JP2006048370A - Method for recognizing pattern, method for generating teaching data to be used for the method and pattern recognition apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pattern recognition method for generating defective data and compensating shortage to improve a recognition ratio and a teaching data generating method to be used for the pattern recognition method. <P>SOLUTION: An appearance inspection apparatus for performing learning on the basis of many teaching data stored in a teaching file, recognizing a pattern and judging a defect is provided with a teaching data generation device 40 for generating new teaching data by deforming one data, the teaching data generation device 40 generates new teaching data by deforming specific teaching data which has the small number of data out of many teaching data stored in the teaching file 35, replenishes the generated teaching data to the teaching file to recognize defects in the teaching data. When teaching data to be generated are image data, affine transformation including the enlargement, contraction and rotation of an image and attribute conversion including brightness, contrast and edge intensity are executed to generate new teaching data. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a pattern recognition method, a teaching data generation method used therefor, and a pattern recognition apparatus. More specifically, a device that inspects and classifies appearance defects of products and parts by capturing images of printed matter and solid objects (paper, film, metal, etc.) using a camera, and pattern recognition and pattern classification. BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method widely applied to the present invention, a teaching data generation method used therefor, and a pattern recognition apparatus.

  For example, print defects and the like have been inspected by human hands before, but an appearance inspection apparatus using pattern recognition has been introduced and has been effective. However, many of the inspection apparatuses only determine the quality of the object, and are not useful for suppressing or correcting the occurrence of defects. Therefore, there is a demand for the development of an inspection apparatus capable of recognizing a print defect, classifying the defect type, estimating the cause of occurrence, and feeding back the result at an early stage.

By the way, in the case of defect inspection of printed matter, the print defect such as “hole”, “stain”, “convex”, “streaks” as shown in FIG. Since the variation of the feature amount is large, the development of such a defect recognition system is a work that requires a great deal of labor of an image processing specialist.
Therefore, the appearance inspection system can be easily operated even by non-experts, and there is an increasing expectation for a learning type system that automatically configures the system while showing actual examples.

  The technique of Patent Document 1 responds to the above. The feature for each type of printing defect is indicated as a defect description using a medium such as a language or a graphic symbol, and the supported defect description is expressed by a membership function. Fuzzy processing is performed, and the output of the fuzzy reasoning unit that performs this fuzzy processing is input to a hierarchical neural network having the fuzzy reasoning unit as an input layer, and the type of defect obtained by the detection is determined by this network. ing. In this way, by registering the class name for description for each feature of the image in advance, new defect registration can be described in sensory terms, and based on the description, the network structure is designed, Teaching for configuring a network as designed can also be provided. As a result, a high recognition rate could be obtained for several types of defects.

  However, the above prior art shows a sufficient recognition rate for printing defects with high occurrence frequency of defects, but the number of data for the printing defects with low occurrence frequency is lower than other defects, so teaching is sufficient. However, it has been pointed out that the defect recognition rate is lowered.

JP-A-7-333170

  In view of the above circumstances, the present invention provides a pattern recognition method capable of achieving a high recognition rate even when the number of data is small due to low occurrence frequency or the like, and a teaching data generation method and a pattern recognition device used therefor With the goal.

A pattern recognition method according to a first aspect of the present invention is a pattern recognition method that learns and recognizes a pattern based on a large number of teaching data stored in a teaching file, and includes data among a large number of teaching data in the teaching file. The specific teaching data with a small number is characterized by supplementing new teaching data generated by transforming the specific teaching data and recognizing the pattern.
A pattern recognition method according to a second aspect of the present invention is the pattern recognition method according to the first aspect, wherein the pattern recognition method performs pattern classification and there is a bias in the number of teaching data for each classification class. Is modified to generate new teaching data.
When the teaching data to be generated is image data, the teaching data generation method of the third invention performs affine transformation including enlargement, reduction, and rotation of the image and attribute conversion including brightness, contrast, and edge strength. And generating new teaching data.
A pattern recognition device according to a fourth aspect of the present invention is a pattern recognition device that learns and recognizes patterns based on a large number of teaching data stored in a teaching file, and generates new teaching data by modifying one data. Among the many teaching data in the teaching file, the specific teaching data with a small number of data is transformed by the teaching data generating device to change the specific teaching data. The teaching data is generated, and the generated teaching data is supplemented to the teaching file for pattern recognition.
The pattern recognition device according to a fifth aspect of the present invention is the pattern recognition device according to the fourth aspect, wherein the teaching data generation device is built in the pattern recognition device.
According to a sixth aspect of the present invention, there is provided a pattern recognition apparatus according to the fifth aspect, wherein the teaching data generation apparatus exists independently of the pattern recognition apparatus and is connected by a network.

According to the present invention, since the number of data can be increased by supplementing teaching data obtained by transforming the data having a small number of data, a high recognition rate can be achieved at the time of pattern recognition.
Also, since new teaching data can be generated by affine transformation and attribute transformation, deformation data having an appropriate pattern can be obtained, so that deformation data can be easily generated.

Next, an embodiment of the present invention will be described with reference to the drawings.
A feature of the present invention is a teaching data generation technique applied to pattern recognition. The generation technique is a pattern used for appearance inspection such as printing defects and detection of ground defects such as paper, film, and metal. Applied to recognition device. The pattern recognition referred to in this specification refers to a recognition technique widely used for the above inspection. Next, before describing the present invention, an appearance inspection apparatus for printing defects which is an application target will be described.
FIG. 1 is a block diagram of an appearance inspection apparatus which is a type of pattern recognition apparatus to which the present invention is applied. The printed matter m in FIG. 1 is wound from roller to roller, and a plurality of identical printed patterns to be inspected are arranged at regular intervals along the longitudinal direction (flow direction) of the printed matter m. It is printed. The appearance inspection apparatus detects and determines a printing defect on the printed matter m, and includes a camera C and a recognition apparatus A.
The camera C is installed above the printed matter m, for example. The camera C is an image pickup means configured by arranging a plurality of photoelectric conversion elements, for example, CCD (charge coupled device) elements in one row (line shape), and outputs an image pickup signal S to the recognition device A. ing.
The teaching data generation device 40 important in the present invention is connected to or incorporated in the recognition device A, and details thereof will be described later.

The recognition device A includes a processing unit 10 and a determination unit 30.
The processing unit 10 processes a signal (imaging signal S) output from the camera C to extract defects included in each print pattern of the printed matter m, and displays the defects extracted based on the processing results. 20 is displayed.
The determination unit 30 is connected to the processing unit 10 and determines the type of defect based on the processing result in the processing unit 10. A teaching data generation device 40 configured by a personal computer or the like is connected to the determination unit 30. The teaching data generation device 40 also serves as an input unit, and an initial teaching operation for the determination unit 30 is performed. Various operations such as these are performed. Reference numeral 41 denotes a monitor.

A specific configuration of the processing unit 10 will be described with reference to FIG.
The processing unit 10 includes an A / D converter 11, an image memory 12 for one period, a comparison circuit 13, and a different part detection circuit 14. The imaging signal S output from the camera C is input to the A / D converter 11 and converted into (multivalued) digital data for each pixel of the camera C. This data is compared with the image memory 12 for one period. And output to the circuit 13 respectively.

  The image memory 12 for one cycle stores output data (image data) from the camera C corresponding to one cycle of the print pattern. The image stored in the memory 12 is a standard image A (see FIG. 3). As this standard image A, a normal image of a print pattern prepared and registered in advance without any defect is used. When the same print pattern continues, the print pattern may be used.

The comparison circuit 13 stores the image data at the address corresponding to the image data from the camera C input as the real captured image B (see FIG. 3) in the standard image one cycle before stored in the image memory 12 for one cycle. A difference portion detection circuit 14 receives an output from the comparison circuit 13 and detects a difference portion between the standard image A and the captured image B. The difference part is analyzed.
That is, as illustrated in FIG. 3, the processing unit 10 superimposes and compares the standard image A and the captured image B, extracts a difference between the images A and B, and detects a defect among the detected differences. Only the image portion is output to the determination unit 30.

  As shown in FIG. 2, the determination unit 30 includes a feature value extraction circuit 31, a plurality of fuzzy circuits 32, a neural network 33, an initial setting circuit 34, and a teaching file 35.

The feature value extraction circuit 31 converts the defect image extracted by the processing unit 10 and input to the feature value extraction circuit 31 into a binary image, and calculates a feature value for numerical evaluation of the defect image by calculation. Circuit.
The feature value of the defect includes its area (size), circularity, linearity, rectangularity, brightness, contrast, edge strength, and the like.
The size feature value is obtained from the total number of pixels of the defect, the circularity feature value is obtained from the ratio of the area to the perimeter of the defect, and the linearity feature value is a ratio of the major axis length to the minor axis length of the defect. The rectangularity characteristic value is obtained from the ratio of the area of the box circumscribing the defect represented by the product of the major axis length and the minor axis length of the defect and the area of the defect. Further, the brightness feature value determines the brightness at the center of gravity of the defect, and both the contrast feature value and the edge strength feature value are determined for recognizing the relationship with the surroundings.

  The feature value extraction circuit 31 has a number of output terminals corresponding to the type of feature value, and a fuzzy circuit 32 is separately connected to each of these output terminals. These fuzzy circuits 32 constitute an input layer of the neural network 33, have a fuzzy reasoning membership function as an internal function, and output a class value (affiliation value) corresponding to a feature value input thereto. For example, if the area characteristic value of the defect is described as a representative (hereinafter the same), the area is classified into five classes (“large”, “slightly large”, “medium”, “slightly small”, “small”, etc.) Class 1 to class 3,... Class n) are determined according to which class the area feature value input from the feature value extraction circuit 31 corresponds to, and is output to the neural network 33. .

  The neural network 33 has a two-stage configuration. That is, a two-stage neural network (hereinafter abbreviated as “two-stage NN”) has an FNN that classifies defect types in the first stage, and classifies defect types here. In the second stage, there are as many as the number of output layer units of the first stage FNN (hereinafter abbreviated as NN), and only the NN corresponding to the classification result in the first stage operates. Here, levels such as the size of defects are classified. The same input value is used for the first and second stages of NN. Note that this configuration is an example, and an arbitrary configuration may be adopted such that the first stage and the second stage are configured by NN.

  The process flow of the two-stage NN will be described with reference to FIG. First, a filtering process is performed on the defect image to extract a feature value. The feature value is input to the first-stage FNN (101) and operated. As a result, the types of defects are classified (102). Next, the level classification NN corresponding to the defect type classified in the first stage is activated to classify the defect level (103). For example, when the first stage FNN classifies a certain defect image as a “stain” defect, the second stage “stain” defect level classification NN is activated (103). As a result, the defect type and the level of the defect are classified (104).

Details of the above-described neural network 33 will be described.
(Outline of first stage FNN)
(1) Configuration of FNN FNN has a structure having a fuzzy inference part in front of a three-layer perceptron type NN. Note that feature values such as “area”, “brightness”, and “circularity” extracted from the defect image by filtering are used as input values.
(2) Function of the fuzzy reasoning part Since there are no numerical rigorous expressions that humans use, such as “large” and “small”, these are expressed as “large” and “small” by applying fuzzy reasoning. Replace with words such as
The fuzzy inference unit is directly connected to the NN input layer, and converts it into a class value into a feature expression by a membership function. This function is generated based on the boundary value between each feature description set in advance.
(3) Intermediate layer unit In FNN, the number of intermediate layer units is determined by the number of feature representations of defects. Further, before teaching the whole by the back propagation method, the intermediate layer is taught by the initial teaching, and the unit configuration direction of the intermediate layer is designated. The initial teaching indicates that teaching data is also given to the intermediate layer to teach the intermediate layer from the input layer and the output layer from the intermediate layer.
(4) Output layer units The number of units in the output layer corresponds to the number of types of defects. In accordance with the type of defect determined here, the level classification NN dedicated to the defect is activated.

(Outline of second level classification NN)
(1) Configuration of level classification NN The second level classification NN has a three-layer structure of an input layer, an intermediate layer, and an output layer. First, only one second level classification NN corresponding to the first level classification result is activated. The activated defect level classification NN receives the same data as the input data used in the first stage, and classifies the defect level.
The reason why the FNN is not used in the second stage is that the level classification NN in the second stage needs to classify the level in one type of defect, and therefore requires an accurate value.
(2) Operation of input layer of level classification NN The same value as the input value used in the first stage FNN is used as the input value of the level classification NN of each defect. Therefore, the number of units in the input layer is the same as the number of input values. However, since the range of each input value is different, it cannot be input as it is. Therefore, here, normalization is performed so that the input value is in the range of 0 to 1 by dividing by the maximum value taken by each input value.
(3) Operation of output layer of level classification NN The number of units of the output layer is the same as the maximum number of levels of defects to be level classified. The result of classification here is the classification result of the two-stage NN.

  As described above, the determination unit 30 is a fuzzy neural network formed by combining a fuzzy inference unit and a neural network, and this network is a three-layer hierarchical neural network. In this hierarchical neural network, the input layer is formed by each fuzzy circuit 32 as a fuzzy inference unit, the intermediate layer is formed by each preceding circuit of the network corresponding to the defect description of the defect, and the output layer corresponds to the type of defect. Network post-stage circuit that outputs the output.

  As shown in FIG. 2, the post-stage circuit of the neural network 33 has a number of output terminals 42a, 42b, 42c... 42n corresponding to the feature values of the defects. A numerical value indicating the degree of affiliation is output. Thus, since the numerical values output from the output terminals 42a, 42b, 42c,... 42n are 0 to 1.0, it can be determined that the closer the output value is to 1.0, the higher the possibility of actually being the defect type. For example, the output end 42a is a defect due to ink splash, the output end 42b is a defect due to a doctor flaw, the output end 42c is a defect due to drowning, the output end 42d is a defect due to insect mixing, and the output end 42e is other than a bug. The output end 42f is allocated to a defect due to dirt, etc., due to a defect due to dust contamination.

Since the determination unit 30 is configured as described above, if the weight values of the weighting circuits of the network front-stage circuit and the rear-stage circuit are learned by the initial teaching, the defect determination as expected can be performed.
Note that the type of feature and the number of classes selected as described above are associated with the configuration of the determination unit 30, but the association needs to be performed by learning. This learning is performed in two stages, initial teaching and in-service teaching (back propagation method normally performed in a neural network). The “association” is performed during the initial teaching process. Is called.

In the initial teaching, the defect description about the defect is taught to the unit of the intermediate layer so that it reacts only to the corresponding feature, and at the same time, the output according to the teaching signal is set in the intermediate layer, and the unit of the output layer is set. Teach the judgment result.
The initial teaching and the teaching during operation are performed using the defect images stored in the teaching file 35 shown in FIGS. A high defect recognition rate can be achieved if there are a sufficient number of defect images stored in this teaching file for both frequent and infrequent defects. As described above, the defect recognition rate also decreases because of the small number.

  Therefore, in order to increase the recognition rate of defects with low occurrence frequency, the number of data of the defect is increased. However, the method is the present invention, and defect data is generated from feature descriptions of defects with low occurrence frequency. Features.

Next, a method for generating teaching data according to the present invention will be described.
As shown in FIG. 5, first, a plurality of defect images to be generated are selected (201). Of course, the selected defect image is an image of a defect having a low occurrence frequency. Next, the feature of the defect is described (202). Next, defect generation is performed from the description contents (203). The last defect image generated is presented to a person and the correct defect image is selected (204).

In the following, since the frequency of occurrence of “streaks” defects among printing defects is low, a teaching data generation method for these streaks will be described with reference to FIGS. 6 and 7.
(Defect feature description step 202)
As the features used for the description, six types of “brightness”, “edge”, “length”, “width”, “area”, and “direction” are used. For example, the teaching image generating apparatus 40 converts the original image indicated by the symbol X in FIG. 6 into the screen indicated by the symbol Z, “the brightness is dark, the edge is clear, the length is slightly short, the width is wide, the direction Defects are generated by writing “Undefined”.
(Defect generation step 203)
When described in this way, a generated image is generated as indicated by a symbol Y in FIG. When the deformation process is performed based on the generated defect image, for example, as shown in FIG. 7, a large amount of image data that greatly compensates for the shortage is obtained.

（３）長さ・幅・面積・方向処理
長さ・幅処理は画像の縦と横に対してアフィン変換を用いた拡大縮小処理を行えばよい。ただし、このとき欠陥の面積が記述で指定された範囲を越えないようにする。方向処理も同様にアフィン変換を用いて回転処理を行えばよい。 Here, details of the feature description step 202 and the defect generation step 203 will be described.
(1) Brightness processing Brightness processing is expressed by the following equation.
g (x, y) = Th + s (f (x, y) −Th) (1)
Here, f is the luminance value of the input image, g is the luminance value of the output image, and Th is the average value of the luminance of 5 pixels around the input image. The range of values to be taken is determined by the description contents such as “very bright” and “dark”, and s is randomly selected from the range.
(2) Edge processing Edge processing differs depending on whether it is emphasized or blurred.
In the process of emphasizing, first, an image in which the maximum value is extracted from the target pixel and its surrounding nine points and an image in which the minimum value is extracted are created. Let these be f _max and f _min , respectively. The following processing is performed by comparing the two images with the original image.
When f _max (x, y) −f (x, y)> f (x, y) −f _min (x, y) g (x, y) = f (x, y) −s (f (x , Y) −f _min (x, y)) (2)
When f _max (x, y) −f (x, y) ≦ f (x, y) −f _min (x, y) g (x, y) = f (x, y) + s (f _max (x , Y) -f (x, y)) (3)
When blurring, use the following formula.

(3) Length / Width / Area / Direction Processing In length / width processing, enlargement / reduction processing using affine transformation may be performed on the vertical and horizontal sides of an image. At this time, however, the defect area should not exceed the range specified in the description. Similarly, the direction processing may be performed by rotating using affine transformation.

In the teaching data generation method of the present invention, it is sufficient if the insufficient teaching data is completed in steps 201 to 203 shown in FIG. 5, and if unrealistic defect data is generated only by computer processing, it is unqualified artificially. A step (204) for removing unnecessary defect data may be included.
If the deficient defect data D completed in this way is accumulated in the teaching file 35 shown in FIG. 3, it will be described later according to teaching contents based on a sufficiently large number of defect patterns even for defects with low occurrence frequency. As a result, the type and level of the defect can be recognized with high accuracy.

The teaching data generation device 40 for generating the teaching data takes in image data as an original image of the defect, and based on the feature description, brightness processing, edge processing, length / width / area / direction It suffices to have a function for processing.
The teaching data generation device 40 may be incorporated in the appearance inspection device as shown in FIG. 2, but is configured as an independent device and is connected to the separately provided appearance inspection device to transmit data. Also good.

Next, the effectiveness when the teaching data is supplemented according to the present invention will be described using experimental data.
(1) Experimental method As input and teaching data, data obtained by actually classifying defects on the production line was used. The number of defect types used in the data is 28 (convex defect 5 level, hole defect 5 level, mura defect 4 level, concave defect 5 level, concave defect (black), streak defect 5 level, dirt defect 3 level). However, only one streak defect was present in all data, and only one streak defect level 3 was present, so 24 types of defects were usable. And if there are more than 250 defect data, 50, 5 or more, 1/5 of the number of data, and if less than 5, create 1 teaching data did. Then, defect generation was performed on the extracted data with less than 50 pieces to make 50 pieces. As a result, the teaching data is 7 types of 1200 data when the defect is generated, and 7 types of 652 data when the defect is not generated. Details of the teaching data are shown in Table 5 of FIG.
Regarding the number of teachings, the first stage FNN performed initial teaching 500 times and backpropagation 5000 times. The second-stage NN performed backpropagation 5000 times for each network. After that, the learned NN was operated on the teaching data and all the data, and it was examined whether or not it matched with the result classified by the original person.
The number of units in the NN intermediate layer is the same as the number of units in the input layer. In addition, the number of divisions of the FNN input is 4, and the value used for the fuzzy function is the maximum value of the input / 5 × n (n = 1 to 4).
For comparison, recognition rates were examined for the following four NN configurations.
Configuration 1: Two-stage NN and defect generation Configuration 2: Two-stage NN and defect generation are not performed Configuration 3: Only one FNN stage and defect generation are performed Configuration 4: Only one FNN stage and defect generation are not performed Defect level Considering that it may be difficult for humans to specify the level depending on the type, we decided to show data when the level is deviated by ± 1 and it is considered as a correct answer.
(2) Results The experimental results of configurations 1 to 4 are shown in Tables 1 to 4 in FIGS.
(3) Consideration Comparing the experimental results of the configurations that generated defects (Tables 1 and 3) and the configurations that did not generate defects (Tables 2 and 4), the recognition of the entire data when defects were generated The rate is decreasing (from 97.38% in Table 2 to 97.2% in Table 1 and from 78.93% in Table 4 to 77.48% in Table 3), but the average recognition rate for each defect is It can be seen that it has improved (from 72.38% in Table 2 to 88.21% in Table 1, from 58.36% in Table 4 to 71.95% in Table 3). The reason why the recognition rate with respect to the entire data has decreased is that the recognition rate of defects with a small number of data has improved, but the recognition rate of defects with a large number of data has decreased. The reason why the average recognition rate is improved is considered to be that the recognition improvement rate of the defect having a small number of data is larger than the recognition reduction rate of the defect having a large number of data. Further, only one streak defect in the data used this time can be recognized when a defect is generated, and cannot be recognized when no defect is generated. In an actual production line, it is considered that the frequency of occurrence of a fatal defect is lower, so that it is considered effective to generate a defect that improves the recognition rate of a defect having a lower frequency of occurrence.
Next, when FNN and defect generation are combined (Table 3) and when two-stage NN and defect generation are combined (Table 1), the recognition rate and the average recognition rate for the entire data are compared, and FNN generates defects. However, the recognition rate for the entire data has decreased by about 1.5% (down from 78.93% in Table 4 to 77.48% in Table 3), while the second-stage NN has only decreased by about 0.4% (Table 2). From 97.38% to 97.02% in Table 1). In addition, when FNN generates defects, the average recognition rate has improved by about 14% (from 58.36% in Table 4 to 71.95% in Table 3), whereas 2-stage NN has improved by about 16%. (Up from 72.38% in Table 2 to 88.21% in Table 1) From this, it is considered that combining the two-stage NN and defect generation can suppress a decrease in the recognition rate with respect to the entire data and greatly improve the average recognition rate for each defect.
From the experimental results in Tables 2 and 4, the recognition rate for the entire data of the two-stage NN is about 97% (see Table 2), while the recognition rate for FNN is about 79% (see Table 4). From this, it can be confirmed that the two-stage NN shows a higher recognition rate than the FNN.

  The present invention is an apparatus and method for imaging an object using a camera and performing pattern recognition, for example, to an apparatus for inspecting and classifying appearance defects of products and parts regardless of the purpose or use thereof. Widely applicable. Therefore, it can be applied to the appearance inspection of printed matter and the inspection of the ground such as paper, film and metal.

1 is a block diagram of an appearance inspection apparatus A to which a teaching data generation apparatus according to the present invention is applied. It is a block diagram of the process part 10 and the determination part 30 of the external appearance inspection apparatus A in FIG. It is function explanatory drawing of the process part 10 and the determination part 30 in FIG. FIG. 3 is an explanatory diagram of a processing flow of the neural network 33 in FIG. 2. It is a flowchart of the teaching data generation method in this invention. It is explanatory drawing of the defect feature description step in a teaching data generation method. It is explanatory drawing of the defect production | generation step in a teaching data production | generation method. It is a table | surface which shows the recognition rate after teaching data replenishment, and is a table | surface in the case of 2 steps | paragraphs NN and teaching data generation. It is a table | surface which shows the recognition rate after teaching data replenishment, and is a table | surface when there is no 2nd stage NN and teaching data generation. It is a table | surface which shows the recognition rate after teaching data supplementation, and is a table | surface in the case of only one stage of FNN and teaching data generation. It is a table | surface which shows the recognition rate after teaching data replenishment, and is a table | surface when only FNN 1 step | paragraph and there is no teaching data generation. It is a table | surface which shows all the teaching data. It is explanatory drawing which shows a printing defect.

Explanation of symbols

C camera 10 processing unit 30 determination unit 31 feature value extraction circuit 32 fuzzy circuit 33 neural network 35 teaching file 40 teaching data generation device

Claims (6)

A pattern recognition method for learning and pattern recognition based on a large number of teaching data stored in a teaching file,
Of the large number of teaching data in the teaching file, specific teaching data having a small number of data is supplemented with new teaching data generated by modifying the specific teaching data, and pattern recognition is performed. Pattern recognition method. When the pattern recognition method performs pattern classification and there is a bias in the number of teaching data for each classification class, the teaching data with a small number of data is transformed to generate new teaching data. The pattern recognition method according to claim 1, wherein: When the teaching data to be generated is image data,
Affine transformation including image enlargement, reduction, rotation,
By performing attribute conversion including brightness, contrast, and edge strength,
A teaching data generation method characterized by generating new teaching data. A pattern recognition device that learns based on a large number of teaching data stored in a teaching file and performs pattern recognition,
A teaching data generation device that generates new teaching data by transforming one data;
Among the many teaching data in the teaching file, for the specific teaching data with a small number of data, the teaching data generation device transforms the specific teaching data and generates new teaching data. A pattern recognition apparatus for recognizing a pattern by supplementing the taught data with the taught data. The pattern recognition apparatus according to claim 4, wherein the teaching data generation apparatus is built in the pattern recognition apparatus. 5. The pattern recognition apparatus according to claim 4, wherein the teaching data generation apparatus exists independently of the pattern recognition apparatus and is connected by a network.

Images