Image inspection device, image inspection method, and image inspection program

Abstract

An image inspection apparatus includes: a learning data set acquisition section configured to acquire a learning data set; a learning feature amount deriving unit for deriving a learning feature amount for learning the image data; a clustering processing section configured to cluster the learning image data based on the learning feature amount and divide the learning image data into a plurality of clusters; an index derivation unit for deriving, for each cluster, the centers of gravity and distances of the plurality of learning feature quantities belonging to the cluster, and deriving an average value [mu] and a variance [sigma] of the plurality of distances corresponding to the plurality of learning feature quantities; an imaging feature amount derivation unit for extracting an imaging feature amount based on the imaging data; a minimum distance deriving unit for deriving respective distances between the imaging feature amount and the plurality of centers of gravity, and deriving a minimum distance X, which is the smallest among the respective distances; the abnormal degree calculation part is used for calculating the abnormal degree alpha based on the minimum distance X, the average value mu and the variance sigma; and the comparison part is used for comparing the abnormal degree alpha with a preset threshold value.

Description

TECHNICAL FIELD

[0001] The present application relates to an image inspection apparatus, an image inspection method, and an image inspection program. BACKGROUND

[0002] In the past, in image inspection of products, in order to suppress false determination of defective products as good products, there has been an image inspection method in which a plurality of determination algorithms are combined to make a final determination of whether or not a product is good (for example, Patent Literature 1).

[0003] However, in the conventional image inspection method, the determination score can be unstable, and it can not be possible to clearly distinguish between good products and defective products by a specific threshold value. Therefore, the accuracy of the final determination of whether or not a product is good needs to be further improved. PRIOR ART DOCUMENTS PATENT LITERATURE

[0004] Patent Literature 1: Japanese Patent Laid-Open No. 2018-165712 SUMMARY PROBLEMS TO BE SOLVED BY THE INVENTION

[0005] An object of the present application is to provide an image inspection apparatus, an image inspection method, and an image inspection program capable of determining whether or not a product is good with high accuracy. SOLUTION TO PROBLEM

[0006] (1) An image inspection apparatus according to an aspect of the present application includes: a learning data set acquisition section configured to acquire a learning data set including a plurality of learning image data; a learning feature quantity derivation section configured to derive a learning feature quantity of the learning image data; a clustering processing section configured to cluster the learning image data based on the learning feature quantity and divide the learning image data into a plurality of clusters; an index derivation section configured to derive, for each of the clusters, a center of gravity of a plurality of the learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantity, and derive an average value μ and a variance σ of a plurality of distances corresponding to the plurality of the learning feature quantities; an imaging feature quantity derivation section configured to extract an imaging feature quantity based on imaging data captured by a camera; a minimum distance derivation section configured to derive each distance between the imaging feature quantity and a plurality of the centers of gravity, and derive a minimum distance X which is the smallest of the each distance; an abnormality degree calculation section configured to calculate an abnormality degree a based on the minimum distance X, the average value μ, and the variance σ; and a comparison section configured to compare the abnormality degree a with a predetermined threshold value. (2) In the above (1), an error processing section can be included, which is configured to execute error processing when the abnormality degree a exceeds the threshold value. (3) In the above (1) or (2), the clustering processing section can perform the clustering multiple times using random numbers, and the index deriving section can derive the center of gravity for each cluster each time the clustering is performed, and can use a representative cluster in which the sum of squares of residuals of the distances in the cluster is smallest as a population for calculating the minimum distance X. (4) In the above (1) or (2), the imaging feature quantity deriving section or the feature quantity deriving section can perform size adjustment, window function processing, FFT processing, or band-pass filter processing. (5) An image inspection method according to one aspect of the present application includes: a learning data set acquisition step of acquiring a learning data set including a plurality of learning image data; a learning feature quantity deriving step of deriving a learning feature quantity of the learning image data; a clustering processing step of clustering and dividing the learning image data into a plurality of clusters based on the learning feature quantity; an index deriving step of deriving, for each of the clusters, a center of gravity of a plurality of the learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantity, and deriving an average value μ and a variance σ of a plurality of the distances corresponding to the plurality of the learning feature quantities; an imaging feature quantity deriving step of extracting an imaging feature quantity based on imaging data captured by a camera; a minimum distance deriving step of deriving each distance between the imaging feature quantity and a plurality of the centers of gravity, and deriving a minimum distance X which is the smallest of the distances; an abnormality degree calculating step of calculating an abnormality degree α based on the minimum distance X, the average value μ, and the variance σ; and a comparing step of comparing the abnormality degree α with a predetermined threshold value. (6) An image inspection program according to one aspect of the present application causes a computer to execute: a learning data set acquisition function of acquiring a learning data set including a plurality of learning image data; a learning feature quantity deriving function of deriving a learning feature quantity of the learning image data; a clustering processing function of clustering and dividing the learning image data into a plurality of clusters based on the learning feature quantity; an index deriving function of deriving, for each of the clusters, a center of gravity of a plurality of the learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantity, and deriving an average value μ and a variance σ of a plurality of the distances corresponding to the plurality of the learning feature quantities; an imaging feature quantity deriving function of extracting an imaging feature quantity based on imaging data captured by a camera; a minimum distance deriving function of deriving each distance between the imaging feature quantity and a plurality of the centers of gravity, and deriving a minimum distance X which is the smallest of the distances; an abnormality degree calculating function of calculating an abnormality degree α based on the minimum distance X, the average value μ, and the variance σ; and a comparing function of comparing the abnormality degree α with a predetermined threshold value. Effects of the Invention

[0007] According to the present application, an image inspection device, an image inspection method, and an image inspection program capable of highly accurately determining whether or not to be qualified can be provided. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a diagram showing an outline of an image inspection device. Figure 2 is a diagram showing an outline of an image inspection method. Figure 3 is a diagram showing a flow of an image inspection method. Figure 4 is a diagram showing a flow of a feature quantity deriving step. DETAILED DESCRIPTION

[0009] (Embodiment) Hereinafter, an embodiment of the present application will be described in detail with reference to the accompanying drawings. Figure 1 is a diagram showing an outline of an image inspection device 100. Figure 2 is a diagram showing an outline of an image inspection method. Note that hereinafter, portions having common functions can be given the same symbols or reference numerals.

[0010] (Image Inspection Device) The image inspection device 100 according to the embodiment is used for inspecting the appearance of a product such as an electronic component, for example.

[0011] As shown in Figure 1 , the image inspection device 100 includes a control unit S. The image inspection device 100 can include a camera C.

[0012] The camera C suitably photographs a product (component) to be inspected. The photographed imaging data (image data) is sent to the control unit S. Note that image data photographed by an external camera not included in the image inspection device 100 can be input to the control unit S of the image inspection device 100 so that the image inspection device 100 performs subsequent inspection on the image data.

[0013] The control unit S is a computer system including a CPU and a memory M (storage medium) such as a RAM, a ROM, and an auxiliary storage device, an external connection interface IF with an input / output device, and a bus B. The control unit S can include a network interface. The memory M stores an image inspection program for causing the control unit S (computer system) to execute the image inspection method according to the present embodiment.

[0014] The CPU is the arithmetic circuit that provides comprehensive control of the control unit S. The CPU reads programs stored in ROM or auxiliary storage devices into RAM. The CPU executes various processes within the programs read into RAM. ROM stores system programs used to control the control unit S. Auxiliary storage devices store application programs used to perform various processes. Examples of auxiliary storage devices include HDDs and SSDs. The external connection interface IF is used to connect various devices to the control unit S. Examples of external connection interfaces IF include connecting a camera C, a display D, and a keyboard K to the control unit S. The network interface's function is to communicate via a network based on CPU control. Bus B communicatively connects the aforementioned functional units constituting the control unit S.

[0015] The image inspection device 100 takes pictures of the product to be inspected using a camera. Based on the captured image data, the control unit S determines whether the product to be inspected is qualified or not (whether the product is good or defective, normal or abnormal).

[0016] like Figure 2 As shown, the image inspection device 100 inspects the product based on the captured imaging data. The image inspection device 100 performs a defective product determination and an abnormal product determination in parallel, in combination, on the imaging data to be inspected, to determine whether the product is good or defective. When the product is determined to be good and normal, the image inspection device 100 outputs an OK (qualified) inspection result; when the product is determined to be defective or abnormal, the image inspection device 100 outputs an NG (qualified) inspection result.

[0017] Defect determination is based on good / defect rules learned from pre-collected images of good and defective products. In defect determination, if an unknown defective image is used as input during the learning process, it is possible to mistakenly classify a product that should be determined as defective as a good product.

[0018] To suppress false positives in defective product identification, it is preferable to perform defective product identification and defective product identification in parallel using a combined approach. Defective product identification includes a process of learning the quality of good products as features solely from images of good products, and then calculating a defective product identification score for the image data to be inspected. Defective product identification is based on whether the defective product identification score exceeds a threshold. The anomaly determination score is normalized to a certain range. The specific calculation method for the anomaly determination score could be, for example, using an autoencoder algorithm, or an algorithm combining kernel density estimation and projected histograms.

[0019] Thus, the image inspection device 100 performs the defective product determination and the abnormal product determination in a combined manner in parallel, can avoid misjudgment due to the defective product determination by the abnormal product determination, and thus realizes high-precision inspection. In addition, the image inspection device 100 can also perform only the abnormal product determination.

[0020] The abnormal product determination score of the normal image can not be stable in the algorithm using the autoencoder. In addition, when the image data to be inspected is a color image, in particular, the abnormal product determination score can not have a clear distinction between the normal image and the abnormal image, resulting in inaccurate abnormal product determination. In addition, when using a method combining kernel density estimation and projection histogram, appropriate normalization processing needs to be performed on the abnormal product determination score.

[0021] Therefore, the image inspection device 100 comprises: a learning data set acquisition unit 10 configured to acquire a learning data set comprising a plurality of learning image data; a learning feature quantity derivation unit 20 configured to derive a learning feature quantity of the learning image data; a clustering processing unit 30 configured to cluster and divide the learning image data into a plurality of clusters according to the learning feature quantity; an index derivation unit 40 configured to derive, for each cluster, a center of gravity of a plurality of learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantity, and derive an average value μ and a variance σ of a plurality of distances corresponding to a plurality of learning feature quantities; an imaging feature quantity derivation unit 50 configured to extract an imaging feature quantity according to imaging data photographed by a camera C; a minimum distance derivation unit 60 configured to derive each distance between the imaging feature quantity and a plurality of centers of gravity, and derive a minimum distance X which is the minimum of each distance; an abnormality degree calculation unit 70 configured to calculate an abnormality degree a according to the minimum distance X, the average value μ and the variance σ; and a comparison unit 80 configured to compare the abnormality degree a with a preset threshold. Thus, even if the product to be inspected has multiple variations, or the variation degree of the good product varies according to the product type, an appropriately normalized evaluation score, i.e. the abnormality degree a, can be calculated, so that the pass or fail of the product to be inspected can be determined based on a stable evaluation score with a clear threshold between the good product and the defective product. Therefore, an image inspection device 100 capable of accurately determining pass or fail can be provided.

[0022] It should be noted that each of the above-mentioned components in the image inspection device 100 can be understood as a module or a unit.

[0023] The abnormality degree a is preferably calculated by the following formula: Thus, the evaluation score, i.e. the abnormality degree a, which is appropriately normalized, can be calculated by considering the minimum distance X, the average value μ and the variance σ. Moreover, the variation degree of the appearance of the product to be inspected according to the product type can be considered, and the abnormality degree a seen by the user is always in the same scale.

[0024] The image inspection apparatus 100 can include an error processing section 90 for performing error processing when the abnormality degree a exceeds the threshold value. The error processing is, for example, displaying on a display that the inspection result of the product to be inspected is NG (abnormal or defective). In this way, the user can be notified that the inspection result of the product is NG.

[0025] The clustering processing section 30 can perform clustering multiple times using random numbers. Also, the index deriving section 40 derives the center of gravity for each cluster each time clustering is performed, and takes the representative cluster in which the sum of the squared residuals of the distances in the cluster is the smallest as the population for which the minimum distance X is calculated. In this way, even if the result of each clustering attempt changes, the center of gravity that best captures the characteristics of the learning data set can be derived, and thus an image inspection apparatus 100 that can determine whether or not to pass with higher accuracy can be provided.

[0026] (Image inspection method) Hereinafter, the flow of the image inspection method will be described. The image inspection method can be implemented using the image inspection apparatus 100. Each step of the image inspection method can be performed by the corresponding each section in the image inspection apparatus 100. Figure 3 is a diagram showing the flow of the image inspection method.

[0027] (1) As shown in Figure 3 , first, a learning data set including a plurality of learning image data is acquired (learning data set acquisition step). The learning image data is an image of a good product. That is, the learning data set can be composed only of good product images. The learning data set acquisition step can be performed by the learning data set acquisition section 10 in the image inspection apparatus 100.

[0028] (2) Next, the learning feature amount of the learning image data acquired in the learning data set acquisition step (process) is derived (learning feature amount derivation step). The learning feature amount can be derived based on the learning image data by an appropriate algorithm. The algorithm for deriving the learning feature amount can be, for example, a method using a difference set of brightness in an image as a feature amount (Haar-Like), a method focusing on the distribution of the direction of the brightness gradient in a grayscale image (HOG), a method extracting a feature amount in two steps of feature point detection and feature amount description (SIFT), a filtering method, a wrapping method, and the like. The learning feature amount derivation step can be performed by the learning feature amount derivation section 20 in the image inspection apparatus 100.

[0029] Here, the learning feature amount derivation step can be performed according to the flow shown below. Note that the imaging feature amount derivation step described later can also be performed according to a flow similar to the learning feature amount derivation step shown here. Figure 4 is a diagram showing the flow of the feature amount derivation step.

[0030] (2-1) As shown in Figure 4As illustrated, resizing processing can be performed on the image data (learning image data or imaging image data). Resizing processing is an operation of changing the size of the image data to unify the input size of the machine learning model.

[0031] (2-2) Next, padding processing can be performed on the image data (learning image data or imaging image data) that has been appropriately resized. Padding processing is processing of adjusting the length by adding meaningless data before or after short data so as to treat the data as fixed length.

[0032] (2-3) Next, window function processing can be performed on the image data (learning image data or imaging image data) that has been appropriately padded. Window function processing is processing of multiplying the image data by a window function so as to extract only information in a limited interval.

[0033] (2-4) Next, FFT processing can be performed on the image data (learning image data or imaging image data) that has been appropriately window function processed. FFT processing is processing of performing Fourier transform on the image data to convert the spatial domain to the spatial frequency domain based on the image data and obtain an amplitude spectrum distribution.

[0034] (2-5) Next, band pass filter processing can be performed on the image data (learning image data or imaging image data) that has been appropriately FFT processed. Band pass filter processing is processing of removing components that have little influence on determination by allowing only frequency components in a specific range to pass, compressing the image data.

[0035] (3) Returning Figure 3 Then, based on the learning feature quantity derived in the learning feature quantity derivation step, the learning image data is clustered and divided into a plurality of clusters (clustering processing step). Clustering is a machine learning method of grouping data based on similarity between data. Clustering can be, for example, hierarchical clustering using an algorithm such as the shortest distance method, the longest distance method, the average distance method, the barycenter method, the Ward method, or the like, or non-hierarchical clustering typified by the k-means method. Each time clustering processing is performed, a random number can be used to select an algorithm for clustering. Each time clustering processing is performed, a random number can also be used to change the algorithm for clustering. Specifically, for example, an algorithm can be defined and then a random number used in the algorithm can be reassigned to perform clustering a plurality of times. The clustering processing step can be performed by the clustering processing section 30 in the image inspection device 100.

[0036] (4) Next, for each cluster divided in the clustering processing step, a center of gravity of the plurality of learning feature amounts belonging to the cluster is derived, and distances between the center of gravity and the respective learning feature amounts are derived. Then, an average value μ and a variance σ of the plurality of distances derived for the plurality of learning feature amounts are derived (index derivation step). The index derivation step can be performed by the index derivation section 40 in the image inspection apparatus 100. The respective steps above can be used as a pre-process for inspection to derive a standard index of a good product.

[0037] (5) Next, imaging feature amounts are extracted based on imaging data captured by a camera (imaging feature amount derivation step). The imaging data is image data obtained by capturing a product to be inspected by the camera. The imaging feature amounts can be derived based on the imaging data by a suitable algorithm. The extraction of the imaging feature amounts can be the same as the method used when the learning feature amounts are derived in the learning feature amount derivation step. The imaging feature amount derivation step can be performed by the imaging feature amount derivation section 50 in the image inspection apparatus 100.

[0038] (6) Next, each distance between the imaging feature amounts and the plurality of centers of gravity is derived, and a minimum distance X which is the smallest of the respective distances is derived (minimum distance derivation step). The minimum distance derivation step can be performed by the minimum distance derivation section 60 in the image inspection apparatus 100. (7) Next, based on the minimum distance derived in the minimum distance derivation step, and the average value μ and the variance σ derived in the index derivation step, an abnormality degree α is calculated using a calculation formula (α = μ - σ) in particular (abnormality degree calculation step). The abnormality degree calculation step can be performed by the abnormality degree calculation section 70 in the image inspection apparatus 100.

[0039] (8) Next, the abnormality degree α calculated in the abnormality degree calculation step is compared with a preset threshold value (comparison step). The comparison step can be performed by the comparison section 80 in the image inspection apparatus 100.

[0040] As described above, the image inspection method includes the respective steps described above. Thereby, it is possible to stabilize the determination score, and clearly divide good products and defective products by a specific threshold value. Thus, it is possible to improve the accuracy of the final determination of whether or not a product is acceptable.

[0041] (9) In addition, error processing can be performed when the abnormality degree α exceeds the threshold value, as appropriate after the comparison step (error processing step). The error processing is, for example, displaying the inspection result of the product to be inspected as NG (abnormal or defective) on the display D. In this way, it is possible to notify the user that the inspection result of the product is NG. The error processing step can be performed by the error processing section 90 in the image inspection apparatus 100.

[0042] (Image inspection program) ​The image inspection program has a function of causing the computer to perform each process performed by each step of the above-described image inspection method. Specifically, the image inspection program has a learning data set acquisition function, a learning feature quantity derivation function, a clustering processing function, an index derivation function, an imaging feature quantity derivation function, a minimum distance derivation function, an abnormality degree calculation function, and a comparison function, corresponding to each step of the above-described image inspection method. In this way, it is possible to stabilize the determination score and clearly divide the good products and the defective products by a specific threshold value. Thus, it is possible to improve the accuracy of the final determination of whether the product is good or not.

[0043] Note that the technical scope of the present application is not limited to the above-described embodiments, and various changes can be made without departing from the gist of the present application.

[0044] In addition, in a range not departing from the gist of the present application, the constituent elements in each of the above-described embodiments can be appropriately replaced with well-known constituent elements. Further, in a range not departing from the gist of the present application, the above-described variants can be appropriately combined.

[0045] As described above, according to the image inspection apparatus 100 of the embodiment, the learning data set acquisition section 10 acquires a learning data set including a plurality of learning image data, the learning feature quantity derivation section 20 derives learning feature quantities of the learning image data, the clustering processing section 30 clusters and divides the learning image data into a plurality of clusters based on the learning feature quantities, the index derivation section 40 derives, for each cluster, a center of gravity of a plurality of learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantities, and derives an average value μ and a variance σ of a plurality of distances corresponding to the plurality of learning feature quantities, the imaging feature quantity derivation section 50 extracts an imaging feature quantity from imaging data captured by the camera C, the minimum distance derivation section 60 derives each distance between the imaging feature quantity and the plurality of centers of gravity, and derives a minimum distance X that is the smallest among the distances, the abnormality degree calculation section 70 calculates an abnormality degree a based on the minimum distance X, the average value μ, and the variance σ, and the comparison section 80 compares the abnormality degree a with a predetermined threshold value. In this way, it is possible to calculate an evaluation score, that is, the abnormality degree a, which is appropriately normalized, and thus it is possible to determine whether the product under inspection is good or not based on a stable evaluation score having a clear threshold value between the good product and the defective product. Therefore, it is possible to provide an image inspection apparatus 100 capable of determining whether the product is good or not with high accuracy.

[0046] According to the image inspection method of the embodiment, the image inspection method includes: a learning data set acquisition step of acquiring a learning data set including a plurality of learning image data; a learning feature quantity derivation step of deriving a learning feature quantity of the learning image data; a clustering processing step of clustering the learning image data and dividing into a plurality of clusters based on the learning feature quantity; an index derivation step of deriving, for each cluster, a center of gravity of a plurality of learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantity, and deriving an average value μ and a variance σ of a plurality of distances corresponding to the plurality of learning feature quantities; an imaging feature quantity derivation step of extracting an imaging feature quantity based on imaging data captured by a camera; a minimum distance derivation step of deriving each distance between the imaging feature quantity and the plurality of centers of gravity, and deriving a minimum distance X that is the smallest among the distances; an abnormality degree calculation step of calculating an abnormality degree α based on the minimum distance X, the average value μ, and the variance σ; and a comparison step of comparing the abnormality degree α with a predetermined threshold value. In this way, an evaluation score that is appropriately normalized, that is, the abnormality degree α, can be calculated, and thus whether or not a product to be inspected is good or not can be determined based on a stable evaluation score having a clear threshold value between good products and defective products. Therefore, an image inspection method capable of determining whether or not a product is good or not with high precision can be provided.

[0047] According to the image inspection program of the embodiment, the image inspection program includes: a learning data set acquisition function of acquiring a learning data set including a plurality of learning image data; a learning feature quantity derivation function of deriving a learning feature quantity of the learning image data; a clustering processing function of clustering the learning image data and dividing into a plurality of clusters based on the learning feature quantity; an index derivation function of deriving, for each cluster, a center of gravity of a plurality of learning feature quantities belonging to the cluster and a distance between the center of gravity and the learning feature quantity, and deriving an average value μ and a variance σ of a plurality of distances corresponding to the plurality of learning feature quantities; an imaging feature quantity derivation function of extracting an imaging feature quantity based on imaging data; a minimum distance derivation function of deriving each distance between the imaging feature quantity and the plurality of centers of gravity, and deriving a minimum distance X that is the smallest among the distances; an abnormality degree calculation function of calculating an abnormality degree α based on the minimum distance X, the average value μ, and the variance σ; and a comparison function of comparing the abnormality degree α with a predetermined threshold value. In this way, an evaluation score that is appropriately normalized, that is, the abnormality degree α, can be calculated, and thus whether or not a product to be inspected is good or not can be determined based on a stable evaluation score having a clear threshold value between good products and defective products. Therefore, an image inspection program capable of determining whether or not a product is good or not with high precision can be provided.

[0048] REFERENCE NUMERALS 100: image inspection apparatus, 10: learning dataset acquisition section, 20: learning feature quantity derivation section, 30: clustering processing section, 40: index derivation section, 50: imaging feature quantity derivation section, 60: minimum distance derivation section, 70: abnormality degree calculation section, 80: comparison section, 90: error processing section, a: abnormality degree, m: mean value, s: variance, B: bus, C: camera, D: display, K: keyboard, M: memory, S: control unit, X: minimum distance.

Claims

1. An image inspection apparatus, comprising: The learning dataset acquisition unit is used to acquire a learning dataset that includes multiple learning image data. The learning feature derivation unit is used to derive the learning feature values ​​of the learning image data. The clustering processing unit is used to cluster the learned image data and divide it into multiple clusters based on the learned feature values; The indicator derivation unit is used to derive, for each cluster, the centroids of multiple learned feature quantities belonging to the cluster and the distances between the centroids and the learned feature quantities, and to derive the average value μ and variance σ of the multiple distances corresponding to the multiple learned feature quantities; The imaging feature deriving unit is used to extract imaging features based on the imaging data captured by the camera. The minimum distance derivation unit is used to derive each distance between the imaging feature quantity and the plurality of centroids, and to derive the minimum distance X among the various distances; An anomaly degree calculation unit is used to calculate the anomaly degree α based on the minimum distance X, the average value μ, and the variance σ; The comparison unit is used to compare the degree of abnormality α with a preset threshold.

2. The image inspection apparatus according to claim 1, It includes an error handling unit for performing error handling when the anomaly level α exceeds a threshold.

3. The image inspection apparatus according to claim 1 or 2, The clustering processing unit performs the clustering multiple times using random numbers. When performing the clustering, the index derivation unit derives the centroid for each cluster and takes the cluster with the smallest sum of squared residuals of distances as the population for calculating the minimum distance X.

4. The image inspection apparatus according to claim 1 or 2, The imaging feature deriving unit or the learning feature deriving unit is subjected to size adjustment, window function processing, FFT processing, or bandpass filtering processing.

5. An image inspection method, comprising: The steps for obtaining the learning dataset are as follows: Obtain a learning dataset that includes multiple learning image data. The learning feature quantification step is to derive the learning feature quantification of the learning image data. The clustering process involves clustering the learned image data into multiple clusters based on the learned feature values. The indicator derivation step involves, for each cluster, deriving the centroids of multiple learned feature quantities belonging to the cluster and the distances between the centroids and the learned feature quantities, and deriving the average value μ and variance σ of the multiple distances corresponding to the multiple learned feature quantities. The imaging feature deriving step extracts imaging features based on the imaging data captured by the camera. The minimum distance derivation step involves deriving the distances between the imaging feature quantity and the multiple centroids, and then deriving the minimum distance X among these distances. The anomaly degree calculation step involves calculating the anomaly degree α based on the minimum distance X, the average value μ, and the variance σ. The comparison step compares the degree of abnormality α with a preset threshold.

6. An image inspection program for causing a computer to perform: The function to acquire a learning dataset includes multiple learning image data. The function of exporting learning features is used to export the learning features of the learning image data. The clustering function clusters the learned image data into multiple clusters based on the learned feature values. The indicator export function, for each cluster, exports the centroids of multiple learned feature quantities belonging to the cluster and the distances between the centroids and the learned feature quantities, and exports the average value μ and variance σ of the multiple distances corresponding to the multiple learned feature quantities; The imaging feature export function extracts imaging features based on the imaging data captured by the camera. The minimum distance export function exports the distances between the imaging feature and the multiple centroids, and exports the minimum distance X among the various distances. The anomaly degree calculation function calculates the anomaly degree α based on the minimum distance X, the average value μ, and the variance σ; The comparison function compares the degree of abnormality α with a preset threshold.

Images