JPH0729010A - Shape classification device - Google Patents

Abstract

PURPOSE:To classify a fuzzy shape, to improve classification and to shorten time required for classification. CONSTITUTION:The learning processing means 10 of a storage/processing part 1 executes a learning processing as against a classification processing in a classification processing means 11 based on standard pattern data which is read from the standard pattern data file 2a of a disk part 2. The learning processing means 10 outputs a neural net weight value used for the classification processing, which is obtained from the learning processing, to the classification processing means 11. The classification processing means 11 classification- processes data being the object of classification, which is read from the classification object data file 2b of the disk part 2, by using the neural net weight value obtained from the learning processing means 10.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a shape classification device, and more particularly to a CAD (Computer Aided Design).
n) relates to a device for classifying the shape data created.

[0002]

2. Description of the Related Art Conventionally, in a classification method for classifying shape data of this kind, a typical method is GT (Grou).
p Technology) method is available. In this GT method, the shape attributes of the shape data are classified into large classification, middle classification,
Level-specific attributes such as small classification are set, and the classification code is set uniquely by the combination of each code.

For example, a sheet metal shape as shown in FIG.
When the classification is performed by the method, the classification attributes are each code as shown in FIG. That is, the codes are the presence / absence of bending, the type of bending, the bending R, the number of bendings, the number of bending peaks, and the number of bending troughs.

An example of the technique of the classification system using the GT method as described above is a "production management system by GT".
(Nikkan Kogyo Shimbun, 1981, 6) p. 13-15
Are detailed in.

[0005]

In the shape classification by the conventional GT method described above, since the level-based attributes are set and the classification code is set by the combination of each code, it is difficult to set the attribute parameter. There is a problem that the number of attribute parameters increases with the sophistication of classification accuracy. Further, there is a problem that the classification result differs from the operator's judgment result due to the encoding.

Therefore, an object of the present invention is to solve the above problems, to provide an ambiguous shape classification, and to provide a shape classification device capable of improving classification accuracy and shortening the time required for classification. is there.

[0007]

A shape classification device according to the present invention is a holding means for holding a standard pattern preset for classifying graphic data to be classified displayed in a three-dimensional shape and its identification information. And a binarizing means for dividing the graphic data into a mesh and binarizing it for each block, and a classification for classifying which standard pattern the graphic data corresponds from the binarized data of the binarizing means. And a means for adjusting the classification processing of the classification means so that the classification means outputs the identification information corresponding to the standard pattern as a classification result during the classification processing for the standard pattern read from the holding means. .

[0008]

An embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention. In the figure, the storage / processing unit 1 of the shape classification device is composed of a learning processing means 10 and a classification processing means 11.

The learning processing means 10 performs learning processing for the classification processing in the classification processing means 11 based on the standard pattern data read from the standard pattern data file 2a of the disk unit 2. The learning processing means 10 outputs to the classification processing means 11 the neural network weight value used in the classification processing obtained from the learning processing.

The classification processing means 11 uses the neural network weight value obtained from the learning processing means 10 to perform classification processing on the classification target data read from the classification target data file 2b of the disk unit 2. Classification processing means 11
Writes the result of the classification process in the classification output file 2c.

The classification processing by the classification processing means 11 includes a mesh division processing for dividing the classification target data into meshes, and a binarization processing for binarizing each block of the classification target data divided by the mesh division processing. It is composed of a process of classifying, by a perceptron, which standard pattern corresponds to the binarized data.

The standard pattern data file 2a contains a standard pattern prepared for learning in advance and its pattern I.
D and D are stored as standard pattern data. The classification target data file 2b also stores patterns of classification targets to be classified as classification target data.

Here, in the shape classification apparatus according to the embodiment of the present invention, the man-machine interface section between the storage / processing section 1 is composed of the CRT 3 and the keyboard 4.

FIG. 2 is a diagram showing an example of standard patterns stored in the standard pattern data file 2a of the disk unit 2 of FIG. In the figure, each of the standard patterns 1 to 3 indicates a three-dimensional pattern.

FIG. 3 is a diagram showing a processing result by the classification processing means 11 of FIG. FIG. 3A is a diagram showing the classification target data that has been mesh-divided, and FIG. 3B is a diagram showing the binarized classification target data.

The classification processing means 11 mesh-divides the pattern A of the classification target data into a plurality of fine blocks (square grid) (16 × 16 = 256) as shown in FIG. 3A.
Blocks).

After that, the classification processing means 11 determines whether or not each of the plurality of divided fine blocks is based on the presence or absence of an entity.
Quantify. In this case, if the block has an entity, "1" is set, and if the block has no entity, "0" is set.

FIG. 4 is a diagram showing the classification processing by the perceptron of the classification processing means 11 of FIG. In the figure, when binarized classification target data is input to each of the nodes 1-1 to 1-256 of the input layer, each node of the hidden layer 2-
1 to 2-4 calculate the value of each of the nodes 1-1 to 1-256 in the input layer.

The nodes 3-1 to 3-4 of the output layer calculate the values of the nodes 2-1 to 2-4 of the hidden layer, and output the pattern ID as the calculation result. Each hidden layer node 2
-1 to 2-4 and the nodes 3-1 to 3-4 of the output layer perform the calculation using the neural network weight value.

FIG. 5 is a flow chart showing a learning process for the classification process in the embodiment of the present invention. A learning process for the classification process according to the embodiment of the present invention will be described with reference to FIGS.

When the shape classifying device is activated, the learning processing means 10 in the storage / processing unit 1 starts the learning process for the classification process. The learning processing means 10 first initializes the pattern ID and the weight value (step 2 in FIG. 5).
1).

Next, the learning processing means 10 reads the standard pattern data of the initialized pattern ID value from the standard pattern data file 2a of the disk unit 2 (step 22 in FIG. 5). At this time, the standard pattern of the three-dimensional shape read from the standard pattern data file 2a is output to the CRT 3 and displayed.

The learning processing means 10 divides the standard pattern displayed on the CRT 3 into a plurality of fine blocks by mesh division processing, and performs binarization processing based on the presence or absence of the substance (step 23 in FIG. 5) (see FIG. 3). ).

The learning processing means 10 converts the binarized data of the standard pattern into the nodes 1-1 to 1- of the input layer of the perceptron.
256 (step 24 in FIG. 5), and the pattern I is calculated by the hidden layer and the output layer of the perceptron.
D is obtained (see FIG. 4).

The learning processing means 10 checks whether the pattern ID obtained from the output layer of the perceptron is the same as the pattern ID of the standard pattern read in step 22 (FIG. 5).
Steps 25, 26).

If the pattern ID obtained from the output layer of the perceptron is the same as the pattern ID of the standard pattern read in step 22, the learning processing means 10 updates the pattern ID to read the next standard pattern ( I = I +
1) (FIG. 5, step 27).

The learning processing means 10 determines whether all standard patterns have been read by updating the pattern ID (see FIG. 5).
If the reading of all the standard patterns is completed in step 28), the learning process is ended. If the reading of all standard patterns is not completed, the process returns to step 22 to learn the next standard pattern.

On the other hand, if the pattern ID obtained from the output layer of the perceptron is not the same as the pattern ID of the standard pattern read in step 22, the learning processing means 10 determines the weight of each layer of the perceptron based on the error of the pattern ID. Change (step 29 in FIG. 5).

When 0 is output when the output result becomes 1, the weight before change is Wi (t), the weight after change is Wi (t + 1), and the input corresponding to the weight is Xi ( t), where α is a constant, the changed weight Wi (t + 1) is obtained from the equation: Wi (t + 1) = Wi (t) + αXi (t). However, the constant α is 0 <α ≦ 1 and controls the weighting speed.

Further, when 1 is output when the output result becomes 0, the changed weight Wi (t + 1) is obtained from the equation Wi (t + 1) = Wi (t) -αXi (t). The weight can be changed by a method other than the above method.

After changing the weight, the learning processing means 10 causes the pattern ID of the standard pattern whose weight has been changed to be the same as the pattern ID of the standard pattern read first (I =
1) It is determined (step 30 in FIG. 5).

As a result, if the same as the pattern ID of the standard pattern read first, the learning processing means 10 returns to step 24 to convert the binarized data of the standard pattern into the nodes 1-1 to 1-1 of the input layer of the perceptron. The pattern IDs are input to 1-256, respectively, and the pattern ID is obtained by calculation of each of the hidden layer and the output layer of the perceptron.

Pattern I of the standard pattern read first
If it is not the same as D, the learning processing means 10 uses the pattern ID.
Are initialized, that is, I = 1 is set, the process returns to step 22, and the first standard pattern is learned again.

In this way, the learning processing means 10
In order to classify the graphic data to be classified displayed in the three-dimensional shape, the classification process of the classification processing means 11 is learned in advance using the standard pattern stored in the standard pattern data file 2a in advance. Can eliminate the classification attribute code found in the conventional GT method. Therefore, ambiguous shape classification can be performed, the classification accuracy can be improved, and the time required for classification can be shortened.

[0036]

As described above, according to the present invention, 3
A binary pattern obtained by holding a standard pattern preset for classifying the graphic data to be classified displayed in the three-dimensional shape and its identification information, dividing the graphic data into meshes, and binarizing each block. In the classification process of the classification unit that classifies which standard pattern the graphic data corresponds to from the digitized data, the classification unit outputs the identification information corresponding to the standard pattern as the classification result during the classification process for the standard pattern. Thus, it is possible to perform ambiguous shape classification, improve classification accuracy, and shorten the time required for classification.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a standard pattern stored in a standard pattern data file of a disc section of FIG.

FIG. 3 (a) is a diagram showing mesh-divided classification target data, and FIG. 3 (b) is a diagram showing binarized classification target data.

FIG. 4 is a diagram showing classification processing by a perceptron of the classification processing means in FIG.

FIG. 5 is a flowchart showing learning processing for classification processing according to an embodiment of the present invention.

FIG. 6 is a diagram showing an example of classification by the GT method of the conventional example.

[Explanation of symbols]

 1 Storage / Processing Unit 2 Disk Unit 2a Standard Pattern Data File 2b Classification Target Data File 2c Classification Output File 10 Learning Processing Means 11 Classification Processing Means

Claims (3)

[Claims] 1. A holding unit for holding a preset standard pattern for classifying graphic data to be classified displayed in a three-dimensional shape and identification information thereof, and dividing each graphic data into meshes into blocks. Binarizing means for binarizing for each, a classifying means for classifying which standard pattern the graphic data corresponds to from the binarized data of the binarizing means, and the standard pattern read from the holding means. The shape classifying device, wherein the classifying unit adjusts the classifying process of the classifying unit so that the classifying unit outputs identification information corresponding to the standard pattern as a classification result during the classifying process. 2. The shape classification device according to claim 1, wherein the classification means is configured to classify the graphic data by a perceptron. 3. The shape classification apparatus according to claim 2, wherein the classification processing of the classification means is adjusted by changing a neural net weight value used for the perceptron.