Method for identifying contents of can

Abstract

To provide a method for identifying contents by using X-ray CT scanning in an unopened state.SOLUTION: A method for identifying contents, comprises: performing X-ray CT scanning on two or more known types of objects that are covered with metal and whose inside cannot be seen; constructing a machine learning model with machine learning software using a three-dimensional image S generated on the basis of generated internal visualization data; and for an unknown object, generating a three-dimensional image which is the source of the learning model in the same manner as above, and identifying the unknown object using the constructed machine learning model. This can provide a method for identifying whether the contents of objects belong to two known types by using X-ray CT scanning without opening the two or more known types of objects that are covered with metal and whose inside cannot be seen.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for determining the contents of an object that is covered with metal and whose interior cannot be seen, using a machine-learned learning model. [Background technology]

[0002] Conventionally, X-ray inspection devices and inspection methods have been provided for production lines of food products and the like, as described in Patent Document 1. Furthermore, Patent Document 2 describes a technology developed for efficiently detecting abnormalities in medical diagnoses by using machine learning on images taken with an X-ray CT (Computed Tomography) device. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2008-175691 [Patent Document 2] Japanese Patent Publication No. 2020-192006 Summary of the Invention [Problem to be solved by the invention]

[0004] Percussion testing is a known non-destructive method for detecting defective metal-encased canned fish. However, percussion testing detects abnormalities in the internal pressure of the can due to leakage or deterioration by detecting the pitch of the sound emitted, and does not provide a method for detecting defects that are merely cosmetically unsightly, such as broken fish flesh or peeling skin. Visual inspections, such as for broken fish flesh or peeling skin, require sample testing by opening the lid, but the inspected products are discarded, and because this is a random inspection, it is not possible to inspect the condition of all cans. If it were possible to detect unintended foreign matter, such as other fish species, in addition to visual inspections such as broken fish flesh or peeling skin, it would be possible to eliminate defective products due to contamination within the factory. Furthermore, examples of canned food identification, in addition to identifying defective products, include identifying the type of fish in the can and the can's manufacturer. In addition, X-ray CT scanning, which is used to inspect canned fish for foreign matter contamination, can determine whether or not foreign matter is present, but it cannot handle the disintegration of fish flesh, peeling of skin, or the identification of fish species. Furthermore, conventional non-destructive inspection methods such as percussion testing and X-ray CT scanning cannot identify cans whose manufacturer is unknown due to peeling or damage to the painted surface or label. Therefore, the present invention aims to provide a method for identifying the contents of metal-covered canned fish even when the can is unopened. [Means for solving the problem]

[0005] In order to solve the above problems, the present invention includes the following aspects. [1] A method for identifying the contents of an object that is covered with metal and whose interior cannot be seen, comprising the steps of acquiring cross-sectional image data of the object using an X-ray CT scan, generating a three-dimensional image from the cross-sectional image data, and constructing a machine learning model trained on the three-dimensional image data. [2] In the method described in [1], the object covered with metal and the inside of which cannot be seen is canned fish. [3] In any of the methods described in [1] to [2], the object covered with metal so that the inside cannot be seen is canned fish meat made from mackerel. [Effects of the Invention]

[0006] A method is provided for determining the contents of an object that is covered with metal and whose interior cannot be seen, using X-ray CT scanning without opening the package. [Brief explanation of the drawings]

[0007] [Figure 1] 1 is a flowchart of a determination method performed in the present invention. [Figure 2] Photographs of the canned foods KA and KB used in the examples taken from above. [Figure 3] 10 is a three-dimensional image of canned food KA and KB in an embodiment. [Figure 4] FIG. 10 is a diagram showing the discrimination results of canned food KX in an example. [Figure 5] FIG. 10 is a diagram showing the average output accuracy of the canned KX in the embodiment. [Figure 6] FIG. 1 is a diagram showing the discrimination results of canned KY in an example. [Figure 7] FIG. 10 is a diagram showing the average output accuracy of the canned KY in the embodiment. [Figure 8] FIG. 10 is a diagram showing the discrimination results of canned KC in an example. [Figure 9] FIG. 10 is a diagram showing the average output accuracy of canned KC in the embodiment. [Figure 10] FIG. 10 is a diagram showing the transition of the average output accuracy of canned food KA, canned food KB, and canned food KC in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0008] The present invention is a method for constructing a learning model using three-dimensional images of two or more known objects whose interiors are covered with metal and whose interiors are not visible, and for discriminating with a probability of 95% or higher whether an unknown object is of the same type as the three-dimensional image that served as the basis for the learning model. An object covered with metal and whose interior is not visible has the property of being able to be penetrated by X-rays of a certain intensity, and some of the penetrated and some of the un-penetrated light can be captured as images. For example, this applies to objects such as canned fish meat, which is covered with aluminum or iron and whose interior is not visible. Specifically, this refers to an object that is penetrated by X-rays of a certain intensity, but the bone in the center of the fish meat is not penetrated and can be captured as an image. This invention uses three-dimensional images of known objects to construct a learning model, and applies this model to unknown objects, making it possible to discriminate between the objects with a probability of 95% or higher. This allows for the non-destructive and high-probability identification of the type of object covered with metal and whose interior is not visible. [Example]

[0009] FIG. 1 shows an example of the overall configuration of a method 10 for identifying the contents of canned food according to the present invention. In step S10, an X-ray CT scan was performed on can K using the following equipment and settings to create a cross-sectional image D. Setting can K: Can K was placed in an appropriate position. X-ray CT scan equipment and settings: Equipment for performing X-ray CT scans was used. An appropriate X-ray source and detector were arranged, and the target can K was set in an appropriate position. Performing X-ray CT scan: An X-ray CT scan was performed on can K. X-rays passed through can K, capturing the internal structure, generating a cross-sectional image D. Obtaining cross-sectional image D: Cross-sectional image D generated by the X-ray CT scan was obtained. This cross-sectional image visualizes the internal structure of can K. Through this procedure, the contents of can K were visualized non-destructively by X-ray CT scanning, and cross-sectional image D was obtained. This image will be used in subsequent steps to build a machine learning model and identify the contents.

[0010] X-ray CT scans were performed using a TOSCANER-32300μHD (Toshiba IT Control Systems). All scans were performed in the L-mode of the device, with a tube voltage of 90 kV and a tube current of 90 μA. In addition, fine mode and cone beam settings were employed. Specifically, the X-ray tube was positioned close enough to scan the entire can, and measurements were taken at that position. This setting allowed the X-ray CT scan to adequately visualize the internal structure of the can. The cross-sectional images and data obtained were used for subsequent data analysis and the construction of machine learning models.

[0011] In step S20, a three-dimensional image S was created based on the cross-sectional image D created in step S10. In this step, 3D imaging was first performed using VR (Volume Rerendering) software. The software used was ExFact® VR version 2.2 (manufactured by Japan Visual Science Co., Ltd.). The cross-sectional image excluding the pull-top portion was imported for 3D imaging. Because it was difficult to visualize the fish meat, the contrast was adjusted to visualize the bones. Using the VR software's functions, the three-dimensional image was rotated 180 degrees from the surface containing the pull-top portion to a vertical view from the opposite side, and a screenshot was taken. The resulting image was then exported as a TIFF file. Next, the TIFF image file was cropped into a square using ImageJ (provided by the National Institutes of Health, a U.S. public institution), converted to a PNG image file, and saved. This process processed and saved the three-dimensional image, providing data for subsequent analysis and visualization.

[0012] This invention proposes a method for nondestructively identifying the contents of canned goods. The following describes the specific steps S10 and S20 performed on multiple cans. First, the following procedures were performed on 45 cans of KA (Japanese boiled mackerel, manufactured by Hoko Co., Ltd.), 45 cans of KB (Aiko-chan® boiled mackerel, manufactured by Ito Foods Co., Ltd.), 12 cans of KX (Japanese boiled mackerel, manufactured by Hoko Co., Ltd., a different can from KA), 12 cans of KY (Aiko-chan boiled mackerel, manufactured by Ito Foods Co., Ltd., a different can from KB), and 12 cans of KC (various manufacturers and products, totaling 12 cans). Photography (Fig. 2): The lids of the cans KA and KB were opened and photographed. Three-dimensional image creation (Fig. 3): Steps S10 and S20 were performed using the cans KA and KB, and three-dimensional images SA and SB were created. Through these procedures, the inventors observed over 100 cans over a period of approximately two years, but concluded that there were no clear differences in the three-dimensional images of each manufacturer, making it difficult to determine the contents.

[0013] In step S30, using the freely available machine learning software Teachable Machine (provided by Google), a learning model S40 was constructed using the 3D images SA and SB, which were PNG files saved in step S20. To construct this learning model, one can or five cans were selected, increasing the number of cans by five until a maximum of 45 cans was reached. Teachable Machine is software that performs machine learning and image classification using a neural network algorithm. While Teachable Machine was used in this example, other machine learning software, such as YOLO (provided by Ultralytics), can also be used. To construct learning model S40, three combinations were selected from randomly selected combinations ranging from one can to 40 cans.

[0014] The PNG-format 3D image data S stored in the learning model S40 was loaded, and the contents were identified based on the can for which each image file showed the highest output accuracy. Specifically, the output accuracy (SRKD) is defined as the numerical value displayed by the machine learning software as a value indicating the probability that the can is one of the cans when building the learning model S40. A can is identified if this output accuracy (SRKD) is 95% or higher. This is because, statistically, an error probability of 5% or less is considered an error. The displayed output accuracy (SRKD) was recorded, and the average accuracy (Score) was calculated and compared for 12 cans of KX, 12 cans of KY, and 12 cans of KC. When building the learning model S40, the number of 3D image files S used for training was increased in increments of 5, and the level at which the average accuracy (Score) exceeded 95% was investigated.

[0015] Figure 4 shows the results of constructing a learning model S40 by selecting five cans of KA and five cans of KB, each containing one can of KX. An output accuracy SRKD of 100% represents the probability that the can is a can of KA. In this experiment, the cans of KX were from the same manufacturer as the cans of KA, so ideally, this value would be 100%. Figure 5 plots the average accuracy Score. In this example, X-ray CT scans were performed on multiple cans of KA and multiple cans of KB in step S10 to obtain cross-sectional data D. When constructing the learning model S40 using the three-dimensional images obtained in step S20, the number of cans of KA and KB used for machine learning was varied from one to five, up to 40 increments of five, in three different combinations. The average accuracy Score of the output accuracy SRKD when classifying 12 cans of KX was plotted. The same 12 cans were repeatedly used for classification when acquiring each data set, and we investigated how the accuracy varied depending on the combination of data and the number of cans used for learning.

[0016] One of the S40 learning models using data from only one can had an average accuracy score of over 60%, while the remaining S40 learning models averaged approximately 40%. For the S40 learning models constructed using five cans, two had accuracy scores exceeding 95%, while the remaining S40 models achieved relatively good accuracy scores of approximately 90%. When the number of cans used to construct the learning models was increased from 10 to 30, these accuracy scores remained stable at approximately 90%, indicating little fluctuation with increasing the number of cans used for machine learning. When 35 or more cans were used to construct the S40 learning model, accuracy consistently remained above 95%. When the accuracy score was above 95%, the error was below 5%, indicating highly accurate estimation. From this, it was concluded that when constructing the learning model S40, even if only five pieces of three-dimensional image data S are selected, highly accurate estimations can be made depending on the selected data, but that approximately 35 or more pieces of three-dimensional image data S are required to ensure high accuracy.

[0017] Similarly, Figure 6 shows the results of constructing a learning model S40 by selecting one can of KY and five cans of KA and KB. In this case, an output accuracy SRKD of 100% represents the probability that the can is a can of KB. In this experiment, the cans used for the KY were from the same manufacturer as the cans used for the KB, so ideally, this value would be 100%. Figure 7 graphs the average accuracy Score of the output accuracy SRKD. In this example, X-ray CT scans were performed on the multiple cans of KA and multiple cans of KB used in the example to generate cross-sectional data D that visualized their interiors in step S10. Next, in step S20, when constructing a learning model S40 using machine learning with three-dimensional images, the number of cans of KA and cans of KB used for machine learning was either one can each, or five cans, or 40 cans each, with five cans incrementing by five, in three different combinations. This graph shows the average accuracy Score when discriminating 12 cans of KY. When obtaining each data set, the same 12 cans were repeatedly used for discrimination, and we investigated how the results would change depending on the combination of data used for learning and the number of learnings.

[0018] The experimental results showed that one of the training models S40, which used only data from a single can's 3D image S, had an average accuracy of over 65%, while the other training models S40 averaged approximately 25% to 40%. Two of the training models S40 constructed using 15 cans achieved accuracy scores exceeding 95%, while the other training models S40 achieved a relatively good accuracy score of approximately 90%. When the number of cans used to construct the training model S40 was increased from 20 to 25, these accuracy scores remained stable at approximately 90%, indicating little fluctuation due to an increase in the number of cans trained. When 30 or more cans were used to construct the training model S40, the accuracy score consistently remained above 95%. Even in this case, when the accuracy score was above 95%, the error was below 5%, demonstrating highly accurate estimation. This suggests that even if only 15 pieces of 3D image data are selected when constructing the learning model S40, estimations can be made with extremely high accuracy depending on the combination of selected data, but that in order to construct a learning model S40 with an accuracy score of 95% or more, approximately 30 or more pieces of 3D image data S are required.

[0019] Figure 8 shows the results of constructing learning model S40 by selecting one canned KC and five canned KA and canned KB. The output accuracy SRKD of 52% and 48% represents the probability that the can is a canned KA or a canned KB. In the experiment, canned KC from a manufacturer other than canned KA or canned KB was used. Figure 9 plots the average accuracy Score (calculated based on the probability that it was manufactured by Company B) of the output accuracy SRKD. In this experiment, X-ray CT scans were performed on multiple canned KA and multiple canned KB in step S10, and cross-sectional data D was generated to visualize the interior. Then, in step S20, when constructing learning model S40 using 3D image S, the number of canned KA and canned KB used for machine learning was either one can or five or five cans each, and the graph shows the average accuracy Score when the model was tasked with discriminating 12 canned KC. When obtaining this data, the same 12 cans were used repeatedly to discriminate between different types of food, and we investigated how this would change depending on the combination of data used for learning and the number of learnings.

[0020] Here, the accuracy score indicates the probability that the canned food is KB. In the experiment, canned food KC, which is from a different manufacturer than canned food KA and canned food KB, was used. Therefore, ideally, the accuracy score should hover around 50%. In reality, the accuracy score increased slightly, but no significant difference was observed even when the number of cans used to build the learning model increased. The average accuracy score ranged from approximately 40% to 60%.

[0021] Here, the average accuracy scores for the three groups of canned foods KX, canned foods KY, and canned foods KC are tabulated and shown in a graph. Specifically, the average accuracy scores shown in Figure 10 are plotted on the vertical axis, with the average accuracy scores for machine learning performed on three combinations: canned foods KX = canned foods KA, canned foods KY = canned foods KB, and canned foods KC = canned foods KB, and the horizontal axis is the number of canned foods KA and canned foods KB that were machine-learned. This allowed us to investigate how the number of canned foods that were machine-learned affects the accuracy score.

[0022] In this example, it was successfully demonstrated that the learning model S40 can distinguish between cans from different manufacturers. It was also shown that increasing the number of cans used in building the learning model S40 increases the accuracy of identifying the cans as being from that manufacturer. The results of this study revealed that building the learning model S40 using approximately 35 pieces of 3D image data S enables highly accurate discrimination with an average accuracy score of 95% or higher. It was also revealed that highly accurate discrimination is possible even with only five pieces of 3D image data S when sorting data.

[0023] As described above, in the Examples, it was possible to distinguish between differences between manufacturers. When machine learning was performed on regular canned fish meat without falling apart or peeling of the flesh and canned fish meat with characteristics different from regular products, including falling apart or peeling of the flesh, the method described in the Examples made it possible to distinguish between regular canned fish meat and canned fish meat with characteristics different from regular products, including falling apart or peeling of the flesh, in the same way as distinguishing between manufacturers.

[0024] Although the present invention has been described by the above single embodiment, the descriptions and drawings that form part of this disclosure should not be understood as limiting the present invention. From this disclosure, various alternative embodiments, examples, and operating techniques will become apparent to those skilled in the art.

[0025] As such, the present invention naturally includes various embodiments not described herein. Therefore, the technical scope of the present invention is defined only by the invention-specifying matters according to the claims, which can be interpreted appropriately from the above explanation. [Explanation of symbols]

[0026] 10 How to determine the contents of a can KA, KB, KAn, KBn, KXn, KYn...Canned food DAn, DBn, DXn, DYn...cross-sectional images SA, SB, SAn, SBn, SXn, SYn...3D images Score: Average accuracy SRKD Output Accuracy

Claims

1. A method for identifying the contents of an object that is covered with metal and whose interior cannot be seen, comprising the steps of acquiring cross-sectional image data of the object using an X-ray CT scan, generating a three-dimensional image from the cross-sectional image data, and constructing a machine learning model that has been trained using the three-dimensional image data.

2. 2. The method according to claim 1, wherein the object covered with metal and whose interior is not visible is a canned fish.

3. 3. The method according to claim 1, wherein the object covered with metal and the inside of which cannot be seen is canned fish meat made from mackerel.

Images