between the output information 33 of the neural network 32 and the teacher information 34 becomes the minimum, and the learning ends.
When new input information 31 is given to the learned neural network 32, the output information 33 of the neural network 32 is information corresponding to the teacher information 34 that is most similar to the already learned input information 31.

In the present invention, the neural network 32
Curvature information that accurately expresses the graphic shape that is the main object of the present invention is given as input information 31 to the. The neural network 32 uses the category information of the target figure as teacher information 34.
Learning the figure by giving as. The network 32 that has finished learning is categorized by performing the classification process and the recognition process of the new figure by being given the curvature information of the new figure as the input information 31. In general, each of the input information 31, the output information 33, and the teacher information 34 is given by an arbitrary number and arbitrary numerical value. The term "curvature" used in the present invention does not mean a mathematically rigorously defined curvature, and in order to make it intuitively easy to understand, the term "curvature information" is used to represent the degree of bending of a figure. There is.

FIG. 3 is a diagram for explaining the learning process of one embodiment of the present invention. In the figure, a computer 2
It is the learning graphic 11 that is input to 0.

[Step A1; Graphic Core Line (Center Line) Extraction Processing] The line graphic drawn on the paper is read as digital information in an electronic computer (hereinafter referred to as a computer) using an existing scanner device. . The learning figure 11 (line figure information) read into the computer 20 is converted into one of the line figures by the core making process (thin line making process) similar to the conventional technique.
By obtaining the core line which is the center line of the pixel width, the core line information which is the connection information of the pixels of one pixel width is obtained.

[Step A2; Category Information Adding Process]
The category information, which is the name and classification name of the learning graphic 11 read in step A1, is computerized by inputting from a keyboard or the like connected to the computer 20.

[Step A3; Core-line Pixel Row Extraction Processing] The core-line pixel row information is obtained by tracking the connection of the pixel rows connected from the core-line information in step A1.

[Step A4; Curvature Information Calculation Processing] The core pixel array information is a series of coordinate values of each pixel. Hereinafter, the pixel row is represented by F and each pixel is represented by d _i . That is, the pixel row F is F = {d _i } (where n is the number of pixels in the pixel row, 1 ≦ i ≦
n {} represents a set of d _i . ). The coordinate value of each pixel is represented by (x _i , y _i ), and the pixel array F is F = {(x _i ,
y _i )}.

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. One of the main objects of the present invention is to extract and use curvature information to accurately express this figure shape. The curvature information is calculated for each pixel as follows.

FIG. 4 is a diagram for explaining the curve information calculation processing of one embodiment of the present invention. In the figure, the pixel di
A method of obtaining the curvature Ci of is described. As shown in FIG. 4A, two pixels (d _(ik) and d _{(i + k)} ) separated from the pixel di by k are set, and the pixel d _(ik) and the pixel d _{(i + k} ) are set. Let L _{k be} the straight line segment that connects ₎ . Next pixel d
A perpendicular line from _i to the straight line L _k is drawn, the leg of the perpendicular line (intersection point of the perpendicular line 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, when the distance B _k is set to a preset parameter E, B _k ≦ E, the value of k is increased to 1, 2, 3, ...
The maximum value of is calculated and the maximum value is set to L. Then, as shown in FIG. 9B, a vector having the pixel d _(iL) as the viewpoint and the pixel d _i as the end point is V-, and the pixel d _i is the viewpoint and the pixel d _{(i + L)} is the end point. Let the vector be V +. The angle θ _i formed by the two vectors (V−, V +) thus created (the angle is expressed in radians, −π ≦ θ _i ≦ π: π
Is a representative of the circular constant), it is assumed to represent the degree of bending in the pixel d _i, a value of the curvature information C _i has been converted by the C _i = θ _i / π in the pixel d _i.

Thus, the curvature information C _i is -1.0≤C
_The value is normalized to _i ≤ +1.0. In the present invention, it is not always necessary to carry out this normalization, and simply use C
_It is also possible to set _i = θ _i . The curvature information (referred to as Fc) of the pixel row F created as described above is Fc =
It is represented by {C _i }.

In the description of the present embodiment, for simplification of the description, the following description will be given assuming that the target line figure (core line pixel row) is a figure that draws a loop. That is, the first pixel ((x ₁ , y ₁ ): start point pixel) and the last pixel ((x _n , y _n ): end point pixel) of the pixel row are adjacent to each other, whereby the start point pixel and the end point pixel It is possible to easily obtain two pixels (d _(ik) , d _{(i + k)} ) that are d pixels away from the start point / end point pixels in the vicinity of the pixel. Further, it has been experimentally confirmed that the above-mentioned preset parameter E may be set to be several times the pixel interval length.

FIG. 5 is a diagram showing an example of curve information according to an embodiment of the present invention. This figure is an example of the curvature information Fc of the pixel row F obtained as described above, and when the number of pixels forming the original figure shown in FIG. The position of the 200th pixel from the gazing point in the curvature information in ( ₁ ) is m ₁ , m ₂ (m ₁ = m ₂ ).

[Step A5: Feature Point Extraction Processing] The feature points of the target graphic are extracted from the curvature information Fc = {C _i } of the pixel row (core line) obtained in the core pixel row extraction process of step A4. .. The feature point extraction processing will be described in detail below. As in the example of FIG. 5, the curvature information is expressed by a curved graph having irregularities, as shown in the shape of FIG. The maximum and minimum points of the graph are extracted as feature points. That is, the change points of the irregularities of the line figure on the two-dimensional plane are calculated as the characteristic points. As for the method of calculating the feature points, as can be easily understood from the example of FIG. 5, the points where the curvature changes from increase to decrease and the points where the curvature changes from decrease to increase may be extracted. Further, when the adjacent feature points are greatly separated, the central point is also set as the feature point. As described below, the purpose of extracting the feature points is to normalize the curvature information in the vicinity of the feature points and allow the neural network 21 to learn, as will be described later. However, an arbitrary predetermined distance is set to the feature points. By providing the neural network 21 with curvature information of pixel positions, the number of pieces of learning target information can be reduced, and the processing time for learning can be shortened. That is, when the neighborhood curvature information for each pixel is given to the neural network 21, the learning performance /
Although the recognition performance after learning is improved, the larger the number of pixels that make up a line figure, the larger the number of learning information (and recognition information), which requires a lot of time, which makes it practical. This is because there is a possibility that the feature points will be lost, and it is necessary to perform feature point extraction processing for a type of thinning processing. Of course, the current technological progress is remarkable, so if the neural network 21 technology (hardware / software technology) progresses to the extent that there is no practical problem, and if the number of pixels of the target line figure is small. The curvature information in the vicinity of each pixel unit can be used as the learning target information / recognition target information, and all the pixels can be used as the feature points.

[Step A6: Normalized Curvature Information Calculation Process (Normalization Process) Using Feature Point as Point of Gaze] With respect to the feature point obtained by the feature point extraction process of Step A5, the curvature information of that feature point The curvature information is normalized so that is the input information to the neural network 21.

The normalization process will be described in detail below. We humans visually see the characteristic shape of an arbitrary figure,
When recognizing, pay attention to the shape of a characteristic portion at a close distance in a moment, and pay attention to the fact that the surrounding area is roughly viewed toward the surroundings (human visual / perceptual characteristics). The curvature information in the vicinity of is fine, and coarser curvature information is obtained and used toward the periphery. That is, when there gazing point characteristic points (pixels) d _m, in the lateral direction of the curvature of the curve graph in Fig. 5 from the fixation point d _m, for example, 1, 2, 3, 4,
5, 6, 7, ... Dividing at every pixel interval. Each of the divided parts is moved to the right by G ₀ , G _R1 , G _R2 , G _R3 , G _R4, ..., G
_Rj , ..., and G _L0 , G _LL , G _L2 , G _L3 , and
_Let G _L4 , ..., G _Lj ,. Specifically, in this example, each pixel in the right direction is G _R0 : d _m G _R1 : d _{(m + 1)} , d _{(m + 2)} G _R2 : d _{(m + 3) to} d _{(m + 5 ) )} G _R3 : d _{(m + 6) to} d _{(m + 9)} G _R4 : d _{(m + 10) to} d _{(m + 14)} G _R5 : d _{(m + 15) to} d _{(m + 20)} ... G _Rj : d _{(m + f (j)) to} d _{(m + f '(j))} ...

Further, each pixel in the left direction is G _L0 : d _m G _L1 : d _(m-1) , d _(m-2) G _L2 : d _{(m-3) to} d _(m-6) G _{L3: d (m-6)} ~d (m-9) G L3: d (m-10) ~d (m-13) G L5: d (m-15) ~d (m-14) ... G LJ : D _{(mf (j)) to} d _{(m-f '(j))} . For example, in the above example, the divided region G _R5 corresponds to six pixels of pixels d _{(m + 15)} , d _{(m + 16) ,} d _{(m + 17),} ... D _{(m + 20)} . In this way, as the distance from the feature point increases, the rough division is performed. The dividing method is not limited to the above example, and can be arbitrarily set according to the property of the figure to be handled. Further, the reason why the distribution is coarser as it goes to the peripheral position away from the feature point in this way is that it has the effect of reducing the input information to the neural network 21 while paying attention to the human visual and perceptual characteristics.

Further, the curvature information C _i within each divided range is made to correspond to each pixel as described above as follows. Pixels in the rightward _{direction, G R0: C m G R1} : C (m + 1), C (m + 2) G R2: C (m + 3) ~C (m + 5) G R3: C (m + _{6) to} C _{(m + 9)} G _R4 : C _{(m + 10) to} C _{(m + 14)} G _R5 : C _{(m + 15) to} C _{(m + 20)} ... G _Rj : C _{(m + f) (j)) to} C _{(m + f '(j))} .

Further, each pixel in the left direction is G _L0 : C _m G _L1 : C _(m-1) , C _(m-2) G _L2 : C _{(m-3) to} C _(m-6) G _L3 : C _{(m-6) to} C _(m-9) G _L3 : C _{(m-10) to} C _(m-13) G _L5 : C _{(m-15) to} C _(m-14) … _GLJ : C _{(mf (j)) to} C _{(m-f '(j))} .

Then, the average value of the curvature is obtained for each divided area. The curvature information of each is C _R0 , C _R1 , C _R2 ,
C _R3 , C _R4 , ..., C _Rj , ..., And C _L0 , C _L1 , C _L2 , C
_{When L3,} C _L4 , ... C _Lj , ... For example, the curvature C _R4 is C _R4 = (C _{(m + 10)} + C _{(m + 11)} + C _{(m + 12)} + C _{(m + 13)} + C
_{(m + 14)} ) / 4.

Attention point C_R0And C_L0The same as
Then, assuming that 20 divided areas are set in the left-right direction,
The curvature information in the case is {C_R1, C_R2,,, C
_R20,} (= {C_R20}) Curvature information for each division range
And And to the left, {C_L1, C_L2,,, C_L20,}
(= {C_L20}) The curvature information for each of the division range units.
Therefore, a total of 41 pieces of curvature information are obtained. Curvature information above
{{C_L20}, C _Ro, {C_R20}} Is the normalized curvature
Information, and the input information of the neural network 21
Become. If you set an arbitrary number of divisions,
{{C_L}, C _R0, {C_R}}.

[Step A7: Neural network
Learning control process] Next, the feature points are normalized as described above.
Curvature information {{C_L}, C_R0, {C _R}}
The network network 21 as input information, and
Gori information by the category information addition process of step A2
Error given as teacher information of neural network 21
The input information is automatically input to the neural network using the difference backpropagation method.
The network 21 is made to perform learning processing. Also, one study
If the learning figure has multiple feature points,
Neural network for normalized curvature information of individual feature points
Give to the work 21 to learn.

[Step A8: Plural figure learning control process] When a plurality of learning target figures exist, they are normalized for each learning figure and for each feature point in each learning figure. Neural network 21
And perform learning processing. Neural network 21
Does not end learning immediately after the input information and the teacher information are presented once, and it is necessary to repeatedly perform learning processing. This is similar to the fact that humans can perceive a figure gradually by seeing the figure (picture) many times in early childhood. That is, this procedure is repeated until learning is completed. The end of learning is the difference (error) between the output information obtained as the output when the input information is given to the neural network 21 and processed.
It is the time when the sum total becomes minimum and there is almost no change in the difference sum.

By the processing procedure of the above learning process, the learning process of a plurality of learning figures is completed, and the neural network 21 has learned (learned) a plurality of figures.

Next, the above neural network 21
A description will be given of the recognition of a recognition figure separately created by using. FIG. 7 is a diagram for explaining a recognition process according to an embodiment of the present invention.

[Step B9: Normalized Curvature Information Calculation Control Processing of Recognition Graphic] First, the recognition graphic 12 created by the neural network 21 is used to perform the processing from step A1 to step A6 described above, and the recognition graphic is executed. 1
Compute 2 normalized curvature information.

[Step B10: Neural Network Recognition Control Processing] The neural network 21 that has finished learning the normalized curvature information of the recognition graphic obtained by the normalized curvature information calculation control processing of the recognition graphic in step B9.
The neural network 21 performs a recognition process by inputting the input information as the input information. As a result of the recognition processing, the output information of the neural network 21 is obtained.

Since the category information is set as the teacher information in the learning process, the output information in the recognition process in step B10 can be obtained as the category information of the recognition graphic given to the neural network 21. That is,
It is possible to obtain, as output information, information about which figure of the figures that the neural network 21 has learned in various ways is most similar to, and thus, the figure recognition can be performed.

In the above embodiment, the learning target figure,
The figure to be recognized is read from the figure drawn on the paper from the scanner device etc. and is input to the computer 20, but it is drawn on the computer while directly drawing it on the display device connected to the computer using a mouse or a tablet device. Even with the inputted graphic information, the core line (center line) of the graphic can be easily obtained as the above-mentioned pixel row (the locus (coordinate value row) of the directly drawn graphic can be easily obtained as the pixel row information. Convertible). In this case, the drawn figures may intersect on the way. Also,
Even when another method is used, the present invention can be applied as long as the core line of the target graphic is obtained as the above-mentioned pixel row.

Further, in the above-mentioned embodiment, the description has been made by taking the figure which draws a loop as an example. However, even in the case of an open figure, that is, a figure in which the start point and the end point of the core pixel row of the figure do not coincide, the same as the above-mentioned example The curvature information can be easily obtained.

[0039]

As described above, according to the present invention, the curvature information that accurately expresses the shape of the figure is learned by using the neural network, and the figure created separately by the learned neural network is recognized. Since this method is used, the automatic recognition of the figure can be realized without creating the figure recognition software corresponding to each figure shape as in the prior art, so that the cost and time required for the software development can be reduced.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a diagram showing a principle structure of a neural network.

FIG. 3 is a diagram for explaining a learning process according to an embodiment of the present invention.

FIG. 4 is a diagram for explaining curve information calculation processing according to an embodiment of the present invention.

FIG. 5 is a diagram showing an example of curve information according to an embodiment of the present invention.

FIG. 6 is a diagram illustrating a recognition process according to an embodiment of the present invention.

[Explanation of symbols]

 1 figure core line extraction means 2 category information giving means 3 core line pixel string extraction means 4 curve information calculation means 5 feature point extraction means 6 normalization means 7 target figure learning means 8 learning control means 9 normalized curvature calculation means 10 recognition result Output means 11 Learning figure 12 Recognition figure 20 Computer 21 Neural network

Claims (1)

[Claims] 1. A graphic core line extracting means for obtaining a center line of the graphic from a line graphic, a category information assigning means for providing category information of the line graphic, and a graphic core line extracting means for the graphic graphic obtained by the graphic core line extracting means. A core line pixel string extracting means for extracting a center line as a pixel string, a curve information calculating means for obtaining curvature information for expressing the shape structure of the graphic for each pixel of the pixel string, and the curve information calculating means. Input point of the neural network by normalizing the curvature information in the vicinity of the feature point obtained by the feature point extracting means, and a feature point extracting means for extracting a point where the obtained curvature information changes as a feature point. The normalization means for generating the normalization information, the normalized curvature information obtained by the normalization means, and the category information of the target graphic obtained by the category information giving means. A plurality of target figures to the neural network by using a target figure learning means for inputting to the neural network and causing the neural network to learn the target figure, and the figure core extraction means to the target figure learning means. A plurality of figure learning control means, and a recognition figure created by the neural network, and the normalization of the recognition figure by each means of the normalization means from the figure core line extraction means excluding the category information giving means. Normalized curvature calculation means for obtaining the normalized curvature information, the curvature information obtained by the normalized curvature calculation means is input to the neural network, and the recognition result of the recognition graphic from the output information from the neural network. Characterized by having a neural network recognition control means for obtaining information Figure shape learning and recognition system that.