JP2023047998A - Inspection apparatus, learning model generation apparatus, and program - Google Patents

Abstract

To provide an inspection apparatus which can identify each of test objects to inspect them, on the basis of an image that represents two or more test objects.SOLUTION: An inspection apparatus includes: feature point information output means which outputs information for identifying positions of multiple feature points in an entire image, the information obtained by inputting the entire image to a trained model equipped in advance, the entire image being an image that represents two or more test objects; feature point set extraction means which extracts, from among the feature points output by the feature point information output means, a feature-point set which is a set of two or more feature points each representing two or more sections of the test objects, and satisfying requirements of positional relationship based on a tendency of skeleton shape of the test objects; and inspection means which performs, based on the feature-point set, inspection of the test objects represented in the entire image, specified by positions of the feature points constituting the feature-point set.SELECTED DRAWING: Figure 1

Description

The present invention relates to an inspection device, a learning model generation device, and a program for inspecting each object based on an image representing two or more objects to be inspected.

An inspection device that cuts out an image for each inspected object from an overall image obtained by photographing a plurality of inspected objects being conveyed by a conveying means, inputs the image into a trained model, and determines whether or not the inspected object is a normal product. are disclosed in Patent Documents 1 and 2.

Japanese Patent No. 6754155 WO2020/189043

In order to extract an image for each inspection object from an image representing two or more inspection objects, it is necessary to be able to identify individual inspection objects in the image. Identification of individual inspected objects in an image is difficult if they are granular objects such as almonds and beans, which are less likely to overlap with other inspected objects during transportation. It is easy because it is photographed in the state. However, in the case of inspected objects that are likely to overlap or come into contact with each other during transportation, images may be taken in a state in which a plurality of inspected objects overlap or come into contact with each other. In this case, the images of the objects to be inspected are connected at overlapping portions or contact portions, and it is not easy to identify individual objects to be inspected in the image.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems. The purpose is to provide a device and a program.

The inspection apparatus of the present invention uses an image representing two or more objects to be inspected as an overall image, and inputs the overall image to a pre-prepared trained model to determine the positions of a plurality of feature points in the overall image. Feature point information output means for outputting identifiable information, and feature point information representing two or more predetermined number of parts of the object to be inspected, satisfying the requirements of the positional relationship based on the tendency of the skeletal shape of the object to be inspected. Feature point set extracting means for extracting a set of feature points from among a plurality of feature points; and inspection means for performing inspection based on the set of feature points.

The positional relationship requirement may be a requirement for an arc formed by sequentially connecting a predetermined number of feature points extracted as a set of temporary feature points from among a plurality of feature points.

The requirement on the arc may be a requirement on the radius of curvature of the arc.

Also, the requirement on the arc may be a requirement on the length of the arc.

The feature point information output means further outputs information of a region represented by each of the plurality of feature points in the whole image obtained by inputting the whole image to the trained model, and the feature point set extraction means After extracting a provisional feature point set consisting of a predetermined number of feature points from the feature points of A feature point set may be extracted by the determination.

A set of feature points representing an object to be inspected is a set of three feature points representing one end, a central portion, and the other end of the object to be inspected. A feature point whose representative part is the central part and two feature points whose representative part is the edge are extracted as a temporary feature point set from the points, and then the positional relationship requirements are satisfied. A set of feature points may be extracted by determining whether or not the

The requirement for the positional relationship may be a requirement that the distance between the feature point whose representative part is the central part and the feature point whose representative part is the end part is within a predetermined range.

The inspection means further comprises a clipping means for clipping a predetermined range with the position of the feature point belonging to the feature point set as a base point from the entire image as an image of the part of the inspection object represented by the feature points, wherein the inspection means is clipped by the clipping means. The object to be inspected may be inspected based on the obtained image.

In the feature point information output means, the information capable of specifying the positions of the feature points obtained by inputting the entire image to the trained model is expressed as the median and mode values of the feature point positions, and the probability density is It may be a probability distribution following a normal distribution.

transportation means for transporting an object to be inspected; irradiation means for irradiating a predetermined electromagnetic wave to the object to be inspected transported by the transportation means; It may further include detection means for detecting electromagnetic waves and outputting detection data, and image generation means for generating an entire image based on the detection data and providing it to the feature point information output means.

The learning model generating apparatus of the present invention includes a learning image that is an image representing two or more objects to be inspected, and positional information of each of a plurality of feature points in the learning image that are correct labels. By executing machine learning using the training data, a plurality of features in the whole image is input by inputting a whole image that is an image representing two or more inspected objects in which a plurality of feature points are unknown. Generate a learning model that outputs information that can identify the position of a point.

The correct label may further include information about the part represented by each of the plurality of feature points in the learning image, and the learning model may further output information about the part represented by each of the plurality of feature points in the whole image. .

The learning model outputs a probability distribution whose probability density follows a normal distribution, in which the positions of the feature points are expressed as the median and the mode, as information that can identify the positions of the feature points. The position of the feature point set in the learning image is the median and mode, and the probability distribution follows the normal distribution. In machine learning, the probability distribution output by the learning model and the probability of being the correct label The parameters of the learning model may be adjusted so as to increase the degree of matching with the distribution.

The inspection apparatus of the present invention may be realized by executing a program describing the functions of the inspection apparatus of the present invention on a computer.

The learning model generation device of the present invention may be realized by executing a program describing the functions of the learning model generation device of the present invention on a computer.

According to the inspection device, the learning model generation device, and the program of the present invention, each object can be identified even in an image representing two or more objects to be inspected that are photographed in a state of overlap or contact. Inspection can be performed.

1 is a configuration diagram of an inspection device 100 and a learning model generation device 200; FIG. It is a figure which shows the state which the to-be-inspected object W1 and the to-be-inspected object W2 partially overlapped. FIG. 4 is a diagram showing an electromagnetic wave transmission image P1 of inspected objects W1 and W2; FIG. 10 is a diagram showing an example in which positions of feature points themselves are output from a trained model; FIG. 10 is a diagram showing an example in which a probability distribution corresponding to feature points is output from a learned model; FIG. 10 is a diagram explaining that the positions of feature points can be specified from the probability distribution output from the trained model; It is a figure which shows the state which the to-be-inspected object W11 and the to-be-inspected object W12 partially overlapped. FIG. 4 is a diagram showing an electromagnetic wave transmission image P11 of inspected objects W11 and W12; It is a figure which shows the example which set the feature point itself to to-be-tested object W11 and W12. FIG. 10 is a diagram showing an example of converting feature points set on the objects to be inspected W11 and W12 into a probability distribution; FIG. 10 is a diagram showing another example of outputting a probability distribution corresponding to feature points from a learning model; FIG. 10 is a diagram illustrating deviations between feature points based on a probability distribution output from a learning model and feature points that are correct labels; FIG. 10 is a diagram showing an example of an arc formed by sequentially connecting feature points belonging to a feature point set; FIG. 11 is a diagram showing another example of an arc formed by sequentially connecting feature points belonging to a feature point set; It is a figure which shows the state which superimposed the feature point group on the electromagnetic wave transmission image P1. FIG. 10 is a diagram showing an example of cutting out a partial image from the entire image; FIG. 10 is a diagram showing another example of cutting out a partial image from the entire image;

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below with reference to the drawings. In the following description, the same members are denoted by the same reference numerals, and the description of members that have already been described will be omitted as appropriate.

FIG. 1 shows a configuration diagram of an inspection device 100 and a learning model generation device 200 of the present invention.

The inspection apparatus 100 includes transport means 110 , irradiation means 120 , detection means 130 , image generation means 140 , feature point information output means 150 , feature point set extraction means 160 and inspection means 170 .

The conveying means 110 is a conveyor of any type that conveys the placed inspection object W in the Y-axis direction at a predetermined speed. The conveying means 110 has a depth in the X-axis direction perpendicular to the Y-axis direction, which is the conveying direction, when viewed from the front.

The irradiating means 120 irradiates a predetermined electromagnetic wave to the inspection object W conveyed by the conveying means 110 . The type of electromagnetic waves to be irradiated may be appropriately selected according to the contents of the inspection, such as X-rays, visible light, and infrared rays.

The detection means 130 is arranged at a position where it can detect the electromagnetic wave emitted from the irradiation means 120 and passed through the object W to be inspected. Output detection data to . FIG. 1 shows an example in which the electromagnetic wave transmitted through the object W to be inspected is arranged to face the irradiating means 120 so that the electromagnetic wave can be detected. The detection means 130 may be provided inside the conveying means 110 as shown in FIG. A slight gap may be provided between the two conveying means 110 provided so that the electromagnetic wave transmitted through the object W to be inspected may be directly detected. When detecting electromagnetic waves passing through the conveyor belt or the like of the conveying means 110, it is preferable to adopt a conveyor belt or the like made of a material having high permeability for the electromagnetic waves emitted from the irradiation means 120. FIG.

The detection means 130 includes a plurality of detection elements arranged in a line in a direction orthogonal to the conveying direction, and each detection element detects the electromagnetic wave that has reached it and outputs detection data. In the example of FIG. 1, the multiple detection elements are arranged in the X-axis direction orthogonal to the Y-axis direction.

The positions of the irradiation means 120 and the detection means 130 are fixed with respect to the transport means 110, and the entire inspection object W being transported in the Y-axis direction by the transport means 110 is arranged in a line in the X-axis direction. By traversing the detection means 130, detection data of the transmitted electromagnetic wave when the inspection object W is viewed on the XY plane can be obtained.

The image generating means 140 generates an electromagnetic wave transmission image that expresses the shape and internal state of the object W to be inspected by shading or the like, based on the detection data of the electromagnetic wave transmitted through the object W obtained by the detecting means 130 . At this time, when two or more objects to be inspected are transported while overlapping each other, an integrated electromagnetic wave transmission image is generated in which the overlapping state is expressed in an XY plan view. For example, when two objects W1 and W2 to be inspected are transported while partially overlapping each other as shown in FIG. 2, an electromagnetic wave transmission image P1 having a contour as shown in FIG. 3 is generated.

An electromagnetic wave transmission image P1 expressing a state in which two or more objects to be inspected are superimposed in this way is provided to the feature point information output means 150 as a whole image.

The feature point information output means 150 is obtained by inputting the whole image provided from the image generation means 140 into a pre-prepared trained model, and is information capable of specifying the position of each of the plurality of feature points in the whole image. to output

Examples of information that can specify the position of a feature point include information on the position of the feature point itself, and a probability distribution whose probability density follows a normal distribution, in which the position of the feature point is expressed as a median value and a mode value. be done.

Specifically, for example, when outputting the positions of three feature points for each object to be inspected, by inputting the electromagnetic wave transmission image P1 shown in FIG. Alternatively, as shown in FIG. 5, each position of the feature points C1 to C6 is expressed as the median and mode of the probability density. Probability distributions H1 to H6, which are heat maps, may be output. In this case, as shown in FIG. 6, the respective centers of probability distributions H1 to H6 can be identified as feature points C1 to C6. A detailed description of the output of the probability distribution from the trained model will be given later.

A learned model is generated in advance by the learning model generation device 200 . The learning model generation device 200 generates teacher data including a learning image that is an image representing two or more inspected objects, and positional information of each of a plurality of feature points in the learning image, which is a correct label. By repeatedly executing machine learning using a large number of , a trained model is generated that outputs information that can specify the positions of each of the plurality of feature points in the entire image based on the input of the entire image. Machine learning may be general machine learning in which a person defines feature amounts, or deep learning in which a machine defines feature amounts.

The device user or the like sets a predetermined number of feature points as correct labels for portions corresponding to the respective objects to be inspected in the learning image. A feature point is one point that constitutes a set of feature points that characterize the skeletal shape of an object to be inspected expressed in the entire image, and also represents a part of the object to be inspected that corresponds to a set position. This is the base point used for clipping. Therefore, the setting position and the number of setting points of the feature points should be determined according to the tendency of the skeleton shape of the object to be inspected, the part to be inspected, and the like. For example, if the skeletal shape is curvilinear, it may be set at three locations, one end, the center, and the other end.

When the electromagnetic wave transmission image P11 shown in FIG. 8, in which the two objects W11 and W12 partially overlap each other as shown in FIG. 7, is a learning image, the portion corresponding to the object W1 Characteristic points C11, C12 and C13 are set at one end, center and other end of the inspection object W2, respectively, and characteristic points C14, C15 and C16 are respectively set at one end, center and other end of the portion corresponding to the object to be inspected W2. is shown in FIG.

The generation and learning of a learning model is generally a process of determining the degree of matching between the output obtained by inputting learning data to the learning model and the correct label of the learning data, and adjusting the parameters to increase the degree of matching to the learning model. is performed by repeating A learning model generated by the learning model generation device 200 that outputs information that can specify the positions of each of a plurality of feature points in the entire image based on the input of the entire image directly outputs the positions of each of the plurality of feature points. It may be a learning model that However, when training a learning model that directly outputs the positions of multiple feature points, the positions of the feature points output from the learning model and the positions of the feature points set in the learning image, which are the correct labels, are both points, and the probability that they match is almost zero. Therefore, the degree of matching cannot be determined, and it is difficult to learn with the general learning method described above.

Therefore, for example, the position in the learning image is a random variable, and the position of the feature point set in the learning image as the correct label is the median value and the mode value, and the probability density follows a normal distribution. and use the transformed correct labels for training. Correct label conversion may be performed using, for example, an appropriately created application. By performing learning using the correct label converted in this way, as an output for the input of the training image, a probability distribution whose probability density follows a normal distribution, in which the position of the feature point is expressed as the median value and the mode value. A learning model is generated that outputs Since both the output from the learning model and the converted correct label are positional probability distributions, it is possible to determine the degree of matching. adjustments can be applied to the learning model. In addition, in the probability distribution output by inputting the whole image to the trained model generated by repeatedly performing learning on the learning model, the position of the feature point is expressed as the median value and the mode value. Therefore, the learned model corresponds to a learned model that outputs information that enables the positions of feature points to be specified based on the input of the entire image.

Both the probability distribution that is the output from the learning model and the probability distribution that is the correct label have the maximum probability density at the center, which is the position of the feature point, and decrease according to a normal distribution as they move away from the center. is the highest density, and the image can be expressed in the form of a heat map in which the density decreases concentrically with increasing distance from the center. Therefore, by creating a heat map image of the probability distribution output from the learning model and the probability distribution of the correct label, the degree of agreement between the median values can be calculated on the image from the degree of separation between the highest density positions of both probability distributions. can judge.

For example, when learning a learning model using the electromagnetic wave transmission image P11 shown in FIG. are converted into probability distributions H11 to H16 represented by a heat map as shown in FIG. Outputs of the expressed probability distributions H21 to H26 are obtained, and feature points C21 to C26 corresponding to the probability distributions H21 to H26 are specified. As a result, positional deviations D1 to D6 between the feature points C21 to C26 based on the probability distribution output from the learning model and the feature points C11 to C16, which are the correct labels, are obtained as shown in FIG. Parameter adjustments can be made to the learning model to increase the match based on.

The feature point set extraction means 160 is a feature point set that is a set of feature points that respectively represent two or more predetermined numbers of parts of the object to be inspected, satisfying the requirements of the positional relationship based on the tendency of the skeletal shape of the object to be inspected. are extracted from the plurality of feature points output by the feature point information output means 150 .

Each feature point belonging to a plurality of feature points output by the trained model is located at a position representative of any part of any inspection object represented in the whole image. It is not clear whether it is a representative feature point. Therefore, the feature point group extracting means 160 groups the feature points of each object based on the tendency of the skeleton shape of the object.

As a specific method for extracting feature point sets, for example, a predetermined number of feature points corresponding to a predetermined number of parts are extracted as a provisional feature point set from among a plurality of feature points, and then extracted. There is a method of defining a requirement regarding an arc formed by sequentially connecting each feature point as a positional relationship requirement, and extracting a feature point set that satisfies this requirement.

Taking the case where the objects W1 and W2 to be inspected shown in FIG. meet the standards. Therefore, for example, it is possible to define requirements related to arcs, such as whether the radius of curvature is a certain value or more and whether the length of the arc is within a certain range, corresponding to these standards.

The electromagnetic wave transmission image P1 for the objects to be inspected W1 and W2 shown in FIG. 3 is input to the learned model, and in the positional relationship shown in FIG. When C1 to C6 are output, two sets of temporary feature points each consisting of three feature points are extracted, and each of the three feature points is sequentially connected to form an arc. For example, as shown in FIG. 13, when the characteristic points C1, C2, and C6 are sequentially connected to form an arc A1, and the characteristic points C3, C4, and C5 are sequentially connected to form an arc A2, the arc A1 is a circular arc. It can be determined that the length is too long and the radius of curvature for arc A2 is too small. Such determination is repeated while changing the combination of the feature points that constitute the set of provisional feature points, and finally, as shown in FIG. By specifying the arc A3 and the arc A4 connecting the feature points C4, C5, and C6 in order, the feature point set G1 consisting of the feature points forming the arc A3 and the feature points forming the arc A4 are identified. A feature point set G2 can be extracted.

In addition, instead of forming an arc, a curve that sequentially connects each feature point is expressed by a piecewise polynomial, and instead of the radius of the arc, the requirements are defined in the form of providing a certain relational expression or range for the coefficients of the piecewise polynomial. may

Extraction of a provisional feature point set prior to extraction of a feature point set by the feature point set extraction means 160 may be performed in consideration of information of a site represented by the feature point. Specifically, when generating and learning a learning model, information about the region represented by each feature point is added to the correct label, and each feature point represents the output of the learned model. A learning model is configured to obtain information on body parts. For example, if a total of three feature points, one at both ends and one at the center, are set for each object to be inspected, each feature point represents the center or the edge as the output of the trained model. Since it is possible to obtain information about the object, the feature point set extraction means 160 extracts a provisional feature point set consisting of a feature point representing the central portion and two feature points representing the end portions. Extract. By adopting an extraction method that takes into account the information of the parts in this way, it is possible to avoid the determination of a provisional feature point set based on a combination of feature points with an unnatural positional relationship, so the extraction processing time can be shortened. Extraction accuracy can be improved. When setting three feature points for each object to be inspected, that is, both ends and the center, the distance between the feature point representing the central portion and the feature point representing the end is a predetermined distance. The provisional feature point set may be determined by further considering the requirement of the positional relationship that it is within the range of . This makes it possible to further improve the extraction accuracy.

Moreover, in addition to the information of the part represented by the feature point, a method of adding information of the direction to the adjacent feature point may be adopted. Specifically, when generating and learning a learning model, for each feature point, information on the part represented by the feature point and information on the direction to the adjacent feature point are added to the correct label, and the learned model's The learning model is configured so that information on the part represented by each feature point and information on the direction to the adjacent feature point can be obtained as outputs. Then, the feature point set extracting means 160 selects the feature points that form the provisional feature point set, taking into account the part information and the direction information to the adjacent feature points. For example, if a total of three feature points are set for each object to be inspected, i.e., at both ends and at the center, each feature point will either represent the center or end as the output of the trained model. If it is the central part, information on the direction to each end can be obtained, and if it is the end, information on the direction to the central part can be obtained. A set is composed of a feature point representing the central portion and two feature points representing the end portions whose orientation information is roughly the same in opposite directions. As a result, the number of temporary feature point sets to be determined can be reduced compared to the case where only the information of the part is taken into consideration, so that the extraction processing time can be further shortened and the extraction accuracy can be further improved.

Note that the extraction of feature point sets from a plurality of feature points by the feature point set extraction means 160 may be performed using a trained model that outputs a feature point set upon input of a plurality of feature points, in addition to the above method. may be taken. In this case, the trained model used in the feature point group extraction means 160 outputs information that can specify the positions of each of the plurality of feature points in the whole image to the learning model generation device 200, for example, based on the input of the whole image. In addition to the function of generating a learning model, a function of generating a learning model that outputs a set of feature points by inputting a plurality of feature points may be provided, and the trained model generated here may be provided, or not shown. A trained model generated by another learning model generating device may be provided. A learning model that outputs a set of feature points by inputting a plurality of feature points contains a large amount of teacher data including information on the positional relationship of the plurality of feature points and information on the feature points that make up the set of feature points that are correct labels. It can be generated by repeatedly performing machine learning using

The inspection means 170 inspects the inspection object represented by the entire image specified from the position of each feature point forming the feature point set, based on the feature point set. For example, it is possible to determine the normality of the degree of curvature and the length of the object to be inspected from the shape and length of an arc connecting each feature point forming the feature point set.

Further, between the feature point group extraction means 160 and the inspection means 170, a predetermined range with the position of the feature point belonging to the feature point group extracted by the feature point group extraction means 160 as a base point is defined as a representative feature point. A clipping means 180 may be provided for clipping the entire image as an image of the part of the object to be inspected, and the inspection means 170 may inspect the object based on the image clipped by the clipping means 180 .

The predetermined range, specifically the shape, size, direction, etc., for cutting out the image of the part of the object to be inspected from the whole image in the cutting out means 180 is appropriately determined according to the shape, size, direction, etc. of the object to be inspected. you can Also, the shape, size, and direction to be cut out for each part may be different.

For example, when two sets of feature points G1 and G2 are extracted as shown in FIG. Then, the feature point sets G1 and G2 are superimposed on the corresponding positions of the electromagnetic wave transmission image P1, and predetermined ranges are cut out from the electromagnetic wave transmission image P1 using each of the feature points C1 to C6 as base points. At this time, the respective images extracted from the feature points C1 to C3 belonging to the feature point group G1 are partial images of one end, center and other end of the inspection object W1, and belong to the feature point group G2. The images cut out from the characteristic points C4 to C6 are partial images of one end, center, and other end of the inspection object W2. FIG. 16 shows an example of cutting out partial images T1 to T3 of one end, center, and other end of the object W1 to be inspected using the feature points C1 to C3 forming the feature point group G1 as base points. An example of cutting out partial images T4 to T6 of one end, center, and other end of the inspection object W2 using the feature points C4 to C6 forming the set of points G2 as base points is shown.

The inspection performed by the inspection means 170 based on the image cut out by the cutout means 180 may be performed by image processing, or may be performed by inputting to a trained model for inspection prepared in advance. Since the image cut out by the cutout means 180 is a partial image of each part of the object to be inspected, the inspection result is the inspection result of each part of the object to be inspected. Thus, it is possible to obtain the inspection result of the entire object to be inspected.

Note that if the inspection of a part of the region is omitted, for example, the both ends are inspected but the central portion is not inspected, there is no need to cut out the image of the part of the region by the clipping means 180. do not have.

The inspection device and learning model generation device of the present invention may be realized by causing a computer to execute a program in which the functions of each device are described. That is, in a computer, the functions of each device may be realized by having the CPU read out and execute a program stored in an arbitrary storage means.

According to the inspection apparatus and the learning model generation apparatus of the present invention described above, each object can be identified even in an image representing two or more objects photographed in a state of overlap or contact. inspection can be carried out

In addition, the present invention is not limited to the above embodiments. Each embodiment is an example, and the present invention can be any one that has substantially the same configuration as the technical idea described in the claims of the present invention and produces similar effects. included in the technical scope of That is, modifications can be made as appropriate within the scope of the technical ideas expressed in the present invention, and forms with such modifications and improvements are also included in the technical scope of the present invention.

100 inspection device 110 transportation means 120 irradiation means 130 detection means 140 image generation means 150 feature point information output means 160 feature point set extraction means 170 inspection means 180 clipping means A1 to A4 arcs C1 to C6, C11 to C16, C21 to C26 Features Points D1 to D6 Positional shifts G1, G2 Characteristic point sets H1 to H6, H11 to H16 H21 to H26 Probability distributions P1, P11 Whole images T1 to T6 Partial images W1, W2, W11, W12 Object to be inspected

Claims (15)

Information capable of specifying the positions of a plurality of feature points in the overall image obtained by inputting the overall image to a pre-prepared trained model with an image representing two or more inspected objects as the overall image. feature point information output means for output;
A feature point set, which is a set of feature points respectively representing a predetermined number of two or more parts of the object to be inspected, satisfying the requirements of the positional relationship based on the tendency of the skeletal shape of the object to be inspected, is defined as the plurality of features. feature point set extraction means for extracting from points;
inspection means for inspecting the inspection object expressed in the overall image, which is specified from the positions of the feature points forming the feature point set, based on the feature point set;
inspection device. 2. The requirement for the positional relationship is a requirement for an arc formed by sequentially connecting the predetermined number of feature points extracted as a set of temporary feature points from the plurality of feature points. The inspection device described in . 3. The inspection apparatus according to claim 2, wherein the requirement for the arc is a requirement for the radius of curvature of the arc. 3. The inspection apparatus according to claim 2, wherein the requirement for the arc is a requirement for the length of the arc. The feature point information output means further outputs information of a part represented by each of the plurality of feature points in the whole image, which is obtained by inputting the whole image into the trained model,
The feature point set extracting means extracts a provisional feature point set composed of the predetermined number of feature points from among the plurality of feature points in consideration of the part information, and extracts the provisional feature point set. 5. The inspection apparatus according to any one of claims 1 to 4, wherein the set of feature points is extracted by determining whether or not the set satisfies the requirements of the positional relationship. The set of feature points indicating the object to be inspected is a set of three feature points representing one end, a central portion, and the other end of the object to be inspected,
The feature point set extracting means extracts, as a provisional feature point set from the plurality of feature points, a feature point whose representative portion is the central portion, two feature points whose representative portions are end portions, 6. The inspection apparatus according to claim 5, wherein the feature point set is extracted by determining whether or not the positional relationship requirements are satisfied after extracting . 6. The requirement for the positional relationship is a requirement that a distance between a feature point representing a central portion and a feature point representing an end portion be within a predetermined range. The inspection device described in . further comprising clipping means for clipping a predetermined range with the position of the feature point belonging to the feature point set as a base point from the entire image as an image of the part of the object to be inspected represented by the feature point;
8. The inspection apparatus according to claim 1, wherein said inspection means inspects said inspection object based on the image cut out by said cutout means. In the feature point information output means, the information capable of specifying the positions of the feature points obtained by inputting the whole image into the trained model is expressed as a median value and a mode value of the positions of the feature points. 9. The inspection apparatus according to any one of claims 1 to 8, wherein the probability density is a probability distribution following a normal distribution. a transport means for transporting the object to be inspected;
irradiating means for irradiating the object to be inspected transported by the transporting means with a predetermined electromagnetic wave;
a detection means arranged at a position capable of detecting an electromagnetic wave irradiated from the irradiation means and having passed through the object to be inspected, for detecting the electromagnetic wave and outputting detection data;
image generation means for generating the entire image based on the detection data and providing it to the feature point information output means;
The inspection device according to any one of claims 1 to 9, comprising: Machine learning is performed using teacher data including a learning image, which is an image in which two or more inspected objects are expressed, and positional information of each of a plurality of feature points in the learning image, which is a correct label. By executing the method, an entire image, which is an image representing two or more inspected objects in which a plurality of feature points are unknown, is input, so that the positions of the plurality of feature points in the entire image can be specified. A learning model generation device that generates a learning model that outputs information. The correct label further includes information on a part represented by each of the plurality of feature points in the learning image,
12. The learning model generating apparatus according to claim 11, wherein the learning model further outputs information of a region represented by each of the plurality of feature points in the entire image. The learning model outputs a probability distribution according to a normal distribution in which the positions of the feature points are expressed as median and mode values as information that can specify the positions of the feature points,
The correct label is a probability distribution whose probability density follows a normal distribution in which the position of the feature point set in the learning image as the correct label is the median value and the mode value,
13. The method according to claim 11, wherein in the machine learning, parameters of the learning model are adjusted so that the probability distribution output by the learning model and the probability distribution of the correct label are more likely to match. The learning model generation device described. A program for causing a computer to function as the inspection apparatus according to any one of claims 1 to 10. A program for causing a computer to function as the learning model generation device according to any one of claims 11 to 13.

Images