Learning-completed model recreation method and learning-completed model recreation system

Abstract

The invention provides a learning completion model recreation method and a learning completion model recreation system. In this method for reproducing a learned model, a first learned model is a common learned model corresponding to a plurality of types of objects to be examined, and this method for reproducing a learned model includes: reproducing an X-ray image of any one of the plurality of types of objects to be examined, with respect to an X-ray image of any one of the plurality of types of objects to be examined; a step for acquiring a first estimation result obtained by inputting the X-ray image into the learned model; acquiring a first corrected estimation result in which the first estimation result is corrected; a step in which the first post-correction estimation results relating to the plurality of types of inspection objects are associated with the respective X-ray images and stored; and a step for reproducing the first learned model using the X-ray image stored in association with the first post-correction estimation result.

Description

TECHNICAL FIELD

[0001] The present application relates to a learning-completed model re-creation method and a learning-completed model re-creation system, and particularly relates to a learning-completed model re-creation method and a learning-completed model re-creation system that re-create a learning-completed model used for analysis of an X-ray image. BACKGROUND

[0002] Conventionally, there is known a device that uses a learning-completed model to analyze an X-ray image. For example, a device that uses such a learning-completed model to analyze an X-ray image is disclosed in Japanese Patent Application Publication No. 2024-029975.

[0003] An X-ray imaging system that uses a learning-completed model to identify at least one of a region of an examination object and a region of an abnormal portion contained in the examination object is disclosed in Japanese Patent Application Publication No. 2024-029975. The X-ray imaging system disclosed in Japanese Patent Application Publication No. 2024-029975 includes a fluoroscopic device and an analysis device. In the configuration disclosed in Japanese Patent Application Publication No. 2024-029975, the analysis device analyzes an X-ray image generated by the fluoroscopic device imaging the examination object. Specifically, in the configuration disclosed in Japanese Patent Application Publication No. 2024-029975, the analysis device is configured to identify the region of the examination object by inputting the X-ray image to the learning-completed model.

[0004] Here, although not described in Japanese Patent Application Publication No. 2024-029975, with changes in the market, the tendency of the types of parts and the like installed in the examination object sometimes changes. When the tendency of the types of parts and the like installed in the examination object changes, the accuracy of the inference result of the examination object based on the learning-completed model decreases due to the difference between the X-ray image used as the teacher data at the time of generating the learning-completed model and the X-ray image at the time of actual examination. SUMMARY

[0005] The present application was achieved in order to solve the problem described above, and an object of the present application is to suppress a decrease in the accuracy of the inference result of the examination object based on the learning-completed model even when the tendency of the types of parts and the like installed in the examination object changes.

[0006] A learning completed model re-creation method for re-creating a learning completed model of a region of an examination target object obtained from an X-ray image of the examination target object, the learning completed model re-creation method including: a step of obtaining the X-ray image; a step of inputting the X-ray image into a first learning completed model for an X-ray image of any one of a plurality of examination target objects, thereby obtaining a first inference result of the region of the examination target object; a step of obtaining a first corrected inference result in which the first inference result is corrected based on a user's operation input; a step of storing the first corrected inference result for the plurality of examination target objects in correspondence with respective X-ray images; and a step of re-creating the first learning completed model using the X-ray images and the first corrected inference result stored in correspondence for the plurality of examination target objects. Further, the re-creation of the learning completed model is a concept including: newly creating the learning completed model using only new teacher data; newly creating the learning completed model using both new teacher data and teacher data already used in learning; and re-learning the already created learning completed model using new teacher data.

[0007] Since the first inference result includes some errors, appropriate correction by a user's operation is required, and by storing the first corrected inference result in advance, the first learning completed model can be re-created using the stored first corrected inference result. That is, the execution of the examination of the examination target object serves as the creation of teacher data required for re-creation of a new first learning completed model. Therefore, the user does not need to perform the work of creating teacher data required for re-creation of a new first learning completed model, and the first learning completed model can be re-created while reducing the burden on the user.

[0008] Further, the first learning completed model is a general learning completed model, and is a model created by inputting a plurality of examination target objects as teacher data. However, since the types of examination target objects for which the user performs examinations sometimes change (trends change) over time, the accuracy of the first inference result based on the general learning completed model (the first learning completed model) decreases. However, in the present application, the first corrected inference result for a plurality of examination target objects is stored, and the first learning completed model is re-created using a plurality of stored first corrected inference results. Since the first corrected inference results used in the re-creation are inference results for a plurality of examination target objects for which the process of outputting the first inference result (that is, the examination is performed) is actually performed, it is considered that the re-created general learning completed model is more suitable for a plurality of examination target objects for which the user also performs examinations in the future. Therefore, by using the re-created general learning completed model, the accuracy of the first inference result can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 This is a block diagram illustrating the overall structure of a learning-completed model re-creation system based on an embodiment of the present invention.

[0010] Figure 2 This is a diagram showing an example of an object being inspected.

[0011] Figure 3 It is a block diagram used to illustrate the functional structure of the control unit.

[0012] Figure 4 This is a diagram showing an example of an X-ray image.

[0013] Figure 5 This diagram illustrates the structure of generating a learned model and the structure of using the learned model to analyze X-ray images.

[0014] Figure 6 This is a graph used to illustrate the inference results output by the model after learning.

[0015] Figure 7 This is an enlarged version of a portion of the inference result.

[0016] Figure 8 This is a diagram used to illustrate the correction points for the inference results.

[0017] Figure 9 This is a diagram used to illustrate the corrected inference results.

[0018] Figure 10 This is a diagram illustrating the structure by which the control unit obtains the first corrected inference result from the X-ray image.

[0019] Figure 11 This is a diagram illustrating the structure by which the control unit obtains the second corrected inference result from the X-ray image.

[0020] Figure 12 This diagram illustrates the structure of the control unit recreating the first learned model.

[0021] Figure 13 This diagram illustrates the structure of the control unit recreating the second learned model.

[0022] Figure 14 This is an example of a screen used to confirm the inference result.

[0023] Figure 15 This is an example image showing the management screen after the learning process is complete.

[0024] Figure 16 This is a flowchart used to illustrate the process of recreating the first learned model.

[0025] Figure 17 This is a flowchart used to illustrate the process of recreating a second, fully learned model. Detailed Implementation

[0026] Hereinafter, embodiments embodying the present invention will be described with reference to the accompanying drawings.

[0027] (Overall structure of the X-ray imaging system)

[0028] Reference Figures 1-17 The learning-completed model re-creation system 100 based on an embodiment of the present invention will be described.

[0029] like Figure 1 As shown, the learning-complete model re-creation system 100 based on this embodiment is a learning-complete model re-creation system for re-creating a learning-complete model 31 obtained from an X-ray image 40 of an object to be inspected 90. The learning-complete model 31 is used, for example, in non-destructive inspection applications to analyze the X-ray image 40 obtained by imaging the interior of the object to be inspected 90 as an object.

[0030] like Figure 2 As shown, the object 90 under inspection is an electronic device including a substrate 91. Electronic components 92 are mounted on the substrate 91. The electronic components 92 are electrically connected to the substrate 91 via multiple solder balls 93 (bumps). The multiple solder balls 93 are arranged in a regular pattern. Specifically, the multiple solder balls 93 are arranged in a grid pattern on the substrate 91 in a regular manner. That is, the electronic components 92 are connected to the substrate 91 via a ball grid array (BGA). Furthermore, multiple solder balls 93 are arranged on one side of the substrate 91. The electronic components 92 include, for example, integrated circuits (ICs) and other electronic circuits. The X-ray image 40 (see reference) used in the learned model re-fabrication system 100 for non-destructive inspection of anomalies such as voids (holes) and bridging of the multiple solder balls 93 is shown. Figure 4 Analysis of ). In addition, on the substrate 91, in addition to electronic components 92, surface-mount electronic components such as resistors or capacitors 94 are also mounted.

[0031] like Figure 1As shown, the learned model re-creation system 100 includes an X-ray imaging device 1 and an image processing device 2. The X-ray imaging device 1 performs X-ray imaging on the object 90 to be inspected. The image processing device 2 generates an X-ray image 40. Furthermore, the image processing device 2 analyzes the generated X-ray image 40 using the learned model 31. Additionally, the image processing device 2 re-creates the learned model 31. Both the X-ray imaging device 1 and the image processing device 2 have communication modules and exchange information via a network or the like.

[0032] The X-ray imaging apparatus 1 includes an X-ray irradiation unit 10 and an X-ray detector 11. The X-ray irradiation unit 10 is configured to irradiate X-rays. In this embodiment, the X-ray irradiation unit 10 irradiates an object 90, which includes a plurality of solder balls 93, with X-rays. The X-ray irradiation unit 10 includes an X-ray tube that is powered by a power supply device (not shown) to irradiate X-rays.

[0033] The X-ray detector 11 is configured to detect X-rays irradiated from the X-ray irradiation unit 10. The X-ray detector 11 outputs an electrical signal corresponding to the detected X-rays. The X-ray detector 11 includes, for example, a flat panel detector (FPD) as an X-ray detector. The X-ray irradiation unit 10 and the X-ray detector 11 are disposed inside a frame (not shown) of the X-ray imaging apparatus 1.

[0034] like Figure 1 As shown, the image processing apparatus 2 includes a control unit 20 and a storage unit 21. The image processing apparatus 2 is, for example, a personal computer communicatively connected to the X-ray imaging apparatus 1. The control unit 20 controls the operation of various parts of the X-ray imaging apparatus 1. For example, the control unit 20 controls the X-ray irradiation performed by the X-ray irradiation unit 10 by controlling a power supply device (not shown). The control unit 20 includes a processor or circuit such as a central processing unit (CPU), read-only memory (ROM), and random access memory (RAM). Alternatively, the control unit 20 may also include a graphics processing unit (GPU) or a processor configured for image processing purposes such as a field-programmable gate array (FPGA).

[0035] The storage unit 21 is configured to contain various programs 30 and parameters executed by the storage control unit 20. The storage unit 21 may include, for example, non-volatile memory such as a hard disk drive (HDD) or a solid state drive (SSD).

[0036] Additionally, the storage unit 21 stores the first completed learning model 31a, the second completed learning model 31b, the first revised inference result 51a, the second revised inference result 51b, the date and time of creation of the completed learning model 60, the specified period 61, and the threshold 62. The first completed learning model 31a is an example of the "completed learning model" in the claims.

[0037] Additionally, the image processing apparatus 2 is connected to a display unit 22 and an operation unit 23. The display unit 22 includes, for example, a liquid crystal monitor. The display unit 22 displays images and text information under control by the control unit 20. The operation unit 23 receives input operations performed by the operator. The operation unit 23 includes, for example, a pointing device such as a keyboard and mouse. The operation unit 23 outputs operation signals to the control unit 20 based on the received input operations.

[0038] like Figure 3 As shown, the control unit 20 includes an image generation unit 20a, an inference result acquisition unit 20b, a corrected inference result acquisition unit 20c, a model re-creation unit 20d, a re-creation acquisition unit 20e, a correction amount acquisition unit 20f, and a display control unit 20g, which are executed by the control unit 20 and stored in the storage unit 21 (see reference). Figure 1 Program 30 in ) (refer to) Figure 1 The function blocks implemented are configured in software form. In other words, program 30 is configured to enable the computer (control unit 20) to execute the various controls performed by function blocks 20a to 20g. Furthermore, these function blocks 20a to 20g may also be equipped with dedicated processors (processing circuits) and contain mutually independent hardware.

[0039] Image generation unit 20a is based on X-ray detector 11 (reference) Figure 1 X-ray images generated from detected X-rays 40 (reference) Figure 4 ).

[0040] (X-ray image)

[0041] like Figure 4 As shown, X-ray image 40 is an image of a substrate 91. Specifically, X-ray image 40 shows a plurality of solder balls 93 arranged in a grid pattern in a regular manner on the substrate 91.

[0042] (Model completed)

[0043] Reference Figure 5 The analysis of the first learned model 31a and the second learned model 31b, as well as the generation and reconstruction of the first learned model 31a and the second learned model 31b, are explained. The common parts of the first learned model 31a and the second learned model 31b are explained as learned model 31. For example... Figure 5 As shown, the learned model 31 is used to analyze X-ray image 40. Specifically, the learned model 31 is used to identify solder balls 93 (see reference). Figure 4 The solder ball area 80a (refer to) Figure 6 In other words, the area of ​​the object 90 being inspected is the solder ball area 80a containing multiple solder balls 93. After learning the model 31, the image processing device 2 (see reference)... Figure 1 (or generated by a computer different from the image processing device 2 and pre-stored in the storage unit 21 (see reference)) Figure 1 )middle.

[0044] In addition, such as Figure 5 As shown, a learned model 31 is generated through machine learning using input teacher data 33 and output teacher data 34 as the dataset. Input teacher data 33 is generated based on teacher X-ray images (not shown). Furthermore, the re-creation of the learned model 31 is performed similarly to its generation, based on input teacher data 33 and output teacher data 34. Teacher X-ray images and X-ray images 40 for analysis (see reference) Figure 4 Similarly, by X-ray imaging device 1 (refer to...) Figure 1 The X-ray image 40, which is taken and generated by the image processing device 2, and the teacher's X-ray image used to generate the completed learning model 31 are images containing the inspection object 90 (solder ball 93) with a common structure.

[0045] Additionally, the output teacher data 34 includes labeling 82 on the area of ​​the solder ball 93 captured in the teacher's X-ray image (see reference). Figure 6 This is generated when the learned model 31 is recreated. The inference result 50 or the corrected inference result 51 output by the learned model 31 (refer to...) Figure 9 ) is used as output teacher data 34.

[0046] Model 31, after learning, was generated using deep learning machine learning. Deep learning includes, for example, machine learning based on U-Net, a Fully Convolutional Network (FCN). Model 31, after learning, is able to identify solder balls 93 (see reference 40) by performing pixel-by-pixel processing on each of the input X-ray images 40. Figure 4 The image is generated by learning the image transformation (image reconstruction) method between the region (solder ball region 80a) and the background region other than solder ball 93.

[0047] like Figure 5 As shown, after learning, model 31 is configured to output inference result 50 when X-ray image 40 is input. Inference result 50 is obtained by analyzing the region of solder ball 93 in X-ray image 40 (solder ball region 80a (reference)). Figure 6 The label image is labeled with label 82 and identified as solder ball area 80a and other areas.

[0048] (Inference result)

[0049] Figure 6 An example of inference result 50 is shown. The common portion of the first inference result 50a and the second inference result 50b is used as inference result 50 for explanation. Inference result 50 is an image of the solder ball region 80a labeled with tag 82. When inputting X-ray image 40 (refer to...) Figure 4 When model 31 (refer to) is completed, Figure 1 The model 31 extracts the region (solder ball region 80a) of the object 90 being inspected contained in the X-ray image 40. Then, after learning, the model 31 outputs an inference result 50 that labels the extracted solder ball region 80a with 82. That is, the inference result 50 is a superimposed image of the X-ray image 40 with 82 superimposed on it.

[0050] (Revision of the inference results)

[0051] Figure 7 This is an enlarged image of a portion of the inference result 50. After learning model 31 (refer to...) Figure 5 When the inference accuracy is low, areas without labels 82 in the solder ball area 80a (label non-overlapping area 80b) and areas with labels 82 outside the solder ball area 80a (label detachment area 80c) may be generated. When the learned model 31 is re-created using the inference result 50 that generated the label non-overlapping area 80b and the label detachment area 80c, the inference accuracy of the re-created learned model 31 is reduced. Therefore, in this embodiment, the corrected inference result acquisition unit 20c (refer to...) Figure 3 This constitutes a revised inference result of 50.

[0052] Specifically, the corrected inference results obtained are as follows: Part 20c Figure 8 As shown in inference result 50c, insufficient labels 82 are added to correction region 81 or unnecessary labels 82 are deleted. Correction region 81 includes regions 80b that do not overlap with the labels (see reference). Figure 7 The corresponding non-overlapping label correction region 81a and the label detachment region 80c (see reference) Figure 7 The corresponding detached label correction area 81b. The non-overlapping label correction area 81a is the area where label 82 needs to be added. Additionally, the detached label correction area 81b is the area where label 82 needs to be deleted. Furthermore, in Figure 8 In the original text, the non-overlapping label correction area 81a and the detached label correction area 81b were marked with shadows for convenience. However, in the actual inference result 50, the non-overlapping label correction area 81a and the detached label correction area 81b were not marked with shadows.

[0053] In this embodiment, the modified inference result acquisition unit 20c acquires the result based on the user's operation input. Figure 9 The corrected inference result 51 is shown. Specifically, when the corrected inference result acquisition unit 20c receives user input to add a tag 82 to the area corresponding to the solder ball area 80a, it adds a tag 82 to the non-overlapping tag correction area 81a. Furthermore, when the corrected inference result acquisition unit 20c receives user input to delete an unwanted tag 82 from the tag removal area 80c, it deletes the tag 82 that has been removed from the tag correction area 81b.

[0054] The corrected inference result acquisition unit 20c corrects the correction region 81 based on user operation input and obtains the result. Figure 9 The corrected inference result 51 is shown. The common portion of the first corrected inference result 51a and the second corrected inference result 51b is described as corrected inference result 51. Corrected inference result 51 is a superimposed image excluding the label non-overlapping area 80b and the label detachment area 80c relative to the solder ball area 80a. Furthermore, with user permission, corrected inference result 51 may include the label non-overlapping area 80b and the label detachment area 80c.

[0055] (First and second learning-complete models)

[0056] The first learned-complete model 31a is a general learned-complete model that can output a first inference result 50a corresponding to various inspection objects 90. That is, the first learned-complete model 31a is a learned-complete model that is designed to output the corresponding inference result 50, i.e., the first inference result 50a, for any X-ray image 40 of any of the various inspection objects 90 input. Furthermore, the first inference result 50a is an example of the "inference result" of the claim.

[0057] The second learning completion model 31b is a dedicated learning completion model that can output a second inference result 50b corresponding to a specific type of inspection object 90 among a variety of inspection objects 90. The second learning completion model 31b is set up separately for each type of inspection object 90.

[0058] When the inference accuracy based on the first learned model 31a decreases, the correction region 81 generated in the first inference result 50a increases. Similarly, when the inference accuracy based on the second learned model 31b decreases, the correction region 81 generated in the second inference result 50b increases. Therefore, in this embodiment, the corrected inference result acquisition unit 20c (refer to...) Figure 3 ) constitutes the first inference result 50a (refer to Figure 10 The first revised inference result 51a was obtained (refer to...) Figure 10 Furthermore, the revised inference result acquisition unit 20c is configured to obtain the second inference result 50b (see reference). Figure 11 The second revised inference result 51b was obtained (refer to...) Figure 11 Furthermore, the first modified inference result 51a is an example of the "modified inference result" of the claim.

[0059] (Obtaining the first revised inference result)

[0060] First, refer to Figure 10 The structure of the inference result acquisition unit 20b and the corrected inference result acquisition unit 20c for acquiring the first corrected inference result 51a from the X-ray image 40 will be described.

[0061] For multiple inspection objects 90 (reference) Figure 2 The X-ray image 40 of any one of the objects 90 is input into the first learned model 31a by the inference result acquisition unit 20b, thereby obtaining the first inference result 50a of the area of ​​the object 90 to be inspected.

[0062] Then, the corrected inference result acquisition unit 20c obtains the result based on the operation unit 23 (refer to...). Figure 1 The user's input corrected the first inference result 50a and the first corrected inference result 51a.

[0063] Then, the control unit 20 accumulates and stores the first corrected inference result 51a of the various inspection objects 90 in a correspondence with their respective X-ray images 40. The control unit 20, for example, accumulates and stores the first corrected inference result 51a and X-ray images 40 in a correspondence using a folder structure in the storage unit 21 (see reference). Figure 1 In this context, the control unit 20 creates folders 24 for each type of the inspected object 90 (see reference). Figure 12 The first corrected inference result 51a and the X-ray image 40 are stored in the created folder 24. Furthermore, when obtaining the first inference result 50a using the first learned model 31a, the user inputs information determining the type of the object 90 to be inspected (e.g., substrate 91 (refer to...)). Figure 2 (e.g., part number). Therefore, the control unit 20 can, based on information determining the type of the object 90 to be inspected, accumulate and store the first corrected inference result 51a in correspondence with the X-ray image 40 for each type of the object 90 to be inspected.

[0064] (Obtaining the second revised inference result)

[0065] Next, refer to Figure 11 The structure of the inference result acquisition unit 20b and the corrected inference result acquisition unit 20c for acquiring the second corrected inference result 51b from the X-ray image 40 will be described.

[0066] The inference result acquisition unit 20b inputs the X-ray image 40 of a specific type of inspection object 90 into the second learning completed model 31b and obtains the inference result 50, i.e., the second inference result 50b, output from the second learning completed model 31b.

[0067] Then, the corrected inference result acquisition unit 20c obtains the result based on the operation unit 23 (refer to...). Figure 1 The user's input corrected the second inference result 50b and the second corrected inference result 51b.

[0068] Then, the control unit 20 accumulates and stores the second corrected inference result 51b of the specific type of inspection object 90 in a correspondence with the X-ray image 40. The control unit 20, for example, accumulates and stores the second corrected inference result 51b in a correspondence with the X-ray image 40 in the storage unit 21 (see, for example, a folder structure). Figure 1 In this context, the control unit 20 creates folders 25 for each type of the inspected object 90 (see reference). Figure 13The second corrected inference result 51b and the X-ray image 40 are stored in the created folder 25. Furthermore, when obtaining the second inference result 50b using the second learned model 31b, the user inputs information determining the type of the object 90 to be inspected (e.g., substrate 91 (refer to...)). Figure 2 (e.g., part number). Therefore, based on the information determining the type of the object 90 to be inspected, the control unit 20 accumulates and stores the second corrected inference result 51b in a corresponding relationship with the X-ray image 40 for each type of object 90 to be inspected.

[0069] Here, as a certain period of time passes since the initial production of the first completed learning model 31a, and with changes in the market, etc., the substrate 91 (refer to) installed on the inspection object 90 is... Figure 2 Electronic components 92 (refer to) Figure 2 The tendency of the types of electronic components 92, etc., may sometimes change. In the case described, due to the change between the tendency of the types of electronic components 92, etc., installed on the substrate 91 when the first learning model 31a was initially made and the tendency of the types of electronic components 92, etc., installed on the substrate 91 in the X-ray image 40 taken during the actual inspection, the inference accuracy of the first inference result 50a based on the first learning model 31a decreases.

[0070] Furthermore, as a certain period passes since the initial creation of the second learning model 31b, due to the changes in the X-ray irradiation unit 10 over the years, even if the same type of examination object 90 is photographed, the accuracy of the second inference result 50b based on the second learning model 31b will decrease due to changes in the contrast of the photographed X-ray image 40, etc. Additionally, if the X-ray irradiation unit 10 is replaced after years of change, even if the same type of examination object 90 is photographed, the contrast of the X-ray image (based on the teacher data from the initial creation of the second learning model 31b) and the actual examination X-ray image 40 will change. Even in this case, the accuracy of the second inference result 50b based on the second learning model 31b will decrease.

[0071] Therefore, after completing the model re-creation department 20d (refer to...) Figure 3 Then, the first learned model 31a is remade. In addition, the second learned model 31b is remade in the remaking section 20d.

[0072] (Re-creation of the model after completing the first learning)

[0073] Reference Figure 12 The structure of the re-creation of the first learned model 31a by the re-creation department 20d after learning is explained.

[0074] The learning-completed model re-production unit 20d uses the X-ray images 40 stored in a correspondence with various inspection objects 90 and the first corrected inference result 51a to re-produce the first learning-completed model 31a.

[0075] Folder 24a stores X-ray images 40 of inspection object 90 of type A and the first corrected inference result 51a. Folder 24b stores X-ray images 40 of inspection object 90 of type B and the first corrected inference result 51a. Folder 24c stores X-ray images 40 of inspection object 90 of type C and the first corrected inference result 51a. Figure 12 In the example shown, item 90 of type A is inspected six times. Item 90 of type B is inspected four times. Item 90 of type C is inspected twice. Furthermore, in... Figure 12 The example shown illustrates a structure in which folder 24 is separated to accumulate and store X-ray images 40 and first corrected inference results 51a in a corresponding relationship for three types of objects 90 to be examined. However, in practice, folder 24 is separated to accumulate and store X-ray images 40 and first corrected inference results 51a in a corresponding relationship for each of three or more types of objects 90 to be examined.

[0076] In this embodiment, the learning-completed model re-creation unit 20d re-creates the first learned-completed model 31a based on the first corrected inference result 51a and X-ray images 40 accumulated and stored in the storage unit 21 for each type of the inspected object 90. Specifically, the learning-completed model re-creation unit 20d re-creates the first learned-completed model 31a using X-ray images 40 and the first corrected inference result 51a for the inspected object 90, which are a predetermined number of types of inspected objects 90 in descending order of the number of times the processing to obtain the first inference result 50a has been performed. For example, in the case of re-creating the first learned-completed model 31a based on the first two inspected objects 90, the first learned-completed model 31a is re-created using a set of X-ray images 40 and the first corrected inference result 51a stored in folder 24a, and a set of X-ray images 40 and the first corrected inference result 51a stored in folder 24b. Furthermore, the user can change the X-ray images 40 and the corresponding first corrected inference results 51a, which are among the first few X-ray images used in the process of obtaining the first inference result 50a.

[0077] Furthermore, in this embodiment, when the learning completed model re-creation unit 20d re-creates the first learning completed model 31a, it does not use the teacher data from the initial creation of the first learning completed model 31a, but re-creates the first learning completed model 31a based on the first corrected inference result 51a and the X-ray image 40.

[0078] Furthermore, in this embodiment, the learning-completed model re-production unit 20d weights the number of X-ray images 40 and the number of first corrected inference results 51a used in re-production based on the number of times the process of obtaining the first inference result 50a has been performed, and re-produces the first learning-completed model 31a based on the weighted X-ray images 40 and the corresponding number of first corrected inference results 51a. When weighting is based on the number of times the process of obtaining the first inference result 50a has been performed, as the ranking of the number decreases, the number of X-ray images 40 and the corresponding number of first corrected inference results 51a used in the re-production of the first learning-completed model 31a are reduced. For example, when re-producing the first learning-completed model 31a using X-ray images 40 of the first two inspection objects 90 that have undergone the process of obtaining the first inference result 50a the most times, and the corresponding first corrected inference results 51a, the first inspection object 90 uses two X-ray images 40 and the corresponding two first corrected inference results 51a. In addition, the second object to be inspected, 90, uses an X-ray image 40 and a corresponding first corrected inference result 51a.

[0079] Furthermore, the re-created first learned model 31a is accumulated and stored in the storage unit 21. That is, the re-created first learned model 31a is stored separately from the first learned model 31a already stored in the storage unit 21. In other words, the first learned model 31a stored in the storage unit 21 is not overwritten by the newly re-created first learned model 31a. When accumulating and storing the first learned model 31a, the date and time of creation of the learned model 60 (see reference) are used. Figure 1 ) are stored in association.

[0080] (Re-creation of the model after completing the second learning phase)

[0081] Next, refer to Figure 13 The structure of the re-creation of the second learned model 31b by the re-creation department 20d after learning is explained.

[0082] The learning-completed model re-creation unit 20d uses the X-ray image 40 of the inspection object 90 for a specific type and the second corrected inference result 51b to re-create the second learning-completed model 31b. Specifically, the learning-completed model re-creation unit 20d re-creates the second learning-completed model 31b based on the second corrected inference result 51b and the X-ray image 40 accumulated and stored in the storage unit 21 for each type of inspection object 90.

[0083] In this embodiment, the learning-completed model re-creation unit 20d uses at least the latest X-ray image 40 and the second corrected inference result 51b to re-create the second learning-completed model 31b. Specifically, the learning-completed model re-creation unit 20d uses the latest X-ray image 40 and a second corrected inference result 51b corresponding to the latest X-ray image 40 to re-create the second learning-completed model 31b. That is, in this embodiment, when re-creating the second learning-completed model 31b, the learning-completed model re-creation unit 20d does not use the teacher data from the initial creation of the second learning-completed model 31b, but instead re-creates the second learning-completed model 31b based on the second corrected inference result 51b and the X-ray image 40.

[0084] Furthermore, the re-created second learned model 31b is accumulated and stored in the storage unit 21. That is, the re-created second learned model 31b is stored separately from the second learned model 31b already stored in the storage unit 21. In other words, the second learned model 31b stored in the storage unit 21 is not overwritten by the newly re-created second learned model 31b. When accumulating and storing the second learned model 31b, the date and time of creation of the learned model 60 (see reference) are used. Figure 1 ) are stored in association.

[0085] (Re-creation to promote the display of information)

[0086] Figure 14 This is the inference result display screen 70, which displays the inference result 50. The inference result display screen 70 is displayed on the display unit 22 (see reference). Figure 1 ).

[0087] The inference result display screen 70 displays the inference result 50 and the learning completion model management screen 71 (see below). Figure 15 The learning process is complete. (Click the "Model Management" button 70a).

[0088] The "Model Management" button 70a, which is displayed on the inference results display screen 70, is a pressable button in the graphical user interface (GUI).

[0089] When the learning completion model management button 70a is pressed, the control unit 20 (refer to) Figure 3 )show Figure 15 The learning completed model management screen shown is 71.

[0090] The learning completed model management screen 71 displays information about the first learned model 31a (see reference). Figure 1 ) and the second completed learning model 31b (refer to Figure 1 The screen displays information about the completed learning model. Additionally, the completed learning model management screen 71 is used when creating the first completed learning model 31a and the second completed learning model 31b. The completed learning model management screen 71 is displayed on the display unit 22 (…). Figure 1 ).

[0091] like Figure 15 As shown, the learning completed model management screen 71 displays a learning button 71a and a display bar 72 that shows the management information of the learned model 31. The learning button 71a is a pressable button displayed in the GUI of the learning completed model management screen 71.

[0092] Display bar 72 includes a general learning completed model display bar 72a and a dedicated learning completed model display bar 72b.

[0093] The general learning-completed model display panel 72a displays information about the first learning-completed model 31a. Additionally, the dedicated learning-completed model display panel 72b displays information about the second learning-completed model 31b.

[0094] In addition, the display bar 72 includes a sequence name display bar 72c, a latest learning date and time display bar 72d, a re-creation promotion information display bar 72e, and a checkbox display bar 72f.

[0095] The sequence name display bar 72c displays the analysis name when the analysis is performed using the learned model 31. Furthermore, since the first learned model 31a is a general learning model, the sequence name display bar 72c does not display the analysis name.

[0096] The latest learning date and time display bar 72d displays the date and time when the learning model 31 was last regenerated. Specifically, the latest learning date and time display bar 72d is based on the date and time stored in the storage unit 21 (see reference). Figure 1The learning completion model creation date and time 60 in the display shows the year, month, day, hour, minute, and second when the learning completion model 31 was last regenerated. In the latest learning date and time display column 72d, the year, month, day, hour, minute, and second when the first learning completion model 31a and the second learning completion model 31b were last regenerated are displayed respectively. Furthermore, if no learning completion model 31 has been regenerated, the latest learning date and time display column 72d displays the year, month, day, hour, minute, and second when the learning completion model 31 was initially created.

[0097] Reproduction promotion information 52 is displayed in the reproduction promotion information display panel 72e. Reproduction promotion information 52 includes first reproduction promotion information 52a and second reproduction promotion information 52b. First reproduction promotion information 52a is information that promotes the reproduction of the first learned model 31a. Second reproduction promotion information 52b is information that promotes the reproduction of the second learned model 31b.

[0098] If it is necessary to recreate the first completed learning model 31a, first re-creation promotion information 52a is displayed. Similarly, if it is necessary to recreate the second completed learning model 31b, second re-creation promotion information 52b is displayed. The display of the first re-creation promotion information 52a and the second re-creation promotion information 52b is controlled by the display control unit 20g (see reference). Figure 3 )conduct.

[0099] The display control unit 20g determines whether the first learned model 31a needs to be remade based on whether a predetermined period 61 has elapsed since the last remake of the first learned model 31a. Specifically, the display control unit 20g determines this by checking whether a predetermined period 61 has elapsed since the last remake of the first learned model 31a. Figure 3 The display control unit 20g compares the elapsed period since the last re-production of the first learned model 31a, obtained from the re-production period acquisition unit 20e, with the specified period 61 to determine whether to display the first re-production promotion information 52a. Specifically, if the elapsed period since the last re-production of the first learned model 31a, obtained from the re-production period acquisition unit 20e, is longer than the specified period 61, the display control unit 20g determines that the first learned model 31a needs to be re-produced. Furthermore, the specified period 61 is, for example, six months or one year. The specified period 61 is stored in the storage unit 21 and can be changed through user input.

[0100] Additionally, during the remanufacturing process, the data obtained from section 20e is stored in storage section 21 (see reference). Figure 1 The learning completion model creation date and time in 60 refers to the period elapsed from the creation date and time of the first learning completion model 31a to the current moment.

[0101] Furthermore, the display control unit 20g determines whether a second learned model 31b needs to be created based on whether the correction amount of the second inference result 50b is above a predetermined amount. Specifically, the display control unit 20g determines whether the correction amount of the second inference result 50b is above a threshold 62 (see reference). Figure 1 The above determines whether a second fully learned model 31b needs to be created. That is, if the correction amount of the second inference result 50b is above the threshold 62, the display control unit 20g determines that a second fully learned model 31b needs to be created.

[0102] Furthermore, the correction amount of the second inference result 50b is obtained by the correction amount acquisition unit 20f. Specifically, the correction amount acquisition unit 20f obtains the correction amount of the second inference result 50b based on the second inference result 50b and the second corrected inference result 51b. More specifically, the correction amount acquisition unit 20f obtains the correction amount of the second inference result 50b by subtracting the second inference result 50b from the second corrected inference result 51b.

[0103] like Figure 15 As shown, the promotional information 52 is an icon. Furthermore, in Figure 15 The learning-completed model management screen 71 shown illustrates an example where both first re-production promotion information 52a and second re-production promotion information 52b are displayed. However, if the elapsed period since the last re-production of the first learning-completed model 31a is less than a predetermined period 61, the first re-production promotion information 52a is not displayed. Furthermore, if the correction amount of the second inference result 50b is less than the threshold 62, the second re-production promotion information 52b is not displayed. Additionally, if either the first re-production promotion information 52a or the second re-production promotion information 52b is displayed, in... Figure 14 The inference results shown in the image 70 also display the re-creation promotion information 52 in the learning completion model management button 70a.

[0104] Checkbox 73 is displayed in checkbox display bar 72f. When the learning button 71a is pressed while checkbox 73 is checked, the learning completed model re-creation unit 20d re-creates the corresponding learning completed model 31. That is, when the learning button 71a is pressed while checkbox 73 in the general learning completed model display bar 72a is checked, the learning completed model re-creation unit 20d re-creates the first learning completed model 31a. In addition, when the learning button 71a is pressed while checkbox 73 in the dedicated learning completed model display bar 72b is checked, the learning completed model re-creation unit 20d re-creates the second learning completed model 31b.

[0105] Additionally, the general learning completed model display panel 72a shows a rollback button 74a and an initialization button 74b. Both the rollback button 74a and the initialization button 74b are GUI buttons.

[0106] When the rollback button 74a is pressed, the control unit 20 performs the process of returning the first learned model 31a to the previous first learned model 31a.

[0107] In addition, when the initialization button 74b is pressed, the control unit 20 performs the process of returning the first learned model 31a to the original first learned model 31a.

[0108] (After completing the first learning stage, the model will be remade and processed.)

[0109] Next, refer to Figure 16 For control unit 20 (refer to) Figure 1 Then recreate the first completed learning model 31a (refer to...) Figure 1 The processing of ) will be explained. Furthermore, Figure 16 The re-creation of the first learned model 31a shown is executed by the control unit 20 and stored in the storage unit 21 (see reference). Figure 1 Program 30 in ) (refer to) Figure 1 (to be carried out)

[0110] In step 101, the control unit 20 acquires the X-ray image 40 (refer to...). Figure 1 ).

[0111] Next, in step 102, the inference result acquisition section 20b (refer to...) Figure 3 ) Obtain the inference result 50 of the area of ​​the object 90 being examined by inputting the X-ray image 40 into the learned model 31 (refer to Figure 6 In this embodiment, the inference result acquisition unit 20b obtains a first inference result 50a for the corresponding inspection object 90 by inputting an X-ray image 40 containing any one of the various inspection objects 90 into the first learned model 31a (see reference). Figure 10 ).

[0112] Next, in step 103, the corrected inference result acquisition section 20c (refer to...) Figure 3 The corrected inference result 51 (refer to) was obtained based on the user's input and the corrected inference result 50. Figure 9 In this embodiment, the corrected inference result acquisition unit 20c acquires a first corrected inference result 51a that corrects the first inference result 50a based on the user's operation input (see reference). Figure 10 ).

[0113] Next, in step 104, the control unit 20 stores the obtained corrected inference result 51 in a corresponding relationship with the X-ray image 40. In this embodiment, the control unit 20 accumulates and stores the obtained first corrected inference result 51a in a corresponding relationship with the X-ray image 40 for each type of object 90 to be inspected. Specifically, whenever a corrected inference result 51 is obtained, the control unit 20 accumulates and stores the corrected inference result 51 and the X-ray image 40 for each type of object 90 to be inspected.

[0114] Next, in step 105, part 20e is obtained during the manufacturing process (see reference). Figure 3 The re-production period acquisition unit 20e obtains the elapsed time from the date and time of the re-production of the first learned model 31a to the current moment. Specifically, the re-production period acquisition unit 20e obtains the date and time information based on the current moment, and the information stored in the storage unit 21 (see reference 20). Figure 1 The learning completed model creation date and time 60 contains the date and time information of the last time the first learning completed model 31a was remade, and the time elapsed from the last remake date and time of the first learning completed model 31a to the current time is obtained.

[0115] Next, in step 106, the control unit 20 determines whether to display the remanufacturing promotion information 52 (see reference). Figure 15 Specifically, the display control unit 20g (refer to...) Figure 3 Determine whether the elapsed period since the last re-creation of the first learned model 31a has exceeded the specified period 61 (refer to...). Figure 1 More specifically, the display control unit 20g compares the elapsed period obtained by the re-production period acquisition unit 20e with the predetermined period 61. If the elapsed period is longer than the predetermined period 61, the display control unit 20g determines that the predetermined period 61 has elapsed since the last re-production of the first learned model 31a. If the predetermined period 61 has elapsed since the last re-production of the first learned model 31a, the process proceeds to step 107. If the predetermined period 61 has not elapsed since the last re-production of the first learned model 31a, the process proceeds to step 108.

[0116] When processing the transition from step 106 to step 107, that is, when it is determined that the elapsed period has exceeded the predetermined period 61, in step 107, the display control unit 20g displays re-production promotion information 52 that prompts the re-production of the learned model 31. Specifically, the display control unit 20g displays the first re-production promotion information 52a after the predetermined period 61 has elapsed since the last re-production of the first learned model 31a.

[0117] Next, in step 108, the control unit 20 determines whether there is an operation input to recreate the learned model 31. Specifically, the control unit 20... Figure 15 In the learning completed model management screen 71 shown, it is determined whether the learning button 71a was pressed while the checkbox 73 of the general learning completed model display bar 72a was checked. If the learning button 71a was pressed while the checkbox 73 of the general learning completed model display bar 72a was checked, the process proceeds to step 109. If the learning button 71a was pressed while the checkbox 73 of the general learning completed model display bar 72a was not checked, or if the learning button 71a was not pressed, the process ends.

[0118] When processing proceeds from step 108 to step 109, in step 109, the learned model re-creation unit 20d re-creates the first learned model 31a using X-ray images 40 stored in a correspondence-based manner with respect to various inspection objects 90 and corrected inference results 51. Specifically, the learned model re-creation unit 20d re-creates the first learned model 31a using X-ray images 40 accumulated and stored according to each type of inspection object 90 and the first corrected inference result 51a. In this embodiment, the inspection object 90 is solder ball 93. Therefore, in step 109, the learned model re-creation unit 20d uses X-ray images 40 as input data and re-creates the first learned model 31a, outputting the solder ball region 80a as the first inference result 50a.

[0119] Furthermore, in step 109, the learning-completed model re-creation unit 20d uses the X-ray image 40 as input data to re-create the first learning-completed model 31a. The output of the first learning-completed model 31a serves as the first inference result 50a, which is used as teacher data when creating the second learning-completed model 31b. Additionally, the type of object 90 for inspection in the first corrected inference result 51a used in step 109 is different from the type of object 90 for inspection in the first corrected inference result 51a used in the previous creation of the first learning-completed model 31a.

[0120] Furthermore, in this embodiment, during the processing of step 109, the learning-completed model re-creation unit 20d uses the X-ray image 40 of the inspection object 90 and the first corrected inference result 51a to re-create the first learning-completed model 31a. The inspection object 90 is a predetermined number of types of inspection objects 90 selected from a variety of inspection objects 90, ordered from highest to lowest number of times the step 103 for obtaining the first inference result 50a was performed. Specifically, the learning-completed model re-creation unit 20d uses the X-ray image 40 of the inspection object 90 and the first corrected inference result 51a to re-create the first learning-completed model 31a. The inspection object 90 is a predetermined number of types of inspection objects 90 selected from a variety of inspection objects 90, ordered from highest to lowest number of times the step 103 for obtaining the first inference result 50a was performed from now until a predetermined period. In addition, the learning completed model re-production unit 20d weights the number of X-ray images 40 and the first corrected inference result 51a used in the re-production based on the number of times the processing 103 to obtain the first inference result 50a has been executed, and re-produces the first learning completed model 31a based on the weighted X-ray images 40 and the first corrected inference result 51a.

[0121] (After completing the second learning phase, the model will be remade and processed.)

[0122] Next, refer to Figure 17 For control unit 20 (refer to) Figure 3 Then create the second completed learning model 31b (refer to...) Figure 1 The processing of ) will be explained. Furthermore, Figure 17 The reprocessing of the second learned model 31b shown is executed by the control unit 20 and stored in the storage unit 21 (see reference). Figure 1 Program 30 in ) (refer to) Figure 1 (to be carried out)

[0123] In step 200, the control unit 20 acquires an X-ray image 40 (refer to...). Figure 4 ).

[0124] Next, in step 201, the inference result acquisition section 20b (refer to...) Figure 3 Obtain the inference result 50 output from the second learned model 31b, i.e., the second inference result 50b (refer to...). Figure 11 Specifically, the inference result acquisition unit 20b obtains a second inference result 50b of the object 90 by inputting the X-ray image 40 of a specific type of object 90 into the second learned model 31b.

[0125] Next, in step 202, the corrected inference result acquisition section 20c (refer to...) Figure 3The second corrected inference result 51b was obtained based on user input, which corrected the second inference result 50b (refer to...). Figure 11 ).

[0126] Next, in step 203, the control unit 20 accumulates and stores the second corrected inference result 51b of the specific type of inspection object 90 in a corresponding manner with the X-ray image 40.

[0127] Next, in step 204, the correction amount acquisition unit 20f (refer to...) Figure 3 Based on the second inference result 50b and the second corrected inference result 51b which is corrected according to the user's input, the correction amount of the second inference result 50b is obtained.

[0128] Next, in step 205, the display control unit 20g (see reference) Figure 3 Determine whether to display the second remanufacturing promotion information 52b (refer to) Figure 15 Specifically, the display control unit 20g determines whether the correction amount of the second inference result 50b obtained by the correction amount acquisition unit 20f is greater than or equal to a predetermined amount. More specifically, the display control unit 20g determines whether the correction amount of the second inference result 50b is equal to or greater than a threshold 62 (see reference). Figure 1 The above steps determine whether the correction amount of the second inference result 50b is greater than or equal to a predetermined amount. If the correction amount of the second inference result 50b is greater than or equal to the threshold 62, the process proceeds to step 206. If the correction amount of the second inference result 50b is less than the threshold 62, the process proceeds to step 207.

[0129] When processing proceeds from step 205 to step 206, in step 206, the display control unit 20g displays second remanufacturing promotion information 52b. In other words, the display control unit 20g displays second remanufacturing promotion information 52b when the correction amount is greater than or equal to a predetermined amount.

[0130] Next, in step 207, the control unit 20 determines whether there is an operation input to recreate the learned model 31. Specifically, the control unit 20... Figure 15 In the learning completed model management screen 71 shown, it is determined whether the learning button 71a was pressed while the checkbox 73 of the dedicated learning completed model display bar 72b was checked. If the learning button 71a was pressed while the checkbox 73 of the dedicated learning completed model display bar 72b was checked, the process proceeds to step 208. If the learning button 71a was pressed while the checkbox 73 of the dedicated learning completed model display bar 72b was not checked, or if the learning button 71a was not pressed, the process ends.

[0131] Next, in step 208, the learned model re-creation unit 20d uses the X-ray image 40 of the specific type of inspection object 90 and the second corrected inference result 51b to re-create the second learned model 31b. Then, the process ends.

[0132] [Variation Example]

[0133] Furthermore, the embodiments disclosed herein should be considered exemplary and not limiting in all respects. The scope of the invention is indicated by the claims rather than the description of the embodiments, and also includes all modifications (variations) within the meaning and scope equivalent to the claims.

[0134] For example, the control unit may be configured to overwrite the modified inference result and X-ray image stored in the storage unit with the latest modified inference result and X-ray image whenever a modified inference result is obtained; that is, only the latest modified inference result and X-ray image are stored. For example, the control unit may not be configured to display reproduction promotion information. For example, the control unit may be configured to reproduce the first completed learning model based on X-ray images and the first modified inference result for all types of inspection objects. For example, the control unit may be configured not to weight the number of X-ray images and the first modified inference result used in reproduction according to the number of times the process of obtaining the first inference result was performed, but instead to reproduce the first completed learning model using one X-ray image and one modified inference result for each inspection object ranked highest according to the type of inspection object and the number of times the process of obtaining the first inference result was performed. For example, the control unit may be configured to reproduce the second completed learning model using teacher data that includes the latest second modified inference result and X-ray image if the latest second modified inference result and X-ray image are included. That is, the control unit can also be configured to create a second learned model based on the latest X-ray image and the corresponding latest second-corrected inference result, and one or more other X-ray images and the corresponding one or more other second-corrected inference results.

[0135] Additionally, for example, in Figure 16 Alternatively, steps 107 and 108 can be omitted, and the first learning completed model can be created after a specified period without displaying the first re-creation promotion information or asking the user.

[0136] Additionally, for example, in Figure 17 Alternatively, steps 206 and 207 can be omitted, and the second learning completed model can be created when the correction amount reaches or exceeds the specified amount, without displaying the second re-creation promotion information or asking the user.

[0137] Alternatively, the control unit may be configured to use the teacher data from the initial creation of the first learned model, along with the first revised inference result and X-ray image, when recreating the first learned model. In this case, the versatility of the first learned model is improved due to the increased variety of teacher data. However, it may be unable to cope with changes in the types of objects being examined, resulting in a decrease in the inference accuracy of the first learned model. That is, when the teacher data used to recreate the first learned model includes the teacher data from the initial creation of the first learned model, versatility and inference accuracy become a trade-off. For example, the control unit may also be configured to use the teacher data from the initial creation of the second learned model, along with the second revised inference result and X-ray image, when recreating the second learned model. However, if the inference accuracy of the second learned model decreases due to changes in the X-ray irradiation unit or the replacement of the X-ray irradiation unit after changes over time, the inference accuracy of the second learned model may not improve when using teacher data that includes the teacher data from the initial creation of the second learned model. Therefore, it is preferable that the control unit is configured such that, when recreating the second completed learning model, the teacher data from the initial creation of the second completed learning model is not used, but the second corrected inference result and the X-ray image are used to recreate the second completed learning model.

[0138] Alternatively, for example, the X-ray image could be an image capturing an object other than solder balls. In this case, the learned model only needs to be configured to perform image conversion capable of identifying the area containing the object other than solder balls in the X-ray image, as well as other areas. For example, the learned model could also be configured to perform image conversion capable of identifying the area containing multiple solder materials in the connection portion of multiple terminals of a Land Grid Array (LGA) package with terminals arranged in a grid pattern, as well as other areas. Furthermore, the learned model could also perform image conversion capable of identifying the area containing multiple terminals, which are not solder materials, as well as other areas.

[0139] Additionally, for example, the present invention can also be applied to structures in which multiple solder balls are arranged on both the surface and back sides of a substrate.

[0140] [form]

[0141] Those skilled in the art will understand that the exemplary embodiments described are specific examples of the following forms.

[0142] (Project 1)

[0143] A method for recreating a learned model is provided for recreating a learned model of a region of an object being examined, obtained from an X-ray image of the object. In this method...

[0144] The learned model is a universal learned model corresponding to various inspection objects, and the method for recreating the learned model includes:

[0145] The steps for obtaining the X-ray image;

[0146] The step of inputting the X-ray image of any one of the various objects to be inspected into the first learned model, thereby obtaining a first inference result of the region of the object to be inspected;

[0147] The step of obtaining a first corrected inference result based on user input, which corrects the first inference result;

[0148] The step of storing the first corrected inference results regarding the various objects to be examined in correspondence with their respective X-ray images; and

[0149] The step of recreating the first learned model using the X-ray images stored in a correspondence-based manner with respect to various of the inspected objects and the first corrected inference result.

[0150] Because the initial inference result may contain some errors, it needs to be appropriately corrected through user operation. By pre-storing the first corrected inference result, the stored first corrected inference result can be used to recreate the first completed learning model. That is, the execution of the inspection of the object also serves as the creation of the teacher data required to recreate the new first completed learning model. Therefore, users do not need to perform the additional task of creating the teacher data required for recreating the new first completed learning model, which reduces the user's burden while recreating the first completed learning model.

[0151] The first learning completion model is a general learning completion model, created by using various inspection objects as teacher data input. However, since the types of inspection objects performed by users sometimes change over time (trends change), the accuracy of the first inference result based on the general learning completion model (first learning completion model) decreases. However, in this invention, the first revised inference result for various inspection objects is stored, and the first learning completion model is recreated using the stored first revised inference results. Since the first revised inference result used in the recreated model is the inference result for various inspection objects that actually underwent the processing that output the first inference result (i.e., the inspection was performed), it is considered that the recreated general learning completion model is more suitable for various inspection objects that users will also perform inspections on in the future. Therefore, by using the recreated general learning completion model, the accuracy of the inference result can be improved.

[0152] (Project 2)

[0153] According to the learning-completed model re-creation method of Project 1, in the re-creation step, the first learning-completed model is re-created using the X-ray image of the inspection object and the first corrected inference result, wherein the inspection object is a variety of inspection objects of a specified number in descending order of the number of times the step of obtaining the first inference result has been performed.

[0154] In this scenario, the various types of objects to be checked in the re-creation of the general learned model are concentrated on those that the user frequently uses the general learned model to output first inference results, while those that are used less frequently are excluded from the scope of objects. Therefore, the re-created general learned model focuses on learning the types that the user frequently uses to output first inference results, thus making it more suitable for various objects that the user will need to check in the future.

[0155] (Project 3)

[0156] According to the learning-completed model re-creation method of Project 1, in the re-creation step, the first learning-completed model is re-created using the X-ray image of the inspection object and the first corrected inference result, wherein the inspection object is a variety of inspection objects of a specified number of types, ordered from most to least number of times the step to obtain the first inference result was performed from now until a specified period.

[0157] In this case, the various types of objects to be checked in the re-creation of the general learning-completed model are concentrated on the types that the user has actually used the general learning-completed model to output the first inference result most frequently from now until the specified period. Therefore, the re-created general learning-completed model is focused on learning the types that the user has actually used the general learning-completed model to output the first inference result most recently, and is thus better suited for the various objects to be checked that the user will also perform checks on in the future.

[0158] (Project 4)

[0159] According to the learning-complete model re-creation method described in Project 2, in the re-creation step, the number of X-ray images and the number of the first corrected inference results used in the re-creation of the first learning-complete model are weighted according to the number of times the step of obtaining the first inference result is executed, and the first learning-complete model is re-created using the weighted X-ray images and the first corrected inference results.

[0160] In this scenario, the more frequently a user actually uses the general learning-completed model to output a first inference result, the greater its weight is assigned; conversely, the fewer frequently a user actually uses the general learning-completed model to output a first inference result, the smaller its weight is assigned. Therefore, regarding the reconstructed general learning-completed model, the more frequently a user actually uses the general learning-completed model to output a first inference result, the greater its influence on the learning process, thus making it more suitable for various inspection objects that the user will need to inspect in the future.

[0161] (Project 5)

[0162] According to the learning-completed model re-creation method described in Project 1, in the re-creation step, it is determined whether the elapsed period since the last re-creation of the first learning-completed model has exceeded a predetermined period. If it is determined that the period has exceeded the predetermined period, the first learning-completed model is re-created using the X-ray images of the various inspection objects and the first corrected inference result.

[0163] It is believed that the longer the period since the last creation of the general learning completion model, the higher the likelihood that the type of object being inspected by the user has changed (the trend has changed). Therefore, in this invention, when a predetermined period has elapsed since the last creation of the general learning completion model, by recreating the first learning completion model, a first learning completion model suitable for the type of object being inspected in which the user recently actually used the general learning completion model to output the first inference result can be used.

[0164] (Project 6)

[0165] According to the learning-completed model re-creation method described in Project 1, the type of the inspection object in the first corrected inference result used in the re-creation is different from the type of the inspection object in the first corrected inference result used in the previous creation of the first learning-completed model.

[0166] The types of objects to be checked in the first corrected inference result used in the re-production are different from the types of objects to be checked in the first corrected inference result used in the production of the previous first learned complete model. Therefore, the learned complete model that is suitable for the types of objects to be checked in the inference result output by the user in the most recent general learned complete model can be used.

[0167] (Project 7)

[0168] According to the learning-complete model re-creation method described in Project 1, the learning-complete model further includes a dedicated second learning-complete model corresponding only to a specific type of inspection object among the various inspection objects, and the learning-complete model re-creation method further includes:

[0169] The step of obtaining the second inference result of the object under inspection by inputting the X-ray image of the specific type of object under inspection into the second learned model;

[0170] The step of obtaining a second corrected inference result based on user input, which corrects the second inference result;

[0171] The step of establishing a correspondence between the second corrected inference result of the specific type of examination object and the X-ray image and storing it; and

[0172] The step of using the X-ray image of the object under examination of the specific type and the second corrected inference result to recreate the second learned model.

[0173] Even with a dedicated learned model for a specific type of object being examined, the accuracy of the inference results can sometimes change due to the deterioration of the hardware used to acquire X-ray images over the years or the replacement of parts. In this invention, a second, corrected inference result is pre-stored for the dedicated learned model, thereby allowing the stored second, corrected inference result to be used to recreate the dedicated learned model (the second learned model). Therefore, by using the recreated dedicated learned model, the accuracy of the second inference result can be improved.

[0174] (Project 8)

[0175] According to the learning-complete model re-creation method described in Project 7, in the step of re-creation of the second learning-complete model, it is determined whether the correction amount of the second corrected inference result is above a predetermined amount. If it is determined to be above a predetermined amount, the second learning-complete model is re-created using the X-ray image of the inspection object of the specific type and the second corrected inference result.

[0176] The fact that the correction amount of the second inference result is greater than a specified amount means that the accuracy of the second inference result has changed significantly due to the deterioration of the hardware used to obtain X-ray images over the years or the replacement of parts. Therefore, a dedicated learned model (second learned model) can be created at an appropriate time to improve the accuracy of the second inference result.

[0177] (Project 9)

[0178] According to the learning-completed model re-fabrication method described in Project 1, the object to be inspected is a substrate with multiple solder balls.

[0179] The X-ray image is an image captured with the substrate in the image.

[0180] The area to be inspected is the area containing the plurality of solder balls.

[0181] In the re-creation step, the X-ray image is used as input data to re-create the learned model of the solder ball region as the inference result.

[0182] (Project 10)

[0183] A system for recreating a learned model is provided for recreating a learned model of a region of an object being examined, obtained from an X-ray image of the object.

[0184] The learned model is a universal learned model corresponding to various inspection objects, and the learned model re-creation system includes:

[0185] An X-ray imaging apparatus includes an X-ray irradiation unit for irradiating X-rays and an X-ray detector for detecting X-rays irradiated from the X-ray irradiation unit; and

[0186] An image processing device generates the X-ray image.

[0187] The image processing device performs the following control:

[0188] Control of obtaining the X-ray image;

[0189] Control over obtaining inference results of regions of the inspected object obtained by inputting the X-ray image into the learned model, based on the X-ray image of any of the various inspected objects;

[0190] Control of the corrected inference result based on user input is obtained;

[0191] Control of storing the corrected inference results obtained regarding various objects under inspection in a correspondence with the X-ray images; and

[0192] The control of recreating the learned model is achieved by using the X-ray images stored in a correspondence-based manner with the corrected inference results for various of the inspected objects.

[0193] The same technical effects as in Project 1 can also be achieved in this invention.

[0194] (Project 11)

[0195] A learned model re-creation program is used to re-create a learned model of a region of an object being examined, obtained from an X-ray image of the object. In the learned model re-creation program...

[0196] The learned model is a universal learned model corresponding to various inspection objects. The learned model is then programmed to enable the computer to perform the following controls:

[0197] Control of obtaining the X-ray image;

[0198] Control over obtaining inference results of regions of the inspected object obtained by inputting the X-ray image into the learned model, based on the X-ray image of any of the various inspected objects;

[0199] Control of the corrected inference result based on user input is obtained;

[0200] Control of storing the corrected inference results obtained regarding various objects under inspection in a correspondence with the X-ray images; and

[0201] The control of recreating the learned model is achieved by using the X-ray images stored in a correspondence-based manner with the corrected inference results for various of the inspected objects.

[0202] The same technical effects as in Project 1 can also be achieved in this invention.

Claims

1. A method for recreating a learned model, used to recreate a learned model of a region of an object being examined, obtained from an X-ray image of the object being examined, wherein in the method for recreating the learned model, The learned model is a general first learned model corresponding to various inspection objects, and the method for recreating the learned model includes: The steps for obtaining the X-ray image; The step of inputting the X-ray image of any one of the various objects to be inspected into the first learned model, thereby obtaining a first inference result of the region of the object to be inspected; The step of obtaining a first corrected inference result based on user input, which corrects the first inference result; The step of storing the first corrected inference results for the various objects to be examined in correspondence with their respective X-ray images; as well as The step of recreating the first learned model using the X-ray images stored in a correspondence-based manner with respect to various of the inspected objects and the first corrected inference result.

2. The method for recreating a learned model according to claim 1, wherein, In the re-creation step, the first learned model is re-created using the X-ray images of the objects to be examined and the first corrected inference result, wherein the objects to be examined are a number of different types of objects to be examined, ordered from most to least number of times the step of obtaining the first inference result has been performed.

3. The method for recreating a learned model according to claim 1, wherein, In the re-creation step, the first learned model is re-created using the X-ray images of the objects to be examined and the first corrected inference result, wherein the objects to be examined are a specified number of types of objects to be examined in descending order of the number of times the step to obtain the first inference result was performed from now until a specified period.

4. The method for recreating a learned model according to claim 2, wherein, In the re-creation step, the number of X-ray images and the number of the first corrected inference results used in the re-creation of the first learned model are weighted according to the number of times the step of obtaining the first inference result is performed, and the first learned model is re-created using the weighted X-ray images and the first corrected inference results.

5. The method for recreating a learned model according to claim 1, wherein, In the re-creation step, it is determined whether the elapsed period since the last re-creation of the first learned model has exceeded a predetermined period. If it is determined that the period has exceeded the predetermined period, the first learned model is re-created using the X-ray images of the various objects under inspection and the first corrected inference result.

6. The method for recreating a learned model according to claim 1, wherein, The type of object to be checked in the first corrected inference result used during the re-creation is different from the type of object to be checked in the first corrected inference result used in the previous creation of the first learned model.

7. The method for recreating a learned model according to claim 1, wherein, The learned model further includes a dedicated second learned model that corresponds only to a specific one of the various inspected objects. The method for recreating the learned model further includes: The step of obtaining a second inference result for the object under inspection by inputting the X-ray image of the specific object under inspection into the second learned model; The step of obtaining a second corrected inference result based on user input, which corrects the second inference result; The step of establishing a correspondence between the second corrected inference result of the specific type of the object being examined and the X-ray image and storing it; and The step of recreating the second learned model using the X-ray image of the specific object under examination and the second corrected inference result.

8. The method for recreating a learned model according to claim 7, wherein, In the step of recreating the second learned model, it is determined whether the correction amount of the second corrected inference result is above a predetermined amount. If it is determined to be above a predetermined amount, the second learned model is recreated using the X-ray image of the specific object to be examined and the second corrected inference result.

9. The method for recreating a learned model according to claim 1, wherein, The object to be inspected is a substrate with multiple solder balls. The X-ray image is an image captured with the substrate in the image. The area to be inspected is the area containing the plurality of solder balls. In the re-creation step, the X-ray image is used as input data to re-create the first learned model of the solder ball region as the first inference result.

10. A system for recreating a learned model, used to recreate a learned model of a region of an object being examined, obtained from an X-ray image of the object being examined, wherein the system for recreating a learned model... The learned model is a universal learned model corresponding to various inspection objects, and the learned model re-creation system includes: An X-ray imaging apparatus includes an X-ray irradiation unit for irradiating X-rays and an X-ray detector for detecting X-rays irradiated from the X-ray irradiation unit. as well as An image processing device generates the X-ray image. The image processing device performs the following control: Control of obtaining the X-ray image; Control over obtaining inference results of regions of the inspected object obtained by inputting the X-ray image into the learned model, based on the X-ray image of any of the various inspected objects; Control of the corrected inference result based on user input is obtained; Control of storing the corrected inference results obtained regarding various objects under inspection in a correspondence with the X-ray images; as well as The control of recreating the learned model is achieved by using the X-ray images stored in a correspondence-based manner with the corrected inference results for various of the inspected objects.

Images