Feature amount selection method and apparatus

Abstract

The present invention provides a feature quantity selection method that selects a feature quantity having a high contribution degree for each defect pattern using minimum expert data and a small number of non-expert data. The feature quantity selection method includes a step of extracting a multi-dimensional feature quantity from expert data including images of various defect shapes; a step of extracting a multi-dimensional feature quantity from non-expert data including images of limited defect shapes; a step of calculating a standard deviation of each dimension of the extracted feature quantity of the expert data and the non-expert data; a step of calculating a ratio of the standard deviation of each dimension of the calculated expert data to the standard deviation of each dimension of the calculated non-expert data; and a step of selecting a given number of standard deviation ratios in descending order of value from the calculated standard deviation ratios, and selecting a feature quantity associated with the selected standard deviation ratio as a feature quantity having a high contribution degree for a specific defect pattern.

Description

Technical Field

[0001] This invention relates to a feature selection method and a feature selection device for use in a learning model for determining the quality of an object to be inspected, such as an inspection device that has machine learning capabilities using neural networks. Background Technology

[0002] In recent years, the development of automated inspection technologies has been advancing, utilizing inspection devices with machine learning capabilities based on neural networks to determine whether various industrial products and components are normal (qualified) or abnormal (unqualified). In such inspection devices, the process learns from image data of the appearance of numerous inspected items already classified as qualified or unqualified, using this data as teacher data. Furthermore, by learning from these classification criteria, the inspection device can classify new inspected items captured by a camera as either qualified or unqualified.

[0003] In the learning of the classification model in such an inspection device, techniques such as SIFT (Scale-Invariant Feature Transform) and CNN (Convolutional Neural Network) are used to extract features from the images of the teacher data, find the common patterns and consistency of normal or abnormal products, and thus classify normal and abnormal products with high accuracy.

[0004] If we consider using images of anomalous products as teacher data to enable a classification model to learn the shapes of various defects, then for example, in the case of castings, defects such as porosity, scratches, and dents exist in several categories (defect patterns) depending on the product characteristics and manufacturing method. Typically, the shape of a defect varies considerably depending on each defect pattern. Therefore, when extracting features from teacher data of anomalous products, understanding the features specific to each defect pattern can sometimes contribute to building a high-accuracy classification model.

[0005] Furthermore, regarding the collection of images of abnormal items, one can consider collecting data selected by experts such as skilled workers and operators with many years of experience (hereinafter referred to as "expert data"), and collecting data selected by non-experts such as newcomers and operators with few years of experience (hereinafter referred to as "non-expert data"). The former requires long-term constraints on skilled workers to obtain a sufficient amount of teacher data, resulting in poor cost efficiency. The latter tends to yield average types and shapes of abnormal items, making it difficult to accurately classify low-probability abnormal items when using only teacher data. Therefore, it is important to collect high-quality teacher data while controlling costs by combining a minimum amount of expert data with easily collectable non-expert data.

[0006] Previously, random forests, as a type of ensemble learning, were known as a method for calculating the importance (contribution) of explanatory variables (features) in classification models. In the learning of classification models in inspection devices, techniques for using random forests to filter features with high contribution rates are known, for example, the technique described in Patent Document 1.

[0007] In Patent Document 1, a temporary classifier is generated using a random forest that uses multiple features as explanatory variables, based on multiple pre-prepared training data sets. Features with low contribution to classification based on this temporary classifier are then identified. Next, a new temporary classifier is generated again using a random forest based on multiple training data sets from which these low-contribution features have been removed. The correct answer rate of the classification is then compared with that of the previously generated temporary classifier. By repeating this process multiple times, a classifier that uses only features with high contribution to classification is finally generated.

[0008] Prior art literature

[0009] Patent documents

[0010] Patent Document 1: Japanese Patent Application Publication No. 2016-109495 Summary of the Invention

[0011] The problem that the invention aims to solve

[0012] However, in order to extract high-contribution features from a classification model with high accuracy using random forests, sufficient quality and quantity of teacher data need to be prepared in advance. Even if random forests are used to extract high-contribution features before sufficient learning data is available, the possibility that these features will not contribute much in a high-accuracy classification model increases.

[0013] Therefore, in situations such as a manufacturing line in its early stages of operation where sufficient teacher data is not available, and in order to efficiently build a high-accuracy classification model, it is necessary to extract features with significant impact for each defect pattern. Random forests cannot be used to extract features in such cases. Therefore, it is necessary to develop new methods that can select features with high contribution for each defect pattern from a small number of defective product images at the beginning of the production line.

[0014] This invention was made to address this problem, and its purpose is to provide a feature selection method that can select the feature quantities with high contribution for each defect pattern using a minimum amount of expert data and a small amount of non-expert data.

[0015] Methods for solving problems

[0016] To achieve this objective, the feature selection method of technical solution 1 of the present invention is based on expert data that includes images containing various defect shapes and is classified according to each defect pattern representing the type of defect, and non-expert data that includes images containing a limited number of defect shapes and is classified according to each defect pattern. The method selects features that contribute highly to a specific defect pattern. Its characteristic is that it includes: an expert feature extraction step, which extracts expert data corresponding to a specific defect pattern and extracts multi-dimensional features from the extracted expert data. Figure 3 Step 4); Non-expert feature extraction process: extract non-expert data corresponding to specific defect patterns, and extract multi-dimensional features from the extracted non-expert data. Figure 3 Step 5); Expert standard deviation calculation process, calculating the standard deviation of each dimension of the extracted expert data features ( Figure 3 Step 6); Non-expert standard deviation calculation process, calculate the standard deviation of each dimension of the feature quantities of the extracted non-expert data ( Figure 3 Step 6); Standard deviation ratio calculation process: The ratio of the standard deviation of each dimension of the feature quantity of the expert data to the standard deviation of each dimension of the feature quantity of the non-expert data is used as the standard deviation ratio. Figure 3 Step 7); and the feature selection process, in which a given number of features are selected from the calculated standard deviation ratios in descending order of value, and the features associated with the selected standard deviation ratio are selected as those that contribute highly to the specific defect mode. Figure 3 Step 9).

[0017] In this feature selection method, data corresponding to the desired defect pattern is extracted from a small amount of expert data. Multi-dimensional features are then extracted from this data, and the standard deviation of each dimension is calculated. Here, the expert data is, for example, data pre-selected by a skilled operator. It is data that comprehensively includes patterns of various possible defect shapes for each defect pattern. Therefore, the features extracted from it have a distributed state that is diffused without omission, and tend to have a large standard deviation.

[0018] On the other hand, non-expert data, such as data selected and accumulated sequentially by inexperienced operators on the manufacturing line, is assumed to contain a large number of data with unclear corresponding defect shapes and limited to average defect shapes that are relatively easy to detect. Therefore, the features extracted from the relatively small amount of non-expert data in the initial collection phase become a distribution pattern biased towards the mean, with a tendency for the standard deviation to decrease.

[0019] This tendency warrants further specific research. For example, in the case of visual inspection of castings, features such as brightness, shading period, and positional information might be considered as characteristics of the object being inspected. However, if these are defined according to each pixel of an image or the position of each rectangular region in a scanned image, it is not uncommon to obtain feature quantities exceeding 1000 dimensions. On the other hand, typically, the number of feature quantities that are important (highly contributing) for classifying defect shapes of each type (defect mode) with multiple defects is limited, and they can be considered to differ according to each defect mode.

[0020] Here, for example, when a feature contributes little to the desired defect pattern, the distribution of the extracted feature is narrower in both expert data containing various defect shapes and non-expert data that can be considered to contain only average defect shapes, and the difference in their standard deviations is also smaller. Therefore, the ratio of the standard deviations in the dimension of this feature is also smaller. On the other hand, when a feature contributes greatly to the desired defect pattern, the distribution of the feature extracted from expert data containing various defect shapes is wider, and the standard deviation becomes larger. On the other hand, the distribution of the feature extracted from non-expert data containing only average defect shapes is narrower, and the standard deviation does not become as large. Therefore, the difference in the standard deviations in the two can be considered larger. Therefore, the ratio of the standard deviations in the dimension of this feature can be considered larger.

[0021] Based on the above insights, in this invention, data corresponding to the desired defect pattern is extracted from a relatively small amount of collected non-expert data. From this data, feature quantities of the same dimension as those extracted from expert data are extracted, and the standard deviation of each dimension is calculated. Then, the ratio of the standard deviation of each feature quantity calculated based on the expert data to the standard deviation of each feature quantity calculated based on the non-expert data for each dimension is calculated as the standard deviation ratio. Then, a given number of feature quantities associated with the selected standard deviation ratios are chosen as those with high contributions to the specific defect pattern. Therefore, through this invention, it is possible to select high-contribution feature quantities for each defect pattern using a minimal amount of expert data and a small amount of non-expert data.

[0022] The feature of the invention involved in technical solution 2 of the present invention is that, in the feature quantity selection method described in technical solution 1, in the feature quantity selection step, a given number is selected from the standard deviation ratios that exceed a given threshold among the calculated standard deviation ratios, in descending order of value, and the feature quantity associated with the selected standard deviation ratio is selected as the feature quantity that contributes highly to the specific defect mode.

[0023] According to this structure, since only the standard deviation ratios exceeding a given threshold among the calculated standard deviation ratios are selected, it is possible to choose the feature quantities that contribute more to a specific defect pattern. Therefore, it is possible to select the feature quantities that contribute more to each defect pattern using a minimal amount of expert data and a small amount of non-expert data.

[0024] The invention involved in technical solution 3 is characterized in that, in the feature quantity selection method described in technical solution 1, the defect mode includes at least one of porosity, scratch, dent, and residual chips.

[0025] According to this structure, since it includes at least one of porosity, scratches, dents, and residual chips as a defect mode, it is particularly possible to select the feature quantity with high contribution for each defect mode using a minimum amount of expert data and a small amount of non-expert data, especially in images of defective castings.

[0026] The invention involved in technical solution 4 of the present invention is characterized in that, in the feature quantity selection method described in technical solution 1, the expert data includes a generated image generated based on the actual image of the defect shape.

[0027] According to this structure, since the minimum amount of expert data includes not only actual images of the defect shape but also generated images based on those actual images using image generation techniques, the number of actual images required to constrain skilled operators can be further reduced. This, in turn, can further reduce the cost associated with collecting expert data.

[0028] The feature selection device involved in technical solution 5 of the present invention is a feature selection device that selects features that contribute highly to a specific defect pattern based on expert data that includes images containing various defect shapes and is classified according to each defect pattern representing the type of defect, and non-expert data that includes images containing a limited number of defect shapes and is classified according to each defect pattern. The device is characterized by comprising: an expert feature extraction unit (feature extraction section 15 in the embodiment, hereinafter the same) that extracts expert data corresponding to a specific defect pattern and extracts multi-dimensional features from the extracted expert data; a non-expert feature extraction unit (feature extraction section 15) that extracts non-expert data corresponding to a specific defect pattern and extracts the multi-dimensional features from the extracted non-expert data; and expert standards. The system includes a deviation calculation unit (standard deviation calculation unit 16), which calculates the standard deviation of each dimension of the feature quantity of the extracted expert data; a non-expert standard deviation calculation unit (standard deviation calculation unit 16), which calculates the standard deviation of each dimension of the feature quantity of the extracted non-expert data; a standard deviation ratio calculation unit (standard deviation ratio calculation unit 17), which calculates the ratio of the standard deviation of each dimension of the feature quantity of the expert data to the standard deviation of each dimension of the feature quantity of the non-expert data as the standard deviation ratio; and a feature quantity selection unit (feature quantity selection unit 18), which selects a given number of features from the calculated standard deviation ratios in descending order of value, and selects the feature quantities associated with the selected standard deviation ratios as the features quantities that contribute highly to the specific defect pattern.

[0029] In this feature selection device, data corresponding to the desired defect pattern is extracted from a small amount of expert data. Multi-dimensional features are then extracted from this data, and the standard deviation for each dimension is calculated. Similarly, data corresponding to the desired defect pattern is extracted from a relatively small amount of collected non-expert data. Features of the same dimension as those extracted from the expert data are extracted from this data, and the standard deviation for each dimension is calculated. Then, the ratio of the standard deviation calculated based on the expert data to the standard deviation calculated based on the non-expert data for each dimension is calculated as the standard deviation ratio. A given number of features associated with these standard deviation ratios, in descending order of value, are selected as those that contribute significantly to the specific defect pattern. Therefore, this invention enables the selection of highly contributing features for each defect pattern using a minimal amount of expert data and a small amount of non-expert data. Attached Figure Description

[0030] Figure 1 This is a diagram illustrating the outline of an inspection system that uses teacher data collected based on features selected by a feature selection device according to an embodiment of the present invention for learning.

[0031] Figure 2 This is a block diagram illustrating a feature quantity selection device according to one embodiment of the present invention.

[0032] Figure 3 This is a flowchart representing the feature selection process of the feature selection device.

[0033] Figure 4 This is a graph used to illustrate the selection of features based on the standard deviation ratio of each dimension of the feature. Detailed Implementation

[0034] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Figure 1 An inspection system is shown that has learned a classification model using data of images of defective products (defective product data) and data of images of acceptable products (acceptable product data) collected based on features selected by the feature selection device 11 described later. This inspection system 1 is installed, for example, in a vehicle parts manufacturing plant, and automatically determines whether a manufactured vehicle part (e.g., a cylinder block) is a normal product (acceptable product) or an abnormal product (defective product) by inspecting its appearance. Hereinafter, the vehicle part to be inspected will be referred to as the "inspection object".

[0035] like Figure 1As shown, the inspection system 1 includes: a conveyor 2 that transports an object G to be inspected along a given direction at a given speed; and an inspection device 3 that determines whether the object G is good or bad when it reaches a given inspection position. Furthermore, although not shown in the figure, objects G determined to be defective by the inspection device 3 are removed from the conveyor 2 or transported to a dedicated storage area for defective products.

[0036] The inspection device 3 is mainly composed of an information processing device consisting of a computer, and includes a control unit 4, an image acquisition unit 5, a storage unit 6, a learning unit 7, an input unit 8, an output unit 9, and a camera 10.

[0037] The control unit 4 includes a CPU and controls the aforementioned units 5-9 of the inspection device 3, as well as the camera 10. The image acquisition unit 5 acquires an image of the appearance of the object to be inspected G captured by the camera 10 as digital data. The storage unit 6 includes ROM and RAM, storing various programs used in the control of the inspection device 3 and storing various data. The learning unit 7 has a learning model that has learned a benchmark for judging whether the object to be inspected G is good or bad. The input unit 8 includes a keyboard and mouse operated by the operator and is configured to input data and signals from external sources. The output unit 9 includes a display device such as a monitor that displays the judgment result of the object to be inspected G.

[0038] Figure 2 A feature selection device 11 according to one embodiment of the present invention is shown. This feature selection device 11 is used to filter high-contribution features for each type (defect pattern) of defect in nonconforming product data, operated by an operator performing an inspection operation on the object G being inspected. Information on the high-contribution features selected by this feature selection device 11 can be transmitted to a teacher data collection device (not shown) to improve the efficiency of teacher data collection, or transmitted to the learning unit 7 of the inspection device 3 to improve the efficiency of classification model learning.

[0039] The feature selection device 11, like the aforementioned inspection device 3, is an information processing device composed of a computer, and includes a non-conforming product image acquisition unit 12, a non-expert data storage unit 13, an expert data storage unit 14, a feature extraction unit 15 (expert feature extraction unit, non-expert feature extraction unit), a standard deviation calculation unit 16 (expert standard deviation calculation unit, non-expert standard deviation calculation unit), a standard deviation ratio calculation unit 17 (standard deviation ratio calculation unit), and a feature selection unit 18 (feature selection unit).

[0040] The non-conforming product image acquisition unit 12 acquires images of the appearance of the inspection object G captured by the same camera 10 as the aforementioned inspection device 3, which are determined by the operator to be non-conforming products, as non-conforming product data.

[0041] The non-expert data storage unit 13 stores non-conforming product data (non-expert data) selected by non-experts (newcomers, operators with short years of experience in inspection operations). On the other hand, the expert data storage unit 14 stores non-conforming product data (expert data) selected by experts (skilled workers, operators with long years of experience in inspection operations).

[0042] Each non-conforming product data item is pre-labeled with a label indicating the type (defect pattern) of the defect occurring in that non-conforming product. The non-expert data storage unit 13 and the expert data storage unit 14 store each non-conforming product data item in a form that allows for classification according to the label. When the object G being inspected is a casting, the defect pattern can be configured to include at least one of porosity, scratches, dents, and residual chips. In this embodiment, the defect pattern is configured to include any one of porosity, scratches, dents, and residual chips.

[0043] In addition to actual nonconforming product data selected by non-experts and experts respectively, non-expert data and expert data may also include pseudo-nonconforming product data generated based on actual nonconforming product data using, for example, VAE (Variational AutoEncoder) or GAN (Generative Adversarial Network).

[0044] The feature extraction unit 15 extracts non-expert data and expert data labeled with tags indicating the desired defect mode selected by the operator from the non-expert data storage unit 13 and the expert data storage unit 14, respectively, and extracts the given feature quantities from the extracted non-expert data and expert data (expert feature extraction process, non-expert feature extraction process).

[0045] For example, in the case of visual inspection of castings, features of the object being inspected can be extracted based on each pixel of the image or each rectangular region of the scanned image, such as brightness, density period, and positional information. Known methods such as SIFT (Scale-Invariant Feature Transform) and CNN (Convolutional Neural Network) can be used for such feature extraction. In this embodiment, for example, 1058-dimensional features are extracted as given features. For convenience, each feature is assigned a feature number from 1 to 1058.

[0046] The standard deviation calculation unit 16 calculates the standard deviation of each dimension of the feature quantity of the extracted expert data as the standard deviation s. exp (n) (Expert standard deviation calculation procedure). Furthermore, the standard deviation of each dimension of the extracted non-expert data's feature quantities is calculated as the standard deviation s. base (n) (Non-expert standard deviation calculation procedure).

[0047] Standard deviation ratio calculation section 17 calculates the standard deviation s exp (n) With standard deviation s base (n) The ratio of dfratio to standard deviation is used as the standard deviation ratio. (n) (Standard deviation ratio calculation process). Additionally, (n) represents the characteristic quantity number, dfratio (n) This represents the ratio of the standard deviations of the characteristic quantity number n.

[0048] Feature selection unit 18 selects the standard deviation ratio dfratio of each calculated dimension. (n) In this process, a given number of features are selected in descending order of their values. The features whose numbers are selected are those that contribute the most to the initially selected desired defect pattern (feature selection process).

[0049] Figure 3 The above-described feature selection device 11 performs a selection process for features that contribute significantly to a desired defect pattern. In this process, firstly, in step 1 (illustrated as "S1," hereinafter the same), a defective product image is acquired. Here, if the acquired defective product image is selected by a non-expert, the defective product image is stored as non-expert data in the non-expert data storage unit 13. On the other hand, if the acquired defective product image is selected by an expert, the defective product image is stored as expert data in the expert data storage unit 14.

[0050] Regarding whether the obtained image of a defective product was selected by an expert or a non-expert, it can be configured to determine, for example, based on the operator's attribute value input when operating the feature selection device 11. Alternatively, it can be configured to obtain an image of a defective product that has been pre-labeled as having been selected by an expert or a non-expert. In this embodiment, both the obtained expert data and non-expert data can be limited, so the structure is set to obtain an image of a defective product that has been pre-labeled as described above.

[0051] Next, in step 2, the defect mode selected by the operator of the feature selection device 11 is determined, and the defect mode (object defect mode) of the object that will be selected as the feature with the highest contribution in this process is determined, and the process proceeds to step 3.

[0052] In step 3, it is determined whether the number of expert data and non-expert data labeled with tags representing the object defect pattern determined in step 2 has reached a given quantity. This given quantity is set to a sufficient number of data points to select features that contribute significantly to the object defect pattern; the number of expert data points and non-expert data points can be set differently. In this embodiment, the necessary number of expert data points is set to, for example, "30," and the necessary number of non-expert data points is set to, for example, "200." If the determination result in step 3 is "No," the process returns to step 1, and the acquisition of the defective product image is performed again. Conversely, if the determination result in step 3 is "Yes," the process proceeds to step 4.

[0053] In step 4, expert data labeled with tags representing object defect patterns is extracted. From the extracted expert data, features of a given number of dimensions are extracted using known methods such as SIFT or CNN. Next, in step 5, non-expert data labeled with tags representing object defect patterns is extracted, and features of the same given number of dimensions as the expert data are extracted from the extracted non-expert data. In this embodiment, the given number of dimensions is 1058, and features numbered 1 to 1058 are extracted for convenience.

[0054] Next, in step 6, the standard deviation of each dimension of the extracted expert data features is calculated as the standard deviation s. exp (n) Furthermore, the standard deviation of each dimension of the extracted non-expert data features is calculated as the standard deviation s. base (n) .

[0055] Next, in step 7, the standard deviation s is calculated. exp (n) With standard deviation s base (n) The ratio of dfratio to standard deviation is used as the standard deviation ratio. (n) (dfratio (n) =s exp (n) / s base (n) That is, calculate the standard deviation ratio dfratio corresponding to each characteristic quantity from 1 to 1058. (n) .

[0056] As mentioned above, when a feature contributes little to the desired defect pattern, the extracted feature has a narrow distribution in both expert and non-expert data, and the difference in their standard deviations is also small. Therefore, the standard deviation of this feature is higher than that of dfratio. (n) This also results in a smaller value. On the other hand, when the feature quantity contributes significantly to the desired defect pattern, the feature quantity extracted from the expert data has a wide distribution and a standard deviation s. exp (n) On the one hand, the distribution of features extracted from non-expert data becomes larger, and on the other hand, the standard deviation s is narrow. base (n) The difference in standard deviations between the two will not become so large, therefore the standard deviation of this characteristic quantity can be considered to be greater than dfratio. (n) This becomes a larger value. Therefore, it can be considered that the standard deviation is greater than dfratio. (n) The larger the feature quantity, the higher its contribution to the object defect pattern.

[0057] In the next step, 8, the calculated standard deviation ratios (dfratios) for each dimension are determined. (n) The value of dref is determined by whether the number of values ​​exceeding a given threshold dref is greater than a given number m. The given number m is the number of features to be selected as features with high contribution, and can be arbitrarily set. In this embodiment, the given number m is set to, for example, "5". Furthermore, the threshold dref is set to the standard deviation ratio dfratio. (n) Values ​​below this threshold can be presumed to be insufficiently significant in their contribution to the determined defect mode. In this embodiment, the threshold dref is set to, for example, 0.4.

[0058] If the judgment result in step 8 is "No", and the standard deviation ratio of the threshold dref exceeds dfratio (n) If the number of features is less than the given quantity m, it is determined that m features with high contribution cannot be selected, and the process returns to step 1 to continue obtaining images of defective products. Alternatively, the given quantity m can be reset when returning to step 1. For example, the given quantity m can be changed to "4", and each process can be executed again.

[0059] On the other hand, if the judgment result in step 8 is "yes" and exceeds the standard deviation ratio of the threshold dref, then... (n) Given a quantity of m or more, if it is determined that m features with high contribution can be selected, proceed to step 9 and calculate the standard deviation ratio (dfratio) of each dimension. (n) Select m values ​​in descending order of their values, and then select the standard deviation ratio dfratio of the selected values. (n)The associated feature quantity is identified as a feature quantity that contributes significantly to the object's defect pattern, and the processing ends. Furthermore, after the processing concludes, the selected feature quantity number, etc., is notified in the feature quantity selection device 11 via a display unit (not shown).

[0060] Figure 4 This is a graph illustrating an example of the standard deviation ratio calculated for each dimension of the feature. In this example, it is possible to identify the standard deviation ratio dfratio that exceeds the threshold dref (set to 0.4). (n) Since there are more than a given number m (set to 5), the standard deviation ratio dfratio exceeding the threshold dref will be used. (n) Five features are selected from the top five. In the example of this figure, features numbered "43", "88", "161", "308", and "349" are selected as features that contribute highly to the determined defect pattern.

[0061] As detailed above, according to this embodiment, expert data and non-expert data labeled with tags representing specific defect patterns are extracted, and the standard deviation s of each dimension of the feature quantity of the extracted expert data is calculated. exp (n) And calculate the standard deviation s of each dimension of the feature quantities of the extracted non-expert data. base (n) Then, the ratio of these standard deviations, dfratio, is calculated. (n) Choose the standard deviation ratio dfratio (n) Features with larger values ​​are considered to contribute more to the defect pattern. Therefore, it is possible to select high-contribution features for each defect pattern using a minimal amount of expert data and a small amount of non-expert data.

[0062] Furthermore, the present invention is not limited to the embodiments described above, and can be implemented in various ways. For example, in one embodiment, it is configured to calculate the standard deviation ratio dfratio. (n) Then, features with high contributions are selected from those exceeding a given threshold dref. However, as a simpler structure, it can also be configured without setting a threshold dref, selecting features from all standard deviation ratios dfratio. (n) The feature quantity with the largest value is selected. Furthermore, the structure of the minor parts of the feature quantity selection device 11 shown in the embodiment is merely illustrative and can be appropriately modified within the scope of the present invention.

[0063] Symbol Explanation

[0064] 1 Inspection System

[0065] 2 Conveyors

[0066] 3 Inspection device

[0067] 4 Control Department

[0068] 5 Image Acquisition Section

[0069] 6 Storage Unit

[0070] 7. Study Department

[0071] 8 Input Section

[0072] 9 Output Section

[0073] 10 cameras

[0074] 11 Feature quantity selection device

[0075] 12 Non-conforming Product Image Acquisition Department

[0076] 13 Non-Expert Data Preservation Department

[0077] 14 Expert Data Preservation Department

[0078] 15 Feature Extraction Section

[0079] 16 Standard Deviation Calculation Section

[0080] 17 Standard Deviation Ratio Calculation Department

[0081] 18 Feature Selection Section

[0082] G is the object being inspected.

Claims

1. A feature selection method, based on expert data comprising images of various defect shapes and classified according to each defect pattern representing a type of defect, and non-expert data comprising images of a limited number of defect shapes and classified according to each said defect pattern, selects features that contribute highly to a specific defect pattern, characterized in that, have: The expert feature extraction process involves extracting the expert data corresponding to the specific defect pattern and extracting multi-dimensional feature quantities from the extracted expert data. The non-expert feature extraction process involves extracting the non-expert data corresponding to the specific defect pattern and extracting the multi-dimensional feature quantities from the extracted non-expert data. The expert standard deviation calculation process calculates the standard deviation of each dimension of the feature quantities of the extracted expert data. The non-expert standard deviation calculation step calculates the standard deviation of each dimension of the feature quantity of the extracted non-expert data; The standard deviation ratio calculation step involves calculating the ratio of the standard deviation of each dimension of the calculated feature quantity of the expert data to the standard deviation of each dimension of the calculated feature quantity of the non-expert data as the standard deviation ratio; and The feature selection process selects a given number of features from the calculated standard deviation ratios that exceed a given threshold, in descending order of value, and selects the features associated with the selected standard deviation ratios as those that contribute highly to the specific defect pattern.

2. The feature selection method according to claim 1, characterized in that, The defect mode includes at least one of porosity, scratches, and residual chips.

3. The feature selection method according to claim 1, characterized in that, The expert data includes generated images based on actual images of the defect shapes.

4. A feature selection device, which selects features that contribute highly to a specific defect pattern based on expert data comprising images of various defect shapes and classified according to each defect pattern representing a type of defect, and non-expert data comprising images of a limited number of defect shapes and classified according to each said defect pattern, characterized in that, have: An expert feature extraction unit extracts expert data corresponding to a specific defect pattern and extracts multi-dimensional features from the extracted expert data. A non-expert feature extraction unit extracts the non-expert data corresponding to the specific defect pattern and extracts the multi-dimensional feature quantities from the extracted non-expert data. The expert standard deviation calculation unit calculates the standard deviation of each dimension of the feature quantity of the extracted expert data; The non-expert standard deviation calculation unit calculates the standard deviation of each dimension of the feature quantity of the extracted non-expert data. The standard deviation ratio calculation unit calculates the ratio of the standard deviation of each dimension of the calculated feature quantity of the expert data to the standard deviation of each dimension of the calculated feature quantity of the non-expert data as the standard deviation ratio. as well as The feature selection unit selects a given number of features from the calculated standard deviation ratios that exceed a given threshold, in descending order of value, and selects the features associated with the selected standard deviation ratios as those that contribute highly to the specific defect pattern.

Images