TWI837586B - Training data creation assistance apparatus and training data creation assistance method - Google Patents

Abstract

The present invention relates to a training data preparation support device, wherein an abnormal data acquisition unit acquires abnormal data, the abnormal data being an image showing an inspection object that is previously determined to have a defect. A normal data acquisition unit acquires normal data showing an image of a normal inspection object in a manner corresponding to the abnormal data acquired by the abnormal data acquisition unit; a differential data generation unit generates differential data, the differential data being a difference between the abnormal data acquired by the abnormal data acquisition unit and the normal data acquired by the normal data acquisition unit in a manner corresponding to the abnormal data; and a prompting unit prompts an image of a portion of the inspection object based on the differential data generated by the differential data generation unit.

Description

Training material production support device and training material production support method

The present invention relates to a training material production support device and a training material production support method for supporting the production of training materials.

For products such as products or foods, inspections are appropriately performed to determine whether they are good products in each step from the manufacturing process to before distribution. For example, Japanese Patent Publication No. 2011-119471 describes a defect inspection device that is used to inspect various defects generated in the manufacturing steps of semiconductor wafers.

In the defect inspection device, a plurality of SEM (scanning electron microscope) images showing a plurality of inspection target wafers are generated. The user specifies an SEM image showing a good wafer from the generated plurality of SEM images as a template. By calculating the difference between each of the plurality of SEM images (inspection images) other than the template and the template, defects on the circuit pattern of the wafer shown in the inspection image are detected.

In a defect inspection device, there is a possibility that an inspection object that does not have defects may be judged as having defects. In this case, although it does not have defects, the inspection object is discarded due to the wrong inspection result, which will cause a decrease in yield. Therefore, it is expected to perform inspection with higher accuracy.

The object of the present invention is to provide a training data preparation support device and a training data preparation support method, which can easily prepare training data for inspection with high accuracy.

(1) A training data preparation support device according to one aspect of the present invention supports the preparation of training data used for re-inspection of an inspection object, and comprises: an abnormal data acquisition unit for acquiring abnormal data, wherein the abnormal data is an image showing an inspection object that has been previously determined to have defects; and a normal data acquisition unit for acquiring normal data in a manner corresponding to the abnormal data acquired by the abnormal data acquisition unit. A training data preparation support device includes a training data preparation support device and a training data preparation support device. The ... Therefore, the user can create training data for rechecking the part by confirming the prompted image. In this way, it is easy to create training data for checking with high accuracy.

(2) The training data preparation support device may further include: an accepting unit that accepts the selection of the differential data presented by the presenting unit; and a registering unit that registers the differential data selected by the accepting unit as the training data. In this case, the differential data selected by the user is registered as the training data. This makes it easier to prepare the training data.

(3) The receiving unit may further receive the correction of the differential data prompted by the prompting unit, and the registering unit may register the corrected differential data as training data. In this case, it is possible to more appropriately correct the part of the image based on the differential data that requires re-inspection of the inspection object. In this way, it is possible to create training data for inspection with higher accuracy.

(4) The normal data acquisition unit can also acquire multiple normal data in a manner corresponding to the abnormal data, and the differential data generation unit can also generate multiple differential data based on the abnormal data and the multiple normal data corresponding to the abnormal data. In this case, multiple differential data are generated from one abnormal data. In this way, the operating efficiency of training data production can be improved.

(5) The normal data acquisition unit may also acquire a plurality of normal data in a manner corresponding to the abnormal data, and the differential data generation unit may also generate differential data based on the abnormal data and the average of a plurality of normal data corresponding to the abnormal data. According to this configuration, even if a noise component unrelated to the defect is accidentally mixed into any of the plurality of normal data, the plurality of normal data is averaged, so that the noise component has almost no effect on the pixel value of the normal data after the average. Therefore, training data for inspection with higher accuracy can be produced.

(6) The normal data acquisition unit may acquire normal data in a manner corresponding to a plurality of abnormal data, and the differential data generation unit may generate differential data based on each abnormal data and the normal data corresponding to the abnormal data. In this case, the training data can be quickly generated using the common normal data.

(7) The normal data acquired by the normal data acquisition unit may also include image data showing an image of an inspection object that has not been determined to have defects in advance. In this case, it is easy to acquire normal data showing an image of a normal inspection object.

(8) The normal data acquired by the normal data acquisition unit may also include master data, which is a design drawing showing the inspection object. In this case, it is easy to acquire normal data showing an image of a normal inspection object.

(9) The normal data acquisition unit can also acquire the master data corrected according to the processing accuracy of the inspection object as normal data. According to this configuration, even if the inspection object area is fine, normal data showing an image of a normal inspection object can be easily acquired.

(10) The abnormal data acquired by the abnormal data acquisition unit and the normal data acquired by the normal data acquisition unit are used to set a non-inspection target area, and the differential data generation unit may exclude the set non-inspection target area to generate differential data. In this case, it is possible to prevent the image displayed by the differential data from including a portion other than the inspection target area. In this way, training data for performing higher-precision inspection can be produced.

(11) The abnormal data acquisition unit can also further acquire defect information, which indicates the type of defect of the inspection object corresponding to the acquired abnormal data, and the differential data generation unit can also provide the generated differential data with the defect information acquired by the abnormal data acquisition unit. In this case, the user does not need to provide defect information to the training data. This can reduce the burden on the user and improve the efficiency of training data production. In addition, since it will not cause errors accompanying the user's operation, more accurate training data can be produced.

(12) The abnormal data acquisition unit may acquire the abnormal data after binarization, and the normal data acquisition unit may acquire the normal data after binarization. In this case, since the data amount of the abnormal data and the normal data is reduced, the training data can be generated at a high speed.

(13) Another aspect of the present invention is a training data production support method, which supports the production of training data used for re-inspection of an inspection object, and includes the following steps: a step of obtaining abnormal data, wherein the abnormal data is an image showing an inspection object that has been previously determined to have defects; a step of obtaining normal data showing an image of a normal inspection object in a manner corresponding to the acquired abnormal data; a step of generating differential data, wherein the differential data is a difference between the acquired abnormal data and normal data acquired in a manner corresponding to the abnormal data; and a step of presenting an image of a portion of the inspection object based on the generated differential data.

According to the training data production support method, an image of a portion of the inspection object is presented based on differential data of the difference between abnormal data of an image of the inspection object that is previously determined to have a defect and normal data of an image of the inspection object that is normal. The portion of the inspection object in the image presented based on the differential data is likely to require re-inspection. Therefore, the user can produce training data for re-inspecting the portion by confirming the presented image. In this way, it is easy to produce training data that can be inspected with higher accuracy.

[1] First Implementation Form (1) Processing System The following uses diagrams to describe a training data production support device and a training data production support method according to an implementation form of the present invention. In the following description, the training data production support device is referred to as a support device. FIG. 1 is a diagram showing the structure of a processing system including a support device according to the first implementation form of the present invention. As shown in FIG. 1 , the processing system 100 includes a processing device 10, an inspection device 20, and a database storage device 30.

The processing device 10 is composed of a CPU (central processing unit) 11, a RAM (random access memory) 12, a ROM (read-only memory) 13, a storage device 14, an operation unit 15, a display device 16, and an input/output I/F (interface) 17. The CPU 11, the RAM 12, the ROM 13, the storage device 14, the operation unit 15, the display device 16, and the input/output I/F 17 are connected to a bus 18.

RAM12 is used as a working area of CPU11. System programs are stored in ROM13. Storage device 14 is a storage medium including a hard disk or a semiconductor memory, and stores a training data production support program (hereinafter referred to as the support program). The support program can also be stored in ROM13 or other external storage devices. CPU11, RAM12 and ROM13 form a support device 40 for executing training data production support processing (hereinafter referred to as support processing). The support processing supports the production of training data.

The operation unit 15 is an input device such as a keyboard, a mouse or a touch panel. The user can provide a predetermined instruction to the support device 40 by operating the operation unit 15. The display device 16 is a display device such as a liquid crystal display device, and displays a GUI (Graphical User Interface) for receiving instructions generated by the user. The input/output I/F 17 is connected to the inspection device 20.

The inspection device 20 is, for example, an AOI (Automatic Optical Inspection) device, and generates a plurality of image data showing images of a plurality of inspection objects by sequentially photographing the inspection objects, and stores each of the generated image data. A unique identification number is given to each of the stored image data.

Although the inspection device 20 is described below using a substrate as an example of an inspection object, the inspection object is not limited to the substrate. The so-called substrate refers to a semiconductor substrate, a liquid crystal display device or an organic EL (Electro Luminescence) display device FPD (Flat Panel Display) substrate, an optical disk substrate, a magnetic disk substrate, a magneto-optical disk substrate, a mask substrate, a ceramic substrate, or a solar cell substrate.

The inspection device 20 processes the stored image data according to a predetermined algorithm to inspect the substrate corresponding to each image data. The inspection device 20 may inspect the substrate corresponding to each image data according to deep learning. During the inspection, it is determined whether the substrate has defects. In addition, for a substrate determined to have defects, the type of the defect is determined.

In the inspection device 20, even a substrate without defects may be judged as having defects. If a substrate is discarded due to an erroneous judgment even though it does not have defects, the yield will be reduced. Therefore, the substrate judged to have defects is re-inspected by supervised learning. The support device 40 supports the production of training data used for re-inspection. The database storage device 30 is a large-capacity storage device including a server. The produced training data is registered in the database storage device 30. The detailed structure of the support device 40 is explained below.

(2) Support device FIG. 2 is a diagram showing the structure of the support device 40 of FIG. 1. FIG. 3 is a diagram showing various data used in the preparation of training data. FIG. 4 is a diagram showing an example of a display screen of the display device 16 during the preparation of training data. As shown in FIG. 2, the support device 40 includes an abnormal data acquisition unit 41, a normal data acquisition unit 42, a differential data generation unit 43, a prompt unit 44, an acceptance unit 45, and a registration unit 46 as functional units. The CPU 11 of FIG. 1 executes a support program stored in the ROM 13 or the memory device 14, etc., thereby realizing the functional unit of the support device 40. It is also possible to realize part or all of the functional unit of the support device 40 by hardware such as an electronic circuit.

The abnormal data acquisition unit 41 acquires image data (hereinafter referred to as abnormal data) from the inspection device 20. The image data is an image of each substrate that has been previously determined to have a defect by the inspection device 20. The image data can display an image of the entire substrate, or a partial image of the substrate if it is the same area. A portion of the image of the substrate based on the abnormal data acquired by the abnormal data acquisition unit 41 is displayed in the upper left portion of FIG. 3.

The normal data acquisition unit 42 acquires image data showing a normal substrate image (hereinafter referred to as normal data) in a manner corresponding to each abnormal data acquired by the abnormal data acquisition unit 41. In this example, the normal data is a predetermined image data showing an image of a substrate that has not been previously determined to have a defect by the inspection device 20, and is acquired from the inspection device 20. In the lower left portion of FIG. 3, a portion of the image of the substrate based on the normal data acquired by the normal data acquisition unit 42 is displayed.

Abnormal data and normal data correspond to each other. For example, when abnormal data is obtained by the abnormal data acquisition unit 41, the normal data acquisition unit 42 may acquire image data having an identification number one before the identification number of the abnormal data and displaying a normal substrate image as normal data. Alternatively, when abnormal data is obtained by the abnormal data acquisition unit 41, the normal data acquisition unit 42 may acquire image data having an identification number one after the identification number of the abnormal data and displaying a normal substrate image as normal data.

The differential data generating unit 43 generates new image data by calculating the difference between each pixel value of the abnormal data acquired by the abnormal data acquiring unit 41 and the normal data acquired by the normal data acquiring unit 42 in a manner corresponding to the abnormal data. Here, the image data generated by the differential data generating unit 43 is referred to as differential data. The right part of FIG. 3 shows a substrate image based on the differential data generated by the differential data generating unit 43. The differential data is an image showing a portion of the substrate that is likely to require re-inspection. Therefore, the differential data can serve as a mark showing a portion of the substrate that requires re-inspection.

The prompting unit 44 prompts the user with each differential data by causing the display device 16 to display a GUI 50 (FIG. 4), which includes an image based on each differential data generated by the differential data generating unit 43. As shown in FIG. 4, the GUI 50 includes an image display area 51, a login key 52, and a correction key 53. A plurality of images of the measurement object are displayed in the image display area 51. In this example, the image based on the differential data is displayed in the image display area 51 in a manner overlapping with the image based on the abnormal data, but it is also possible to display only the image based on the differential data in the image display area 51.

The receiving unit 45 receives an instruction to register the differential data. Specifically, the receiving unit 45 receives the differential data selected to be registered from the operating unit 15 in the GUI 50 displayed by the prompting unit 44 as training data. The user can select an arbitrary image using the operating unit 15 while recognizing the image displayed in the image display area 51, and by operating the registration key 52, the instruction to select the differential data displayed in the image as training data is supplied to the receiving unit 45. The registration unit 46 registers the differential data selected by the receiving unit 45 in the database storage device 30 as training data.

In addition, the receiving unit 45 can receive corrections to the differential data. The user selects any image in the image display area 51 using the operating unit 15, and operates the correction key 53 to instruct the receiving unit 45 to correct the differential data corresponding to the image. In addition, the user can also use the operating unit 15 to color the portion that needs to be re-examined in the selected image.

When the receiving unit 45 accepts the designation, the differential data generating unit 43 adds a mark indicating the portion of the selected differential data that accepts the designation. In this way, the differential data is corrected. After the differential data is corrected, when the register key 52 is operated, the registering unit 46 registers the corrected differential data as training data in the database storage device 30.

(3) Support processing FIG. 5 is a flowchart showing the support processing performed by the support device 40 of FIG. 2 . The support processing of FIG. 5 is performed by the CPU 11 of FIG. 1 executing the support program stored in the ROM 13 or the memory device 14 on the RAM 12 . The support processing is described below using the support device 40 of FIG. 2 and the flowchart of FIG. 5 .

First, the abnormal data acquisition unit 41 acquires each abnormal data from the inspection device 20 (step S1). Then, the normal data acquisition unit 42 acquires normal data corresponding to each abnormal data acquired in step S1 from the inspection device 20 (step S2). Steps S1 and S2 may also be executed simultaneously.

Then, the difference data generating unit 43 generates each difference data based on the abnormal data and normal data corresponding to each other obtained in steps S1 and S2 (step S3). Then, the prompting unit 44 prompts the user with each difference data by causing the display device 16 to display an image based on each difference data generated in step S3 (step S4).

Next, the accepting unit 45 determines whether to accept the correction of any of the differential data presented in step S4 (step S5). When the correction of the differential data is not accepted, the accepting unit 45 proceeds to step S7. When the correction of any of the differential data is accepted, the differential data generating unit 43 corrects the differential data by giving a mark indicating that the designated part of the differential data is accepted for correction (step S6), and proceeds to step S7.

In step S7, the receiving unit 45 determines whether there is an instruction to register any of the differential data presented in step S4 or the differential data corrected in step S6 (step S7). When there is no instruction to register the differential data, the receiving unit 45 proceeds to step S9. When there is an instruction to register any of the differential data, the registering unit 46 registers the indicated differential data as training data in the database storage device 30 (step S8), and proceeds to step S9.

In step S9, the registering unit 46 determines whether there is an instruction to terminate (step S9). The user can instruct to terminate or continue by performing a predetermined operation using the operating unit 15. When there is no instruction to terminate, the registering unit 46 returns to step S5. When further registering differential data, the user instructs to continue without instructing to terminate. When the instruction to terminate is given, the registering unit 46 terminates the support processing.

(4) Effect In the support device 40 of the present embodiment, the abnormal data acquisition unit 41 acquires abnormal data, which is an image showing an inspection object that has been previously determined to have a defect. The normal data acquisition unit 42 acquires normal data showing an image of a normal inspection object in a manner corresponding to the abnormal data acquired by the abnormal data acquisition unit 41. The differential data generation unit 43 generates differential data, which shows the difference between the abnormal data acquired by the abnormal data acquisition unit 41 and the normal data acquired by the normal data acquisition unit 42 in a manner corresponding to the abnormal data. In the image displayed based on the differential data, there is a high possibility that a part of the inspection object needs to be re-inspected.

Therefore, based on the differential data generated by the differential data generating unit 43, a partial image of the inspection object is presented by the presenting unit 44. The accepting unit 45 accepts the selection of the differential data presented by the presenting unit 44. The registering unit 46 registers the differential data selected by the accepting unit 45 as training data in the database storage device 30. In this case, the user selects the differential data corresponding to the desired image while recognizing the presented image, and the selected differential data can be registered as training data. In this way, it is easier to create training data.

In addition, the receiving unit 45 further receives the correction of the differential data prompted by the prompting unit 44. The registering unit 46 registers the corrected differential data as training data. In this case, the user can more appropriately correct the part of the image that needs to be re-inspected according to the differential data. In this way, training data that can implement higher-precision inspection can be prepared.

(5) Variations FIG. 6 is a diagram showing various data used in the preparation of training data in the first variation. In the first variation, the user can pre-register a non-inspection target area outside the inspection target area in the inspection device 20. When the non-inspection target area is registered, the inspection device 20 sets the non-inspection target area in the generated image data.

Therefore, as shown in the upper left portion of FIG. 6 , a non-inspection object area is set in the abnormal data obtained by the abnormal data acquisition unit 41. Similarly, as shown in the lower left portion of FIG. 6 , a non-inspection object area is set in the normal data obtained by the normal data acquisition unit 42. In this case, as shown in the right portion of FIG. 6 , when the set non-inspection object area is excluded, the differential data generation unit 43 generates differential data by calculating the difference between each pixel value of the abnormal data and the normal data corresponding to the abnormal data. The pixel value of the non-inspection object area in the differential data can also be set to 0.

In this case, it is possible to prevent the portion outside the inspection target area from being marked as a portion that needs to be re-inspected. Thus, training data for inspection with higher accuracy can be produced. In addition, since the portion outside the inspection target area is no longer inspected, the substrate can be re-inspected at a high speed by using the produced training data.

FIG. 7 is a diagram showing various data used for preparing training data in the second modification. As shown in the upper left portion of FIG. 7 , the abnormal data acquisition unit 41 acquires abnormal data and acquires defect information of the defect type of the substrate displayed in the image displayed by the abnormal data. As shown in the right portion of FIG. 7 , the differential data generation unit 43 provides the generated differential data with the defect information acquired by the abnormal data acquisition unit 41. In the image display area 51 of the GUI 50, the image can also display the type of defect (e.g., color) based on the differential data.

In this case, the user does not need to provide defect information for the training data. This can reduce the burden on the user and improve the efficiency of training data production. In addition, since there will be no errors caused by the user's operation, more accurate training data can be produced.

Fig. 8 is a diagram showing various data used for preparing training data in the third modification. In the third modification, the user can pre-set the content of the binarization process for the image data in the inspection device 20. When the binarization process is set, the inspection device 20 generates the binarized image data.

Therefore, as shown in the upper left part of FIG8 , the abnormal data acquisition unit 41 acquires the abnormal data that has been binarized. Similarly, as shown in the lower left part of FIG8 , the normal data acquisition unit 42 acquires the normal data that has been binarized. As shown in the right part of FIG8 , the differential data generation unit 43 generates differential data by calculating the difference between each pixel value of the binarized abnormal data and the normal data that has been binarized corresponding to the abnormal data.

In this case, since the amount of image data is reduced, training data can be produced at high speed. In addition, by using the produced training data, the substrate can be re-inspected at high speed.

[2] Second Implementation Form In the first implementation form, although the normal data acquisition unit 42 acquires normal data in a manner that one normal data corresponds to one abnormal data, this implementation form is not limited to this. The following describes the differences between the support processing in the second to fourth implementation forms and the support processing in the first implementation form.

FIG. 9 is a diagram showing various data used for preparing training data in the second embodiment. In this embodiment, as shown in the lower left portion of FIG. 9 , a plurality of normal data are obtained in such a manner that a plurality of normal data corresponds to one abnormal data. The differential data generating unit 43 generates differential data by calculating the difference between each pixel value of one abnormal data and each normal data corresponding to the abnormal data. Therefore, as shown in the right portion of FIG. 9 , a plurality of differential data are generated corresponding to one abnormal data.

According to this structure, a plurality of differential data can be generated from one abnormal data, thereby improving the efficiency of training data creation.

[3] Third Implementation Form FIG. 10 is a diagram showing various data used for preparing training data in the third implementation form. In this implementation form, as shown in the lower left portion of FIG. 10 , a plurality of normal data are obtained in such a manner that a plurality of normal data corresponds to one abnormal data. The differential data generating unit 43 generates differential data by calculating the difference between each pixel value of one abnormal data and the average of a plurality of normal data corresponding to the abnormal data. Therefore, as shown in the right portion of FIG. 10 , one differential data is generated corresponding to one abnormal data.

According to this configuration, even if a noise component unrelated to a defect is accidentally mixed into any of a plurality of normal data, the plurality of normal data are averaged, so the noise component has little effect on the pixel value of the averaged normal data. Therefore, by using the averaged normal data, training data for inspection with higher accuracy can be prepared.

[4] Fourth embodiment FIG. 11 is a diagram showing various data used for preparing training data in the fourth embodiment. In this embodiment, as shown in the upper left portion of FIG. 11, normal data is obtained in a manner such that one normal data corresponds to a plurality of abnormal data. The differential data generating unit 43 generates differential data based on each abnormal data and the normal data corresponding to the abnormal data. Therefore, as shown in the right portion of FIG. 11, it generates a plurality of differential data corresponding to the plurality of abnormal data.

According to this configuration, since it is not necessary to obtain normal data every time abnormal data is obtained, training data can be quickly prepared using common normal data. In this embodiment, image data showing an image of a normal substrate prepared in advance can also be obtained by the normal data acquisition unit 42 as normal data. In this case, step S2 in the support process can also be executed before step S1.

The image data prepared in advance may be any of the image data generated by the inspection device 20, but may also be master data such as CAD data showing the design drawing of the substrate. Here, in the image data generated by the inspection device 20 and the master data, the width of the pattern of the substrate in the image or the radius of curvature of the corner of the pattern may differ due to the processing accuracy of the substrate by etching, etc. Therefore, the master data may be corrected by changing the width of the pattern in the image or the radius of curvature of the corner of the pattern, etc., according to the processing accuracy of the substrate.

In addition, even in the first to third embodiments, although image data showing an image of a substrate that has not been determined to have a defect by the inspection device 20 in advance is obtained as normal data from the inspection device 20, the present embodiment is not limited to this. At least one of the normal data may be pre-prepared master data or corrected master data.

[5] Other implementation forms In the above implementation forms, the support device 40 includes the receiving unit 45 and the registration unit 46, but this implementation form is not limited to this. The support device 40 may not include the receiving unit 45 and the registration unit 46. In this case, the user can also create training data for re-examining the part by confirming the image prompted by the GUI 50. In this way, it is easy to create training data that can implement high-precision inspection.

[6] Correspondence between the constituent elements of the claim and the various parts of the implementation form Although the following describes the correspondence between the constituent elements of the claim and the various elements of the implementation form, the present invention is not limited to the following example. As the constituent elements of the claim, various other elements having the structure or function described in the claim may be used.

In the above-mentioned embodiment, the support device 40 is an example of a training data preparation support device, the abnormal data acquisition unit 41 is an example of an abnormal data acquisition unit, and the normal data acquisition unit 42 is an example of a normal data acquisition unit. The difference data generation unit 43 is an example of a difference data generation unit, the prompt unit 44 is an example of a prompt unit, the receiving unit 45 is an example of a receiving unit, and the registration unit 46 is an example of a registration unit.

10: Processing device 11: CPU (Central Processing Unit) 12: RAM (Random Access Memory) 13: ROM (Read Only Memory) 14: Memory device 15: Operation unit 16: Display device 17: Input/output I/F (Interface) 18: Bus 20: Inspection device 30: Database storage device 40: Support device 41: Abnormal data acquisition unit 42: Normal data acquisition unit 43: Differential data generation unit 44: Prompt unit 45: Acceptance unit 46: Login unit 50: GUI 51: Image display area 52: Login key 53: Correction key 100: Processing system

FIG. 1 is a diagram showing the configuration of a processing system including a support device of the first embodiment of the present invention. FIG. 2 is a diagram showing the configuration of the support device of FIG. 1 . FIG. 3 is a diagram showing various data used for the preparation of training data. FIG. 4 is a diagram showing an example of a display screen of a display device in the preparation of training data. FIG. 5 is a flowchart showing the support processing performed by the support device of FIG. 2 . FIG. 6 is a diagram showing various data used for the preparation of training data in the first variant. FIG. 7 is a diagram showing various data used for the preparation of training data in the second variant. FIG. 8 is a diagram showing various data used for the preparation of training data in the third variant. FIG. 9 is a diagram showing various data used for preparing training data in the second embodiment. FIG. 10 is a diagram showing various data used for preparing training data in the third embodiment. FIG. 11 is a diagram showing various data used for preparing training data in the fourth embodiment.

15: Operation Department

16: Display device

20: Inspection device

30: Database storage device

40: Support devices

41: Abnormal Data Acquisition Department

42: Normal data acquisition department

43: Differential data generation unit

44: Prompt Department

45: Reception Department

46: Registration Department

100:Processing system

Claims (12)

A training data preparation support device supports the preparation of training data used for re-inspection of an inspection object; the device comprises: an abnormal data acquisition unit, which acquires abnormal data, the abnormal data being an image showing an inspection object previously determined to have defects; a normal data acquisition unit, which acquires normal data showing an image of a normal inspection object in a manner corresponding to the abnormal data acquired by the abnormal data acquisition unit; and a differential data generation unit, which generates differential data as image data, the differential data being obtained by calculating the abnormal data acquired by the abnormal data acquisition unit. The inspection unit comprises a detection unit for detecting the abnormal data obtained by the detection unit and a detection unit for detecting the abnormal data, and a detection unit for detecting the abnormal data. The detection unit displays an image of a portion of the inspection object by comparing the difference of each pixel value of the normal data obtained by the normal data acquisition unit in a manner corresponding to the abnormal data; a display unit displays an image of a portion of the inspection object based on the differential data generated by the differential data generation unit; an acceptance unit accepts the selection of differential data registered as training data from the differential data corresponding to the image prompted by the acceptance unit; and a registration unit registers the differential data selected by the acceptance unit as training data. The training data preparation support device of claim 1, wherein the receiving unit further receives the correction of the differential data prompted by the prompting unit, and the registering unit registers the corrected differential data as training data. The training data preparation support device of claim 1 or 2, wherein the normal data acquisition unit acquires a plurality of normal data in a manner corresponding to the abnormal data, and the differential data generation unit generates a plurality of differential data based on the abnormal data and a plurality of normal data corresponding to the abnormal data. The training data preparation support device of claim 1 or 2, wherein the normal data acquisition unit acquires a plurality of normal data in a manner corresponding to the abnormal data, and the differential data generation unit generates differential data based on the average of the abnormal data and the plurality of normal data corresponding to the abnormal data. In the training data preparation support device of claim 1 or 2, the normal data acquisition unit acquires normal data in a manner corresponding to a plurality of abnormal data, and the differential data generation unit generates differential data based on each abnormal data and the normal data corresponding to the abnormal data. In the training data preparation support device of claim 1 or 2, the normal data acquired by the normal data acquisition unit includes image data showing an image of an inspection object that has not been previously determined to have defects. In the training data production support device of claim 1 or 2, the normal data acquired by the normal data acquisition unit includes master data showing a design drawing of the inspection object. As in the training data preparation support device of claim 7, the normal data acquisition unit acquires the master data corrected according to the processing accuracy of the inspection object as the normal data. The training data preparation support device of claim 1 or 2, wherein a non-inspection target area is set in the abnormal data obtained by the abnormal data acquisition unit and the normal data obtained by the normal data acquisition unit, and the differential data generation unit excludes the set non-inspection target area to generate differential data. In the training data preparation support device of claim 1 or 2, the abnormal data acquisition unit further acquires defect information, which indicates the type of defect of the inspection object corresponding to the acquired abnormal data, and the differential data generation unit provides the generated differential data with the defect information acquired by the abnormal data acquisition unit. The training data production support device of claim 1 or 2, wherein the abnormal data acquisition unit acquires the abnormal data that has been binarized, and the normal data acquisition unit acquires the normal data that has been binarized. A training data preparation support method for supporting the preparation of training data used for re-inspection of an inspection object comprises the following steps: a step of obtaining abnormal data, the abnormal data being an image showing an inspection object previously determined to have defects; a step of obtaining normal data showing an image of a normal inspection object in a manner corresponding to the acquired abnormal data; a step of generating differential data as image data, the step of generating differential data as image data; The differential data is a step of displaying an image of a portion of the inspection object by calculating the difference between each pixel value of the acquired abnormal data and the normal data acquired in a manner corresponding to the abnormal data; a step of presenting an image of a portion of the inspection object based on the generated differential data; a step of accepting selection of differential data registered as training data from the differential data corresponding to the presented image; and a step of registering the selected differential data as training data.