JP2020153765A - Inspection device, abnormality detection method, computer program, learning model generation method, and learning model - Google Patents

Abstract

To provide an inspection device, an abnormality detection method, a computer program, a learning model generation method, and a learning model with which it is possible to detect the abnormality of an inspection object with good accuracy.SOLUTION: The inspection device comprises: an acquisition unit 112 for acquiring an image obtained by irradiation of an inspection object placed on a conveyor 61 with an electromagnetic wave; a determination unit 114 for determining presence of abnormality of the inspection object on the basis of a determination value calculated on the basis of the image and a threshold; a learning inspection unit 117 for inputting the image obtained by irradiation of the inspection object with the electromagnetic wave to a learning model 152 that outputs information relating to the presence of abnormality of the inspection object when the image obtained by irradiation of the inspection object with the electromagnetic wave is inputted in accordance with the determination result of the determination unit 114, and acquiring the information relating to the presence of abnormality; and an output unit 118 for outputting the information relating to the presence of abnormality.SELECTED DRAWING: Figure 2

Description

The present invention relates to an inspection device for detecting an abnormality in an inspection object such as food, an abnormality detection method, a computer program, a learning model generation method, and a learning model.

Inspection waves such as X-rays, terahertz waves, infrared rays, and visible light are used for inspection of foods such as meat and beans, inspection of packages containing contents in packaging materials, and inspection of industrial products such as electronic parts. Inspection equipment is used. In this inspection device, the inspection object is irradiated with the inspection wave, the inspection wave that has passed through the inspection object is detected, and the inspection image of the inspection object is acquired.

When the inspection object has a foreign substance, there is a dark part exceeding the threshold value in the inspection image. Conventionally, the presence or absence of foreign matter has been determined based on the area and shape of the dark portion.
The presence or absence of abnormalities in the shape of the inspection object was determined based on the area and peripheral length of the blobs (lumps) in the inspection image. The area of the blob is calculated by counting the pixels constituting the blob, and the peripheral length of the blob is calculated based on the arrangement of the background pixels forming the periphery of the blob (for example, Patent Document 1 and the like).

Japanese Unexamined Patent Publication No. 2005-31069

In the conventional abnormality detection method, the presence or absence of an abnormality is determined by comparing the inspection image of the inspection object acquired based on the predetermined image processing algorithm with the threshold value, but depending on the type of the inspection object and the foreign matter, The difference in pixel value between the normal part and the foreign matter may be small, and there is a problem that the detection accuracy is not good.
Similarly, when detecting an abnormality in the shape of the inspection object, the difference between the pixel values of the blob pixel and the background pixel is small, the edge is not clear, and it may not be detected accurately.

The present invention has been made in view of such circumstances, and provides an inspection device, an abnormality detection method, a computer program, a learning model generation method, and a learning model capable of accurately detecting an abnormality in an inspection object. With the goal.

The inspection device according to one aspect of the present invention includes an acquisition unit that acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and the image. When a judgment unit that determines the presence or absence of an abnormality in the inspection object based on the judgment value and the threshold value calculated based on the above, and an image obtained by irradiating the inspection object with electromagnetic waves according to the judgment result of the judgment unit are input. In addition, a learning inspection unit that acquires information on the presence or absence of the abnormality by inputting an image obtained by irradiating the inspection object with electromagnetic waves into a learning model that outputs information on the presence or absence of the abnormality of the inspection object, and the abnormality. It is provided with an output unit that outputs information regarding the presence or absence of.
Here, the "image obtained by irradiation with electromagnetic waves" includes an image based on electromagnetic waves and an image obtained by image processing the image.

In the abnormality detection method according to one aspect of the present invention, an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is acquired, and the inspection is performed based on a judgment value and a threshold value calculated based on the image. Learning to determine the presence or absence of an abnormality in an object and output information regarding the presence or absence of an abnormality in the inspection object when an image obtained by irradiating the inspection object with electromagnetic waves is input according to the determination result of the presence or absence of the abnormality. A computer is made to execute a process of inputting the acquired image into the model and outputting information indicating the presence or absence of abnormality of the inspection object.

The computer program according to one aspect of the present invention acquires an image obtained by irradiating an inspection object placed on a conveyor with electromagnetic waves, and the inspection object is based on a judgment value and a threshold value calculated based on the image. A learning model that determines the presence or absence of abnormalities in the above, and outputs information on the presence or absence of abnormalities in the inspection object when an image obtained by irradiating the inspection object with electromagnetic waves is input according to the determination result of the presence or absence of the abnormality. The computer is made to execute a process of inputting the acquired image and outputting information indicating the presence or absence of abnormality of the inspection object.

In the method of generating a learning model according to one aspect of the present invention, an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is acquired, and the image and information indicating the presence or absence of abnormality of the inspection object are obtained. When the teacher data recorded in association with and is acquired and the image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves is input based on the teacher data, the abnormality of the inspection object is abnormal. Generate a learning model that outputs information about the presence or absence.

The learning model according to one aspect of the present invention has an input layer into which an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is input, and an output for outputting information regarding the presence or absence of abnormality in the inspection object. An intermediate in which parameters are learned using teacher data recorded by associating an image obtained by irradiating a layer and an inspection object placed on a conveyor with an electromagnetic wave and information indicating the presence or absence of an abnormality in the inspection object. When an image obtained by irradiating an inspection object with an electromagnetic wave is input to the input layer, information regarding the presence or absence of abnormality of the inspection object is output from the output layer through calculations by the intermediate layer. Make your computer work as you do.

According to the present invention, with respect to an image based on an image obtained by irradiating an inspection object with electromagnetic waves, an image is input to a learning model according to the presence or absence of an abnormality in the inspection object determined based on a judgment value and a threshold value. To get the presence or absence of abnormality. For example, when it is uncertain whether or not there is an abnormality in the inspection object determined based on the threshold value, by inputting an image into the learning model and acquiring the presence or absence of the abnormality, the presence or absence of foreign matter, the presence or absence of the shape abnormality, etc. can be accurately determined. The presence or absence of abnormality can be detected.

It is a perspective view which shows the structure of the abnormality detection system which concerns on Embodiment 1. FIG. It is a block diagram which shows the structure of an abnormality detection system. It is explanatory drawing which shows the state which blob was extracted about the inspection object which is a fish meat. It is explanatory drawing which shows the state which cut out each blob. It is explanatory drawing which shows the state which the blob is surrounded by the bounding box. It is explanatory drawing of the process of resizing. It is explanatory drawing which shows the inspection image including the blob which was judged that the abnormality value is 0.5 or more by the abnormality detection model. It is explanatory drawing which shows an example of the record layout of the inspection image DB. It is explanatory drawing about the generation processing of an abnormality detection model. It is a flowchart which shows an example of the processing procedure of the generation processing of the abnormality detection model by a control unit. It is a flowchart which shows an example of the processing procedure of the abnormality detection processing by a control unit. It is explanatory drawing which shows an example of the display screen in a display. It is a flowchart which shows an example of the processing procedure of the relearning process by the control part of a PC. It is a flowchart which shows an example of the processing procedure of the relearning process of the modification by the control part of a PC. It is a block diagram which shows the structure of the abnormality detection system which concerns on Embodiment 2. FIG. It is explanatory drawing which shows an example of the display screen which displays the selection button of levels 1 to 4. It is explanatory drawing which shows an example of the display screen which displays the threshold value a and b of levels 1 to 4. It is a flowchart which shows an example of the processing procedure of the abnormality detection processing by a control unit. It is a block diagram which shows the structure of the abnormality detection system which concerns on Embodiment 3. It is a flowchart which shows an example of the processing procedure of the process of setting a threshold value by a control unit. It is explanatory drawing which shows an example of the display screen for inputting the threshold of a level.

Hereinafter, the present invention will be described in detail with reference to the drawings showing the embodiments thereof.
(Embodiment 1)
FIG. 1 is a plan view showing the configuration of the abnormality detection system 10 according to the first embodiment, and FIG. 2 is a block diagram showing the configuration of the abnormality detection system 10.
The abnormality detection system 10 includes an information processing device 1 as an inspection device and an X-ray inspection machine 2. The information processing device 1 and the X-ray inspection machine 2 may be integrated. In the present embodiment, the inspection object 4 is placed on the transfer belt 61 of the transfer unit (conveyor) 6 of the X-ray inspection machine 2, and the inspection object 4 is abnormal based on the image captured by using X-rays. The abnormality detection system 10 for detecting the above will be described. The abnormality detection system 10 determines the presence or absence of an abnormality in the inspection object 4 based on a predetermined image processing algorithm, and estimates the presence or absence of an abnormality based on the abnormality detection model 152 according to the result. An example of the inspection product 4 is food such as raw meat and fish meat.

The X-ray inspection machine 2 includes a horizontally long rectangular parallelepiped lower housing 23 and an upper housing 21 provided on the lower housing 23 and smaller than the lower housing 23.
A transport unit 6 is provided in the lower housing 23. The transport portion 6 is provided with an upstream roller 63 outside one end in the longitudinal direction of the lower housing 23, and a downstream roller 62 outside the other end. Two lower rollers 64, 65 are provided between the upstream roller 63 and the downstream roller 62 below the upstream roller 63 and the downstream roller 62, and the conveyor belt 61 is attached to these rollers. It has been bridged. A driving force from a transfer motor (not shown) is applied to either the upstream roller 63 or the downstream roller 62, and when the object to be inspected is inspected, the transfer belt 61 rotates clockwise at a constant speed V. Go around.

The upper housing 21 houses an X-ray irradiation unit 3 that irradiates X-rays as electromagnetic waves. The X-ray irradiation unit 3 has an X-ray tube 31. Examples of electromagnetic waves other than X-rays include terahertz waves, infrared rays, and visible light.
A display 22 is provided on the front surface of the upper housing 21. The display 22 is a display device such as an LCD (Liquid Crystal Display).
An X-ray detection unit 7 is provided below the transport belt 61 that moves between the upstream roller 63 and the downstream roller 62. The X-ray detector 7 is called a TDI (Time Delay Integration) camera or TDI sensor, and includes a scintillator that emits fluorescence according to the intensity of X-rays and a large number of photodiodes that detect the fluorescence emitted by the scintillator. have.

In the TDI camera, one detection line 71 is configured by arranging photodiodes linearly in the X direction orthogonal to the Y direction, which is the moving direction of the object to be inspected, and a plurality of detection lines are provided. The plurality of detection lines 71 are arranged side by side in parallel in the Y direction, which is the moving direction of the object to be inspected. In FIG. 1, each detection line 71 is shown as an independent line sensor element, but in reality, a plurality of photodiodes are regularly arranged in the X direction and the Y direction in the housing of one TDI camera. One detection line 71 is composed of photodiodes arranged side by side and arranged in the X direction. An X-ray image is composed of the data of each line received and output by each detection line 71.

The information processing device 1 is an information processing device capable of transmitting and receiving various types of information processing and information, and is, for example, a personal computer, a server device, or the like. The X-ray inspection machine 2 is communicatively connected to the information processing device 1 via a network N such as a LAN. When the information processing device 1 is a server device, the network N may be the Internet. Further, a cloud computer connected so as to be communicable via a network N such as the Internet may execute the processing of the information processing device 1.
In the present embodiment, it is assumed that the information processing device 1 is a personal computer, and in the following description, it is read as PC1 for the sake of brevity.

The PC 1 acquires an image obtained by capturing an image of the inspection object 4, and performs a process of detecting the presence or absence of an abnormality in the inspection object 4 based on the image. In the present embodiment, the PC 1 has been trained to detect (identify) an abnormality of the inspection object 4 from the image by machine learning according to the determination result of the abnormality determination unit (determination unit) 114 described later. The presence or absence of abnormality is detected using 152.
The PC 1 obtained an image obtained by irradiating the inspection object 4 with X-rays using an X-ray inspection machine 2 (a region surrounding the blob 41 was cut out as described later, and the region was subjected to predetermined image processing. The inspection image) is input to the abnormality detection model 152, and the identification result indicating the presence or absence of the abnormality of the blob 41 is acquired as an output.
Hereinafter, the type of abnormality of the inspection object 4 is the presence or absence of foreign matter, the PC1 inputs the inspection image into the abnormality detection model 152, and the probability value of normal (without foreign matter) and the probability value of abnormality (with foreign matter) by the softmax function are used. Will be output, and the case where the probability value of abnormality is acquired as an abnormal value will be described.
The PC 1 selects whether or not to discard the inspection object 4 according to the acquired abnormal value.

The PC 1 includes a control unit 11, a main storage unit 12, a communication unit 13, a display unit 14, an auxiliary storage unit 15, and an input unit 16.
The control unit 11 has one or more CPUs (Central Processing Units), MPUs (Micro-Processing Units), GPUs (Graphics Processing Units), and other arithmetic processing units, and stores the program P stored in the auxiliary storage unit 15. By reading and executing, various information processing, control processing, and the like related to the PC1 are performed. Each functional unit of FIG. 2 is executed by the control unit 11 operating based on the program P.

As functional units, the control unit 11 includes an image generation unit 111, an image processing unit (acquisition unit) 112, a target identification unit (specific unit) 113, an abnormality determination unit (determination unit) 114, a cutting unit 115, and an image processing unit 116. It has a learning inspection unit 117, a display unit (output unit) 118, a sorting unit (output unit) 119, a pass / fail judgment reception unit (second reception unit) 122, and a re-learning unit 123.

The image generation unit 111 passes through the inspection object 4 and generates an X-ray image based on the X-rays detected by the X-ray detection unit 7.
The image processing unit 112 performs image processing on the X-ray image by a smoothing filter, a feature extraction filter, or the like.

The target identification unit 113 analyzes the blob 41 and identifies the blob (region). The target identification unit 113 extracts, for example, an emphasis line obtained by enhancing the brightness change (edge detection) of each pixel, and specifies a portion including the emphasis line as a blob.
The target identification unit 113 may specify a blob whose brightness change rate is equal to or higher than a predetermined value as compared with the peripheral region.
The target identification unit 113 may obtain a standard deviation of the brightness and specify a set of pixels having a portion where the brightness changes by a predetermined level or more as compared with the standard deviation as a blob, and a singular part having a different brightness. The portion containing the blob may be specified as a blob.

The abnormality determination unit 114 determines whether or not the blob is a foreign substance based on the determination value and the threshold value for the blob specified by the target identification unit 113. Examples of the judgment value include the number of images of the blob surrounded by the emphasized line, the area, and the like. The normal side threshold value (first threshold value) a and the abnormal side threshold value (second threshold value) b are set as the judgment values. Every time (a <b), if the calculated judgment value is less than the threshold value a, it is determined that the singular portion is not a foreign substance and is normal, and if the judgment value exceeds the threshold value b, it is determined that the blob is a foreign substance. When the judgment value is a or more and b or less, an abnormal value is acquired by the learning inspection unit 117 described later.
FIG. 3 shows, as an example, a state in which a blob 41 having a judgment value of a or more and b or less is extracted from the inspection object 4 which is fish meat. The blobs are highlighted and filled by image processing.

The cutout unit 115 cuts out the blob 41 having a determination value of a or more and b or less so that it fits within a rectangle of a predetermined size, surrounds it with a bounding box 81, and generates a cutout image. Here, the image surrounding the blob 41 is an image before the enhancement process. The image surrounding the blob 41 may be an image after enhancement processing.
FIG. 4 shows a state in which each blob 41 is cut out.
The image processing unit 116 performs processing such as normalization on the cutout image to generate an inspection area (inspection image) 8 to be input to the abnormality detection model 152.

FIG. 5 is an explanatory diagram showing a state in which the blob specified by the target specifying unit 113 is cut out by the cutting unit 115, and the inspection image 8 formed by surrounding the blob 41 with the bounding box 81 is generated. The target identification unit 113 of the control unit 11 detects the blob 41 as described above. In each inspection image 8, each blob 41 is surrounded by a bounding box 81 of the same size.
FIG. 6 is an explanatory diagram of the resizing process by the cutout portion 115.
When the inspection object 4 protrudes from the bounding box 81, the cutout portion 114 of the control unit 11 is resized so that the blob 41 fits in the bounding box 81. When the blob 41 is enlarged, a complementary process is performed.

The learning inspection unit 117 inputs each inspection image into the abnormality detection model 152 and acquires an abnormality value.

The display unit 118 outputs an abnormal value to the display 22, and displays the presence or absence of an abnormality in the inspection image on the display 22.
FIG. 7 shows an inspection image including a blob whose outlier is determined to be 0.5 or more by the anomaly detection model 152.

The sorting unit 119 compares the abnormal value of the inspection image 8 with the threshold value, selects whether or not to discard the inspection object 4 including the blob 41 showing the abnormal value, stops the transport unit 6, and inspects. Mark the object 4. For example, when the abnormal value of one blob 41 is 0.5 or more, the sorting unit 119 considers the inspection object 4 containing the blob 41 as an abnormal product and sorts it when it is discarded. The outlier threshold is not limited to 0.5. It should be noted that the present invention is not limited to the case where the inspection object 4 is discarded when one blob 41 having an abnormal value equal to or higher than the threshold value is included. When a predetermined number or more of blobs 41 show an abnormal value equal to or more than a threshold value, the inspection object 4 may be discarded.

The pass / fail judgment reception unit 122 receives the pass / fail judgment of the judgment of the abnormality detection model 152 input by the operator through the input unit 16.
The re-learning unit 123 re-learns the abnormality detection model 152 by using the teacher data in which the inspection image 8 and the above-mentioned good / bad judgment input from the operator are associated with each other.

The main storage unit 12 is a temporary storage area for SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), flash memory, etc., and temporarily stores data necessary for the control unit 11 to execute arithmetic processing. Remember. The communication unit 13 is a communication module for performing processing related to communication, and transmits / receives information to / from the outside.

The auxiliary storage unit 15 is a large-capacity memory, a hard disk, or the like, and stores a program P and other data necessary for the control unit 11 to execute processing. In addition, the auxiliary storage unit 15 stores the inspection image DB 151 and the abnormality detection model 152. The inspection image DB 151 is a database that stores information on the inspection object 4 that is the abnormality detection target.

The program P stored in the auxiliary storage unit 15 may be provided by a recording medium 156 in which the program P is readablely recorded. The recording medium 156 is, for example, a portable memory such as a USB (Universal Serial Bus) memory, an SD (Secure Digital) card, a micro SD card, and a compact flash (registered trademark). The program P recorded on the recording medium 156 is read from the recording medium 156 using a reading device (not shown) and installed in the auxiliary storage unit 15. Further, when the information processing device 1 includes a communication unit capable of communicating with an external communication device, the program P stored in the auxiliary storage unit 15 may be provided by communication via the communication unit.
The auxiliary storage unit 15 may be an external storage device connected to the PC1. Further, the PC 1 may be a multi-computer composed of a plurality of computers, or may be a virtual machine virtually constructed by software.

The input unit 16 receives an input such as a threshold value of a determination value described later by an operator.

FIG. 8 is an explanatory diagram showing an example of the record layout of the inspection image DB 151. The inspection image DB 151 includes an inspection image ID string, a coordinate information string, an outlier string, and a judgment value string. The inspection image ID column stores an inspection image ID for identifying each inspection image 8. The coordinate information sequence stores coordinate information indicating the position of the blob 41 in the X-ray irradiation region on the transport belt 61 in association with the inspection image ID. The coordinate information is the X coordinate and the Y coordinate at the four corners of the bounding box 81 when the center in the X direction and the Y direction is the origin between the upstream roller 63 and the downstream roller 62 in FIG. As described above, the outlier is the probability value of the anomaly. The determination value is a determination value calculated by the abnormality determination unit 114.

FIG. 9 is an explanatory diagram relating to the generation process of the abnormality detection model 152. FIG. 9 conceptually shows a process of performing machine learning to generate an abnormality detection model 152.

As the abnormality detection model 152, the control unit 11 of the PC1 learns the image feature amount of the abnormality in the inspection image 8 of the blob 41, inputs the inspection image 8, and outputs information indicating the presence or absence of the abnormality in the blob 41. A neural network is constructed (generated). For example, the neural network is a CNN (Convolution Neural Network), which includes an input layer that accepts the input of the inspection image 8, an output layer that outputs the identification result of the presence or absence of an abnormality, and an intermediate layer that extracts the image feature amount of the inspection image 8. Has.

The input layer has a plurality of neurons that receive input of pixel values of each pixel included in the inspection image 8, and passes the input pixels to the intermediate layer. The intermediate layer has a plurality of neurons for extracting the image features of the inspection image, and the extracted image features are passed to the output layer. In FIG. 9, the number of layers of the intermediate layer is set to 3, but the number of layers is not limited to this. For example, when the abnormality detection model 152 is CNN, the intermediate layer alternates between a convolution layer that convolves the pixel values of each pixel input from the input layer and a pooling layer that maps the pixel values convoluted by the convolution layer. The feature amount of the image is finally extracted while compressing the pixel information of the inspection image 8. In FIG. 9, the description of the convolution layer and the pooling layer is omitted. The output layer has two neurons that output the identification result of identifying the abnormality in the blob 41, and identifies the presence or absence of the abnormality in the blob 41 based on the image feature amount output from the intermediate layer. One neuron outputs the probability value of normal (without foreign matter) in blob 41, and the other neuron outputs the probability value of abnormality (with foreign matter) in blob 41.

In the present embodiment, the anomaly detection model 152 is described as being a CNN, but the anomaly detection model 152 is not limited to the CNN, and is a neural network other than the CNN, an SVM (Support Vector Machine), a Bayesian network, and a regression tree. It may be a trained model constructed by another learning algorithm.

The control unit 11 has a teacher data (whether it has a foreign substance or does not have a foreign substance and is normal) in which a plurality of inspection images 8 and information indicating an abnormality of the blob 41 in each inspection image 8 are associated with each other. Learning is performed using (abnormal data, normal data). For example, as shown in FIG. 9, the teacher data is labeled with the ID of the inspection image 8 and the type of abnormality (presence or absence of foreign matter) and rank (1: abnormal, 0: normal) with respect to the inspection image 8 of the blob 41. It is the attached data.

The control unit 11 inputs the inspection image 8 which is the teacher data to the input layer, performs arithmetic processing in the intermediate layer, and acquires an identification result indicating the presence or absence of an abnormality in the blob 41 from the output layer. The identification result output from the output layer is a continuous probability value (for example, a value in the range of "0" to "1") when the softmax function is used. The identification result may be a value that discretely indicates the presence or absence of an abnormality (for example, a value of "0" or "1").

The control unit 11 compares the identification result output from the output layer with the information labeled for the inspection image 8 in the teacher data, that is, the correct answer value, so that the output value from the output layer approaches the correct answer value. Optimize the parameters used for arithmetic processing in the intermediate layer. The parameters include, for example, the weight between neurons (coupling coefficient), the coefficient of the activation function used in each neuron, and the like. The method of optimizing the parameters is not particularly limited, but for example, the control unit 11 optimizes various parameters by using the backpropagation method.

The control unit 11 performs the above processing on each inspection image 8 included in the teacher data to generate the abnormality detection model 152. When the control unit 11 acquires the inspection image 8 of the blob 41, the control unit 11 detects the presence or absence of foreign matter in the blob 41 by using the abnormality detection model 152.

FIG. 10 is a flowchart showing an example of a processing procedure of the generation processing of the abnormality detection model 152 by the control unit 11.
The control unit 11 acquires teacher data in which the inspection images 8 of the plurality of blobs 41 and the information indicating the abnormality of the blobs 41 in each inspection image 8 are associated with each other (S11).

The control unit 11 uses the teacher data to generate an abnormality detection model 152 (learned model) that outputs the abnormality presence / absence information of the blob 41 when the inspection image 8 of the blob 41 is input (S12). Specifically, the control unit 11 inputs the inspection image 8 which is the teacher data into the input layer of the neural network, and acquires the identification result of identifying the presence / absence of foreign matter in the inspection image 8 from the output layer. The control unit 11 compares the acquired identification result with the correct answer value of the teacher data (information labeled for the inspection image 8), and in the intermediate layer, the identification result output from the output layer approaches the correct answer value. Optimize the parameters (weights, etc.) used in the arithmetic processing of. The control unit 11 stores the generated abnormality detection model 152 in the auxiliary storage unit 15 and ends a series of processes.

FIG. 11 is a flowchart showing an example of the processing procedure of the abnormality detection process by the control unit 11.
The control unit 11 of the PC 1 passes through the inspection object 4 and generates an X-ray image based on the X-rays detected by the X-ray detection unit 7 (S21).
The control unit 11 performs image processing on the X-ray image by a smoothing filter, a feature extraction filter, or the like (S22).
The control unit 11 performs blob analysis as described above to identify the blob (S23). The control unit 11 extracts an emphasis line obtained by enhancing the brightness change of each pixel (edge detection), and identifies a portion including the emphasis line as a blob.

The control unit 11 determines whether or not the blob is a foreign substance (S24). The control unit 11 calculates, for example, a determination value such as the number of images and an area of the blob surrounded by the highlighted line, and makes a determination based on the calculated determination value and the threshold value. The control unit 11 determines whether it is abnormal (NG: judgment value> threshold value b), normal (OK: judgment value <threshold value a), or intermediate (a ≦ judgment value ≦ b).

The control unit 11 determines in S24 whether or not it is determined to be NG (abnormal) or OK (normal) (S25). When the control unit 11 determines that it is NG or OK (S25: YES), it determines whether or not it is OK (S26).
When the control unit 11 determines that the result is OK (S26: YES), the learning test is skipped and the process ends.
When the control unit 11 determines that it is not OK, that is, NG (S26: NO), the learning test is skipped and the process proceeds to S33.

When the control unit 11 determines that it is not NG or OK (S25: NO), the control unit 11 cuts out the blob 41 determined not to be NG or OK so as to be inside a rectangle of a predetermined size, and generates a cut-out image. (S27). The cut-out image is an image of the blob 41 before the filling process.
The control unit 11 performs processing such as normalization on the cut-out image to generate an inspection image (S28).

The control unit 11 inputs the inspection image 8 to the abnormality detection model 152, outputs a normal probability value and an abnormality probability value, and acquires the abnormality value (S29).
The control unit 11 determines whether or not the abnormal value is equal to or greater than the threshold value (S30). When the control unit 11 determines that the abnormal value is equal to or higher than the threshold value (S30: YES), the control unit 11 outputs the abnormal value to the display 22 and displays the inspection image 8 in a state of being surrounded by the red bounding box (BB) 82 (S30: YES). S31). When the control unit 11 determines that the abnormal value is not equal to or higher than the threshold value (S30: NO), the control unit 11 displays the inspection image 8 in a state where it is not surrounded by the red bounding box (BB) 82 (S32).
The control unit 11 identifies the inspection object 4 based on the abnormal value of the inspection image 8 and the inspection image ID, and if it is determined to be NG in S26, selects whether or not to discard the inspection object 4 ( S33), the process is terminated.

As described above, according to the present embodiment, even when the difference between the pixel values of the foreign matter and the other part is small in the inspection image 8 and it is difficult to determine the presence or absence of the foreign matter by the conventional image processing method. By using the abnormality detection model 152, foreign matter can be discriminated with high accuracy.
According to the present embodiment, with respect to the blob 41 acquired based on the predetermined image processing algorithm, the determination value and the determination result of whether or not the blob 41 is a foreign substance acquired based on the threshold values a and b are accurate. If this is not the case, the presence or absence of foreign matter can be detected with high accuracy by inputting the inspection image 8 into the abnormality detection model 152 and acquiring the abnormality value.
It is not necessary for the operator to visually detect the abnormality, and the efficiency of the abnormality detection work can be improved.
In the above-described embodiment, the presence or absence of a foreign substance is mentioned as the type of abnormality, but the present invention is not limited to this, and other abnormalities such as an abnormality in shape may be detected.

When the judgment value is represented by a numerical value from 0 to 1, for example, 0 may be the threshold value a and 1 may be the threshold value b. In this case, when the judgment value is 0, it is normal, and when it is 1. It is abnormal.
Further, the learning test is not limited to the case where the judgment value is a or more and b or less, and the learning test may be performed when the judgment value is a or more and the learning test is performed when the judgment value is b or less. May be done.
The image processing of S28 can be omitted. The inspection image 8 may be an image after the blob 41 has been enhanced.

FIG. 12 is an explanatory diagram showing an example of a display screen 5 on the display 22.
The display screen has a first screen 51 on the left side in FIG. 12 and a second screen 52 on the right side.
The control unit 11 displays the inspection image 8 on the first screen 51 in association with the inspection object 4 mounted on the transport belt 61 in association with the position of the transport belt 61. On the second screen 52, the control unit 11 displays the inspection images 8 corresponding to the blobs 41 in the order of the inspection image IDs, and the abnormal inspection image 8 is surrounded by the red bounding box 82. In the case of FIG. 12, when the inspection image 8 related to the blob 41 is surrounded by the bounding box 82, the blob 41 is a foreign substance, and the inspection object 4 including the blob 41 is an abnormal product.

The operator can select any of the inspection images 8 on the second screen 52 by the selection button 56 consisting of ">" and "<" for sequentially moving in the row direction of the inspection image 8. The inspection image 8 may be selected by a pointing device such as a mouse.
The second screen 52 also includes an OK button 53 that is pressed when the operator determines that the estimation result (judgment) of the abnormality detection model 152 is correct, and an NG button 54 that is pressed when the determination is incorrect. It also has the Other button 55, which is used when inputting an instruction for selecting the level of presence / absence of abnormality, which will be described later.

For the inspection image 8 selected by the selection button 56, the control unit 11 receives input from the operator as to whether or not the judgment of the abnormality detection model 152 is good or bad by the OK button 53 or the NG button 54, and when a predetermined number or more of data are collected, the control unit 11 , The inspection image 8 and the quality of the judgment input from the operator can be used as teacher data to relearn the abnormality detection model 152.

FIG. 13 is a flowchart showing an example of the processing procedure of the above-mentioned relearning process by the control unit 11 of the PC1.

The control unit 11 accepts the determination of the quality of the determination of the abnormality detection model 152 based on the pressing of the OK button 53 or the NG button 54 on the second screen 52 by the operator for the selected inspection image 8 (S41). When the control unit 11 determines that the quality determination is not accepted (S41: NO), the determination process is repeated.
When it is determined that the control unit 11 has received a predetermined number of good / bad judgments (S41: YES), the inspection image 8 of the blob 41 that has detected the abnormality corresponds to the good / bad judgment of the control unit 11 input from the operator. Acquire the attached teacher data (S42).
The control unit 11 uses the teacher data to generate an abnormality detection model 152 that outputs the abnormality presence / absence information of the blob 41 when the inspection image of the blob 41 is input (S43).
With the above, it is relearned. As a result, the abnormality detection model 152 is updated as the abnormality detection process is continued, and more accurate abnormality detection can be performed.
The judgment of the quality of the abnormality detection model 152 by the operator may be performed based on the X-ray image instead of the inspection image 8.

FIG. 14 is a flowchart showing an example of the processing procedure of the re-learning process of the modified example by the control unit 11 of the PC1.

The control unit 11 accepts the determination of the quality of the determination of the abnormality detection model 152 based on the pressing of the OK button 53 or the NG button 54 on the second screen 52 by the operator for the selected inspection image 8 (S51). When the control unit 11 determines that the quality determination is not accepted (S51: NO), the determination process is repeated.
When it is determined that the quality determination has been accepted (S51: YES), the control unit 11 determines whether or not the determination has been accepted (S52). When the control unit 11 determines that the rejection is not accepted (S52: NO), the process returns to S51.
When the control unit 11 determines that the result is rejected (S52: YES), the teacher who associates the inspection image 8 of the blob 41 that detected the abnormality with the determination result when a predetermined number of the results of the determination of no are obtained. Acquire data (S53).
The control unit 11 uses the teacher data to generate an abnormality detection model 152 that outputs the abnormality presence / absence information of the blob 41 when the inspection image of the blob 41 is input (S54).
With the above, it is relearned. This makes it possible to perform accurate abnormality detection with a smaller amount of teacher data.

(Embodiment 2)
In the PC1 of the abnormality detection system 10 according to the second embodiment, the control unit 11 detects the presence or absence of foreign matter based on four image processing algorithms, and the auxiliary storage unit 15 stores the program P corresponding to each image processing algorithm. However, the same processing as that of the PC 1 according to the first embodiment is performed except that a plurality of abnormality detection models 152 to 155 are stored. The types of foreign substances detected based on the four image processing algorithms may be all the same, some may be the same, or all may be different. Examples of the type of foreign matter include bone, metal, wood, synthetic resin and the like. When the types of foreign substances are the same, the method of calculating the judgment value is different.
Further, the image processing algorithm may be used to detect different types of abnormalities in the inspection object 4. Types of anomalies include the presence or absence of foreign matter, abnormal shape, and the like.
The number of image processing algorithms is not limited to four.

For example, in the level 1 image processing algorithm, as in the first embodiment, the target identification unit 113 specifies the portion including the highlighted line extracted by highlighting the brightness change (edge detection) of each pixel as the blob 41. .. The abnormality determination unit 114 makes a determination based on, for example, determination values such as the number of images and the area of the blob 41 surrounded by the highlighted line. The threshold values a and b are set as the judgment values, and if the judgment value is less than the threshold value, for example, it is determined that the singular portion is not a foreign substance.
In the level 2 image processing algorithm, the target identification unit 113 extracts a singular portion whose brightness change rate is equal to or higher than a predetermined value as the blob 41 as compared with the peripheral region. The abnormality determination unit 114 makes a determination based on, for example, a determination value such as the number of pixels whose rate of change is equal to or greater than a predetermined value.
In the level 3 image processing algorithm, the standard deviation of the brightness is obtained, and the singular portion where the brightness changes by a predetermined level or more as compared with the standard deviation is specified as the blob 41. In the level 4 image processing algorithm, the singular portion having different brightness is specified as the blob 41. The abnormality determination unit 114 determines whether or not the shape of these peculiar portions can be, for example, the shape of a bone by an image matching process.

FIG. 15 is a block diagram showing the configuration of the abnormality detection system 10. The same parts as those in FIG. 2 are designated by the same reference numerals, and detailed description thereof will be omitted.
The control unit 11 has a level reception unit (first reception unit) 120. The level corresponds to the image processing algorithm for detecting the presence or absence of foreign matter, and here, it is assumed that there are four types of levels 1 to 4.
The level reception unit 120 accepts the selection of the level by the worker. As shown in FIG. 16, when the operator presses the Other button 55, the control unit 11 displays the buttons 57a to 57d of "level 1" to "level 4" on the first screen 51. The operator selects the desired level of buttons.
Based on the image processing algorithm corresponding to the selected level, the target identification unit 113 identifies the blob 41, and the abnormality determination unit 114 determines whether the blob 41 is abnormal or normal. Depending on the image processing algorithm, not only the processing of the target identification unit 113 and the abnormality determination unit 114 may be different, but also the processing of the image processing unit 112 and the image processing unit 116 may be different.

The abnormality determination unit 114 calculates a determination value for the blob 41 specified by the target identification unit 113 based on the image processing algorithm. At level 1, for example, when the target identification unit 113 specifies as the blob 41 the portion including the emphasis line extracted by emphasizing the brightness change of each pixel, the abnormality determination unit 114 is the blob 41 surrounded by the emphasis line. Judgment values such as the number of images and area are calculated, and the judgment is made based on the calculated judgment values. A threshold value a on the normal side and a threshold value b on the abnormal side are set for the judgment value (a <b), and if the calculated judgment value is less than the threshold value a, the singular part is judged to be normal and not a foreign substance, and the judgment value is determined. If is greater than or equal to the threshold value b, it is determined that the singular portion is a foreign substance and the blob 41 is abnormal.

As shown in FIG. 17, when the operator presses the Other button 55 twice in succession, the control unit 11 displays the threshold values a and the threshold values b of "level 1" to "level 4" on the first screen 51. indicate. The judgment value of each level is calculated based on each image processing algorithm, the content differs depending on the level, each threshold value a and threshold value b also differ, and the range of normal and abnormal is also different.

Similar to the above, the control unit 11 inspects the abnormality by learning the image feature amount of the abnormality in the inspection image 8 related to the blob 41 generated by the image processing algorithm corresponding to the level 1 as the abnormality detection model 152. A neural network is generated in which image 8 is input and information indicating the presence or absence of foreign matter is output. The neural network is a CNN, and has an input layer that accepts the input of the inspection image 8, an output layer that outputs the identification result of the presence or absence of an abnormality, and an intermediate layer that extracts the image feature amount of the inspection image 8.

The control unit 11 performs learning by using the teacher data in which the plurality of inspection images 8 and the information indicating the abnormality of the blob 41 in each inspection image 8 (whether it has a foreign substance or is normal) are associated with each other. The teacher data is data in which the ID of the inspection image 8, the type of abnormality (presence or absence of foreign matter), and the rank (1: abnormality, 0: normal) are labeled with respect to the inspection image 8 of the blob 41.
The control unit 11 inputs the inspection image 8 which is the teacher data to the input layer, and after the arithmetic processing in the intermediate layer, the identification result indicating the presence or absence of the abnormality of the blob 41 from the output layer (normal probability value and abnormality (foreign matter)). (Yes) Probability value) is output, and the probability value of abnormality is acquired as an abnormal value. The identification result output from the output layer is a continuous probability value (for example, a value in the range of "0" to "1") when the softmax function is used.

The control unit 11 compares the identification result output from the output layer with the information labeled for the inspection image 8 in the teacher data, that is, the correct answer value, so that the output value from the output layer approaches the correct answer value. Optimize the parameters used for arithmetic processing in the intermediate layer. The parameters include, for example, the weight between neurons (coupling coefficient), the coefficient of the activation function used in each neuron, and the like. The method of optimizing the parameters is not particularly limited, but for example, the control unit 11 optimizes various parameters by using the backpropagation method.

The control unit 11 performs the above processing on each inspection image 8 included in the teacher data to generate the abnormality detection model 152. When the control unit 11 acquires the inspection image 8 of the blob 41, the control unit 11 detects the presence or absence of an abnormality in the blob 41 by using the abnormality detection model 152.
Similarly, the abnormality detection models 153 to 155 input the inspection image 8 by learning the image feature amount of the abnormality in the inspection image 8 related to the blob 41 generated by the image processing algorithm corresponding to the levels 2 to 4. To generate a neural network that outputs information indicating the presence or absence of foreign matter.

The learning inspection unit 117 inputs the inspection image 8 into one of the abnormality detection models 152 to 155 according to the selected level, outputs the above two probability values, and outputs the probability value (outlier value) of having a foreign substance. get.

FIG. 18 is a flowchart showing an example of the processing procedure of the abnormality detection process by the control unit 11.
The control unit 11 of the PC 1 passes through the inspection object 4 and generates an X-ray image based on the X-rays detected by the X-ray detection unit 7 (S61).
The control unit 11 determines whether or not the operator has accepted the level as described above (S62). When the control unit 11 determines that the level is not accepted (S62: NO), the determination is repeated.
When the control unit 11 determines that the level has been accepted (S62: YES), the control unit 11 reads the program P of the corresponding level from the auxiliary storage unit 15 (S63).
The control unit 11 performs image processing on the X-ray image by a smoothing filter, a feature extraction filter, or the like based on the image processing algorithm corresponding to the received level (S64).
The control unit 11 performs blob analysis and identifies the blob 41 based on the image processing algorithm corresponding to the received level (S65).

The control unit 11 calculates a determination value (S66).
The control unit 11 determines whether or not a ≦ determination value ≦ b (S67). When the control unit 11 determines that a ≤ determination value ≤ b is not satisfied (S67: NO), the determination value is less than a or exceeds b, and therefore determines whether or not the determination value <a (S68).
When the control unit 11 determines that the determination value <a (S68: YES), the control unit 11 has no foreign matter and is normal, so the learning test is skipped and the process ends.
When the control unit 11 has a determination value <a, that is, a determination value> b, has a foreign substance, and definitely determines that the abnormality is abnormal (S68: NO), the learning test is skipped and the process proceeds to S76.

When the control unit 11 determines that a ≦ judgment value ≦ b (S67: YES), the control unit 11 cuts out the blob 41 having a ≦ judgment value ≦ b so as to be inside a rectangle of a predetermined size, and generates a cut-out image. (S69).
The control unit 11 performs processing such as normalization on the cut-out image to generate an inspection image (S70).
The control unit 11 selects the abnormality detection model 152 corresponding to the received level (S71).
The control unit 11 inputs the inspection image 8 to the abnormality detection model 152, outputs a normal probability value and an abnormality probability value, and acquires the abnormality value (S72).
The control unit 11 determines whether or not the abnormal value is equal to or greater than the threshold value (S73). When the control unit 11 determines that the abnormal value is equal to or higher than the threshold value (S73: YES), the control unit 11 outputs the abnormal value to the display 22 and displays the inspection image 8 in a state of being surrounded by the red bounding box (BB) 82 (S73: YES). S74). When the control unit 11 determines that the abnormal value is not equal to or higher than the threshold value (S73: NO), the control unit 11 displays the inspection image 8 in a state where it is not surrounded by the red bounding box (BB) 82 (S75).
Based on the abnormal value of the blob 41 and the inspection image ID, the control unit 11 identifies the inspection object 4 when it is determined to be NG in S68, and selects whether to discard the inspection object 4 (S76). ), End the process.
In this embodiment, an image processing algorithm can be selected to detect the presence or absence of anomalies for a desired type of anomaly at a desired level.
Further, when the target identification of S65 is performed based on the X-ray image, the image processing of S64 can be omitted. The image processing of S70 can also be omitted.

(Embodiment 3)
The PC 1 according to the third embodiment performs the same processing as the PC according to the first embodiment except that the control unit 11 has the threshold value setting unit (setting unit) 121.
FIG. 19 is a block diagram showing the configuration of the abnormality detection system 10. The same parts as those in FIG. 15 are designated by the same reference numerals, and detailed description thereof will be omitted.
The control unit 11 has a threshold value setting unit 121. When the operator accepts the setting of the threshold value a and the threshold value b for any of the levels by the input unit 16, the threshold value setting unit 121 sets (updates) the threshold value.

FIG. 20 is a flowchart showing an example of a processing procedure of the threshold value setting process by the control unit 11.
The control unit 11 determines whether or not the operator has accepted the selection of the level (S81). When the operator presses the Other button 55 once, the control unit 11 displays the first screen 51 shown in FIG. 16 described above, and the operator selects a level. When the control unit 11 determines that the level is not accepted (S81: NO), the control unit 11 repeats this determination process.
When the control unit 11 determines that the selection of the level has been accepted (S81: YES), the control unit 11 determines whether or not the threshold value a and the threshold value b have been accepted for this level (S82). As shown in FIG. 21, when the level is selected, the control unit 11 displays the threshold value a and the threshold value b of the level on the first screen 51. In FIG. 21, the screen when level 1 is selected is displayed. The operator inputs the threshold a and the threshold b of the selected level. When the control unit 11 determines that the threshold value a and the threshold value b are not accepted (S82: NO), this determination process is repeated.
When the control unit 11 receives the threshold value a and the threshold value b (S82: YES), the control unit 11 sets the threshold value a and the threshold value b, displays the threshold values a and b of levels 1 to 4 in FIG. 17 (S83), and ends the process. To do.

According to the present embodiment, when the detection result by the abnormality determination unit 114 is estimated to be uncertain based on the set threshold value, the presence or absence of the abnormality can be satisfactorily determined using the learning model.

As described above, the inspection device according to one aspect of the present invention includes an acquisition unit that acquires an image obtained by irradiating an inspection object placed on a conveyor with electromagnetic waves, and a determination value calculated based on the image. When the determination unit that determines the presence or absence of abnormality of the inspection object based on the threshold value and the image obtained by irradiating the inspection object with electromagnetic waves according to the determination result of the determination unit are input, the abnormality of the inspection object The learning inspection unit that acquires the information on the presence or absence of the abnormality by inputting the image obtained by irradiating the inspection object with electromagnetic waves into the learning model that outputs the information on the presence or absence of the abnormality, and outputs the information on the presence or absence of the abnormality. It is provided with an output unit.

According to the above configuration, with respect to the image obtained by irradiating the inspection object with electromagnetic waves, the image is input to the learning model according to the presence or absence of the abnormality of the inspection object judged based on the judgment value and the threshold value, and the abnormality is found. Get the presence or absence. For example, when it is uncertain whether or not there is an abnormality in the inspection object determined based on the threshold value, by inputting an image into the learning model and acquiring the presence or absence of the abnormality, the presence or absence of foreign matter, the presence or absence of the shape abnormality, etc. can be accurately determined. The presence or absence of abnormality can be detected.

In the above-mentioned inspection device, the determination unit may apply an image processing algorithm to the image to calculate a determination value and determine the presence or absence of an abnormality in the inspection object.

According to the above configuration, when the detection result is estimated to be uncertain based on the judgment value and the threshold value calculated by applying a predetermined image processing algorithm, the presence or absence of an abnormality can be detected with high accuracy.

In the above-mentioned inspection device, an image processing algorithm is applied to an image obtained by irradiating the inspection object with an electromagnetic wave to provide a specific unit for specifying a region for determining the presence or absence of an abnormality in the inspection object. The unit determines the presence or absence of an abnormality in the inspection object based on the determination value and the threshold value calculated based on the image of the region to which the image processing algorithm is applied, and the learning inspection unit determines the determination result of the determination unit. When the image of the region to which the image processing algorithm is applied or the image of the region obtained by irradiating the inspection object with an electromagnetic wave is input, information regarding the presence or absence of an abnormality in the region is output. Even if the image of the region to which the image processing algorithm is applied or the image of the region obtained by irradiating the inspection object with an electromagnetic wave is input to the training model, information regarding the presence or absence of the abnormality is acquired. Good.

According to the above configuration, the region for determining the presence or absence of an abnormality is specified, the presence or absence of an abnormality is determined for the specified region, and the learning model is applied according to the determination result, so that the detection of the presence or absence of an abnormality can be made more efficient.

The above-mentioned inspection device may include a first reception unit that accepts selection of an image processing algorithm.

According to the above configuration, the image processing algorithm can be selected according to the content of the abnormality, and the information regarding the presence or absence of the abnormality of the inspection object can be acquired with high accuracy.

The above-mentioned inspection device may include a setting unit for setting the threshold value.

According to the above configuration, it is determined whether or not to apply the learning model based on the threshold value. Therefore, when the determination result by the determination unit is estimated to be uncertain, the learning model is used to satisfactorily determine the presence or absence of abnormality. You can judge.

In the above-mentioned inspection device, when the determination value is equal to or more than the first threshold value and equal to or less than the second threshold value, the learning inspection unit may input the image into the learning model.

According to the above configuration, for example, when it is presumed that the judgment value is definitely normal when the judgment value is less than the first threshold value and abnormal when the judgment value exceeds the second threshold value, the judgment value is the first. When it is 1 threshold value or more and 2nd threshold value or less, the learning model can be applied to satisfactorily acquire the presence / absence information of abnormality.

In the above-mentioned inspection device, when the determination value is less than the first threshold value, the determination unit determines that the region is normal, and when the determination value exceeds the second threshold value, the determination unit determines. The region may be determined to be abnormal, and the learning inspection unit may not input an image into the learning model.

According to the above configuration, the presence / absence information of abnormality can be acquired satisfactorily and efficiently by applying the learning model, except when the inspection object is surely normal or abnormal.

In the above-mentioned inspection device, the output unit may include a second reception unit that displays the image and accepts the quality of the output result of the output unit with respect to the displayed image.

According to the above configuration, the worker can acquire the quality of the judgment result of the learning model and relearn the image.

The above-mentioned inspection device may include a re-learning unit that relearns the learning model based on the image and information on the quality.

According to the above configuration, it is possible to relearn the quality of the judgment of the output result received by the worker as teacher data and generate a learning model with better accuracy.

In the above-mentioned inspection device, the re-learning unit may re-learn when the second reception unit receives a rejection.

According to the above configuration, when the second reception unit accepts a rejection, the image is relearned from the teacher data in which the judgment of the output result is good or bad, so that the accuracy can be improved with a smaller amount of teacher data. A learning model that is good can be generated.

In the abnormality detection method according to one aspect of the present invention, an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is acquired, and the inspection is performed based on a judgment value and a threshold value calculated based on the image. Learning to determine the presence or absence of an abnormality in an object and output information regarding the presence or absence of an abnormality in the inspection object when an image obtained by irradiating the inspection object with electromagnetic waves is input according to the determination result of the presence or absence of the abnormality. A computer is made to execute a process of inputting the acquired image into the model and outputting information indicating the presence or absence of abnormality of the inspection object.

According to the above configuration, information on the presence or absence of abnormality is acquired using the learning model according to the judgment result. Therefore, for example, when the judgment result is estimated to be uncertain, an image is input to the learning model to obtain the presence or absence of abnormality. By acquiring, it is possible to accurately detect the presence or absence of foreign matter, the presence or absence of abnormalities such as shape abnormalities.

The computer program according to one aspect of the present invention acquires an image obtained by irradiating an inspection object placed on a conveyor with electromagnetic waves, and the inspection object is based on a judgment value and a threshold value calculated based on the image. A learning model that determines the presence or absence of abnormalities in the above, and outputs information on the presence or absence of abnormalities in the inspection object when an image obtained by irradiating the inspection object with electromagnetic waves is input according to the determination result of the presence or absence of the abnormality. The computer is made to execute a process of inputting the acquired image and outputting information indicating the presence or absence of abnormality of the inspection object.

According to the above configuration, information on the presence or absence of abnormality is acquired using the learning model according to the judgment result. Therefore, for example, when the judgment result is estimated to be uncertain, an image is input to the learning model to obtain the presence or absence of abnormality. By acquiring, it is possible to accurately detect the presence or absence of foreign matter, the presence or absence of abnormalities such as shape abnormalities.

In the method of generating a learning model according to one aspect of the present invention, an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is acquired, and the image and information indicating the presence or absence of abnormality of the inspection object are obtained. When the teacher data recorded in association with and is acquired and the image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves is input based on the teacher data, the abnormality of the inspection object is abnormal. Generate a learning model that outputs information about the presence or absence.

According to the above configuration, the presence or absence of the abnormality can be accurately acquired when an image acquired by applying a predetermined image processing algorithm corresponding to the content of the abnormality of the inspection object is input.

The learning model according to one aspect of the present invention has an input layer into which an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is input, and an output for outputting information regarding the presence or absence of abnormality in the inspection object. An intermediate in which parameters are learned using teacher data recorded by associating an image obtained by irradiating a layer and an inspection object placed on a conveyor with an electromagnetic wave and information indicating the presence or absence of an abnormality in the inspection object. When an image obtained by irradiating an inspection object with an electromagnetic wave is input to the input layer, information regarding the presence or absence of abnormality of the inspection object is output from the output layer through calculations by the intermediate layer. Make your computer work as you do.

According to the above configuration, the presence or absence of the abnormality can be accurately acquired when an image acquired by applying a predetermined image processing algorithm corresponding to the content of the abnormality of the inspection object is input.

The embodiments disclosed this time should be considered to be exemplary in all respects and not restrictive. The scope of the present invention is indicated by the scope of claims, not the above-mentioned meaning, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims.
For example, the inspection product is not limited to food, and may be a package, an industrial product, or the like.
The electromagnetic wave irradiating the inspection object 4 is not limited to X-rays, and may be terahertz waves, infrared rays, visible light, or the like.
Further, the target specifying unit 113 may extract the blob 41 in which a ≦ determination value ≦ b.

1 PC (inspection device)
10 Anomaly detection system 11 Control unit 12 Main storage unit 13 Communication unit 14 Display unit 15 Auxiliary storage unit P program 151 Inspection image DB
152-155 Abnormality detection model 156 Recording medium 2 X-ray inspection machine 22 Display 3 X-ray irradiation unit 4 Inspection object 8 Inspection image 81, 82 Bounding box 5 Display screen 6 Conveyor unit 61 Conveyance belt 7 X-ray detection unit 71 Detection line

The inspection device according to one aspect of the present invention is based on an acquisition unit that acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and the brightness of each pixel in the image. A judgment value is calculated based on a specific part that specifies a blob that determines the presence or absence of an abnormality in the inspection object and a blob that is specified in the specific part, and the calculated judgment value is the first threshold value on the normal side and the second threshold value on the abnormal side. and whether a determination unit whether the intermediate value between the threshold value, if the determination section determines that the is an intermediate value, when input images including the identified blobs, the inspected anomaly A learning inspection unit that inputs an image containing a specified blob into a learning model that outputs information regarding the presence or absence of the abnormality and acquires information regarding the presence or absence of the abnormality, and an output unit that outputs information regarding the presence or absence of the abnormality. Be prepared.
Here, the "image obtained by irradiation with electromagnetic waves" includes an image based on electromagnetic waves and an image obtained by image processing the image.

The abnormality detection method according to one aspect of the present invention acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and the inspection object is based on the brightness of each pixel in the image. The blob for determining the presence or absence of the abnormality is specified, and it is determined whether or not the judgment value calculated based on the identified blob is an intermediate value between the first threshold value on the normal side and the second threshold value on the abnormal side . If it is determined that the an intermediate value, when entering an image including a blob identified, the learning model for outputting information about the presence or absence of abnormality of the inspected object, receives an image that includes the identified blobs Then, the computer is made to execute the process of outputting the information regarding the presence or absence of the abnormality.

The computer program according to one aspect of the present invention acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and based on the brightness of each pixel of the image, the inspection object. identify determining blob the presence or absence of an abnormality, the determination value calculated based on the identified blob, it is determined whether the intermediate value between the second threshold value of the first threshold and the abnormal side of the normal side, the If it is determined that the intermediate value, when entering an image including a blob identified, the learning model for outputting information about the presence or absence of abnormality of the inspected object, and enter the image containing the identified blobs , The computer is made to execute the process of outputting the information regarding the presence or absence of the abnormality.

The method of generating a learning model according to one aspect of the present invention is an abnormality of the inspection object identified based on the brightness of each pixel in the image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves. The teacher data recorded by associating the image including the blob for determining the presence or absence of the test object with the information indicating the presence or absence of the abnormality of the inspection object is acquired, and the judgment value calculated based on the specified blob based on the teacher data is obtained. If it is determined that the intermediate the second threshold value of the first threshold and the abnormal side of the normal side, when the image including the blobs identified is input, outputs the information about the presence or absence of abnormality of the inspected Generate a learning model.

In the learning model according to one aspect of the present invention, among the images obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves, the presence or absence of abnormality of the inspection object specified based on the brightness of each pixel is determined. When it is determined that the judgment value calculated based on the blob to be determined is between the first threshold value on the normal side and the second threshold value on the abnormal side, the input layer into which the image including the specified blob is input. , Parameters are learned using the output layer that outputs information regarding the presence or absence of abnormality in the inspection object, the image containing the specified blob, and the teacher data recorded in association with the information indicating the presence or absence of abnormality in the inspection object. When an image containing the specified blob is input to the input layer, the output layer outputs information regarding the presence or absence of abnormality of the inspection object through a calculation by the intermediate layer. Make your computer work.

As described above, the inspection device according to one aspect of the present invention includes an acquisition unit that acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and the brightness of each pixel in the image. Based on the specific unit that identifies the blob that determines the presence or absence of abnormality in the inspection object, and the determination value calculated based on the blob specified by the specific unit, the calculated determination value is the first threshold value on the normal side. an intermediate of whether the determination unit is a value of the second threshold value of the abnormal side, if the determination section determines that the is an intermediate value, when input images including the identified blobs , An image containing the specified blob is input to the learning model that outputs information on the presence or absence of abnormality of the inspection object, and the learning inspection unit that acquires the information on the presence or absence of the abnormality and outputs the information on the presence or absence of the abnormality. It is provided with an output unit.

According to the above configuration, with respect to the image obtained by irradiating the inspection object with electromagnetic waves, the image is input to the learning model according to the presence or absence of the abnormality of the inspection object judged based on the judgment value and the threshold value, and the abnormality is found. Get the presence or absence. For example, when it is uncertain whether or not there is an abnormality in the inspection object determined based on the threshold value, by inputting an image into the learning model and acquiring the presence or absence of the abnormality, the presence or absence of foreign matter, the presence or absence of the shape abnormality, etc. The presence or absence of abnormality can be detected.
According to the above configuration, when the detection result is estimated to be uncertain based on the judgment value and the threshold value calculated by applying a predetermined image processing algorithm, the presence or absence of an abnormality can be detected with high accuracy.
According to the above configuration, the blobs for determining the presence / absence of abnormality are specified, the presence / absence of abnormality is determined for the specified blob, and the learning model is applied according to the determination result, so that the detection of the presence / absence of abnormality can be made more efficient.

In the above inspection device, the specifying unit applies the images processing algorithms, identifying the blob.

The above-mentioned inspection device may include a setting unit for setting the first threshold value and the second threshold value .

The abnormality detection method according to one aspect of the present invention acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and the inspection object is based on the brightness of each pixel in the image. The blob for determining the presence or absence of the abnormality is specified, and it is determined whether or not the judgment value calculated based on the identified blob is an intermediate value between the first threshold value on the normal side and the second threshold value on the abnormal side . If it is determined that the an intermediate value, when entering an image including a blob identified, the learning model for outputting information about the presence or absence of abnormality of the inspected object, receives an image that includes the identified blobs Te, before outputting the information as to whether or not Kikoto normal to execute the process to the computer.

The computer program according to one aspect of the present invention acquires an image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave, and based on the brightness of each pixel of the image, the inspection object. identify determining blob the presence or absence of an abnormality, the determination value calculated based on the identified blob, it is determined whether the intermediate value between the second threshold value of the first threshold and the abnormal side of the normal side, the If it is determined that the intermediate value, when entering an image including a blob identified, the learning model for outputting information about the presence or absence of abnormality of the inspected object, and enter the image containing the identified blobs , before outputting the information as to whether or not Kikoto normal to execute the process to the computer.

The method of generating a learning model according to one aspect of the present invention is an abnormality of the inspection object identified based on the brightness of each pixel in the image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves. The teacher data recorded by associating the image including the blob for determining the presence or absence of the test object with the information indicating the presence or absence of the abnormality of the inspection object is acquired, and the judgment value calculated based on the specified blob based on the teacher data is obtained. If it is determined that the intermediate the second threshold value of the first threshold and the abnormal side of the normal side, when the image including the blobs identified is input, outputs the information about the presence or absence of abnormality of the inspected Generate a learning model.

In the learning model according to one aspect of the present invention, among the images obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves, the presence or absence of abnormality of the inspection object specified based on the brightness of each pixel is determined. When it is determined that the judgment value calculated based on the blob to be determined is between the first threshold value on the normal side and the second threshold value on the abnormal side, the input layer into which the image including the specified blob is input. , Parameters are learned using the output layer that outputs information regarding the presence or absence of abnormality in the inspection object, the image containing the specified blob, and the teacher data recorded in association with the information indicating the presence or absence of abnormality in the inspection object. When an image containing the specified blob is input to the input layer, the output layer outputs information regarding the presence or absence of abnormality of the inspection object through a calculation by the intermediate layer. Make your computer work.

Claims (14)

An acquisition unit that acquires an image obtained by irradiating an inspection object placed on a conveyor with electromagnetic waves, and an acquisition unit.
A determination unit that determines the presence or absence of an abnormality in the inspection object based on the determination value calculated based on the image and the threshold value, and
By irradiating the inspection object with electromagnetic waves, the learning model that outputs information on the presence or absence of abnormalities in the inspection object when an image obtained by irradiating the inspection object with electromagnetic waves is input according to the judgment result of the determination unit. A learning inspection unit that inputs the obtained image and acquires information on the presence or absence of the abnormality.
An inspection device including an output unit that outputs information regarding the presence or absence of the abnormality. The inspection device according to claim 1, wherein the determination unit calculates a determination value by applying an image processing algorithm to the image, and determines the presence or absence of an abnormality in the inspection object. An image processing algorithm is applied to an image obtained by irradiating an inspection object with electromagnetic waves to provide a specific portion for specifying a region for determining the presence or absence of an abnormality in the inspection object.
The determination unit
Based on the judgment value and the threshold value calculated based on the image of the area to which the image processing algorithm is applied, the presence or absence of abnormality of the inspection object is determined.
The learning inspection department
When the image of the region to which the image processing algorithm is applied or the image of the region obtained by irradiating the inspection object with an electromagnetic wave is input according to the determination result of the determination unit, the abnormality of the region is found. An image of the region to which the image processing algorithm is applied or an image of the region obtained by irradiating the inspection object with an electromagnetic wave is input to a learning model that outputs information regarding the presence or absence of the abnormality. The inspection device according to claim 1 or 2 for acquiring information. The inspection device according to claim 3, further comprising a first reception unit that accepts selection of an image processing algorithm. The inspection device according to any one of claims 1 to 4, further comprising a setting unit for setting the threshold value. The inspection device according to any one of claims 1 to 5, wherein when the determination value is equal to or greater than a first threshold value and is equal to or less than a second threshold value, the learning inspection unit inputs the image into the learning model. When the determination value is less than the first threshold value, the determination unit determines that the region is normal, and determines that the region is normal.
When the determination value exceeds the second threshold value, the determination unit determines that the region is abnormal, and determines that the region is abnormal.
The inspection device according to any one of claims 3 to 5, wherein the learning inspection unit does not input an image into the learning model. The output unit displays the image and
The inspection device according to any one of claims 1 to 7, further comprising a second reception unit that receives the quality of the output result of the output unit with respect to the displayed image. The inspection device according to claim 8, further comprising a re-learning unit that relearns the learning model based on the image and information on the quality. The inspection device according to claim 9, wherein the re-learning unit relearns when the second reception unit receives a rejection. The image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves is acquired.
The presence or absence of abnormality in the inspection object is determined based on the determination value calculated based on the image and the threshold value.
When an image obtained by irradiating the inspection object with electromagnetic waves is input according to the determination result of the presence or absence of the abnormality, the acquired image is input to the learning model that outputs information on the presence or absence of the abnormality of the inspection object. And outputs the information indicating the presence or absence of abnormality of the inspection object.
Anomaly detection method that causes a computer to perform processing. The image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves is acquired.
The presence or absence of abnormality in the inspection object is determined based on the determination value calculated based on the image and the threshold value.
When an image obtained by irradiating the inspection object with electromagnetic waves is input according to the determination result of the presence or absence of the abnormality, the acquired image is input to the learning model that outputs information on the presence or absence of the abnormality of the inspection object. And outputs the information indicating the presence or absence of abnormality of the inspection object.
A computer program that lets a computer perform processing. An image obtained by irradiating an inspection object placed on a conveyor with an electromagnetic wave is acquired, and teacher data recorded by associating the image with information indicating the presence or absence of an abnormality in the inspection object is acquired.
Learning to generate a learning model that outputs information on the presence or absence of abnormality of the inspection object when an image obtained by irradiating the inspection object placed on the conveyor with electromagnetic waves is input based on the teacher data. How to generate a model. An input layer into which an image obtained by irradiating an inspection object placed on a conveyor with electromagnetic waves is input, and
An output layer that outputs information regarding the presence or absence of abnormalities in the inspection object, and
An image obtained by irradiating an inspection object placed on a conveyor with electromagnetic waves and an intermediate layer whose parameters have been learned using teacher data recorded by associating information indicating the presence or absence of abnormality of the inspection object. Prepare,
When an image obtained by irradiating an inspection object with electromagnetic waves is input to the input layer, the computer is configured to output information regarding the presence or absence of abnormality of the inspection object from the output layer through calculations by the intermediate layer. A learning model that works.