JP7356962B2 - Machine learning device and method, and film thickness measuring device and method - Google Patents

Description

The present invention provides a machine learning device and a machine learning method for machine learning a machine learning model for recognizing polishing marks formed on a film, and a film thickness measuring device for measuring film thickness using the machine learning device and machine learning method. and a method for measuring film thickness.

When a film is attached to a member, if the film is not formed to a predetermined thickness, the purpose of attaching the film to the member will not be achieved. Therefore, the film thickness is measured. One method for measuring this film thickness is, for example, a film thickness measuring method disclosed in Non-Patent Document 1. This non-patent document 1 discloses a so-called calo test in which the film thickness is determined from polishing marks formed on the film by a ball.

M. G. Gee, et al. , “Ball Creating or Micro-Abrasion Wear Testing of Coatings”, Measurement Good Practice Guide No. 57, [Searched on September 15, 2020], Internet < URL: ht tp://eprintsubscriptions. npl. co. uk/2545/1/mgpg57. pdf>

By the way, in the film thickness measurement method disclosed in the above-mentioned Non-Patent Document 1, there are various cases in which the polishing marks are not necessarily perfect circles and are distorted. The outer circle (the boundary line between the film surface and the polishing marks) and the inner circle (the boundary line between the member and the polishing marks (the interface line between the film and the member)) of the polishing marks were manually set, and the film thickness was measured. has been done. For this reason, automation is difficult with the film thickness measuring method disclosed in Non-Patent Document 1. In particular, when measuring film thickness on a production line, automation is desired.

The present invention was made in view of the above-mentioned circumstances, and its purpose is to provide a machine learning device and a machine learning method that enable automation of film thickness measurement, and a machine learning device and a machine learning method equipped with the machine learning device and the machine learning method. An object of the present invention is to provide a film thickness measuring device and a film thickness measuring method for measuring film thickness.

As a result of various studies, the present inventors have found that the above object can be achieved by the following present invention. That is, a machine learning device according to one aspect of the present invention includes a learning data acquisition unit that acquires learning data, and a machine learning implementation unit that performs machine learning on a machine learning model based on the learning data acquired by the learning data acquisition unit. The machine learning model identifies a film part in the polishing mark and a member part in the polishing mark after the polishing mark is formed on the surface of a film formed on the surface of a predetermined member by a spherical polishing method. The learning data includes a film image obtained by capturing the film to include the polishing marks, and teacher data associated with each pixel of the film image, and the training data includes a film image in the polishing marks. It includes a first answer indicating that it is a part, and a second answer indicating that it is a part of the member in the polishing mark.

Such a machine learning device generates a machine learning model that identifies a film part in a polishing mark and a member part in the polishing mark. For this reason, by including the machine learning device in the film thickness measuring device, it is possible to use a machine learning model machine learned by the machine learning device from a target image taken of the film to be measured so as to include polishing marks. Accordingly, the film portion in the polishing mark and the member part in the polishing mark can be automatically identified, and based on the identification result, the film thickness can be calculated by, for example, the so-called Calotest film thickness calculation method. Therefore, the machine learning device described above enables automation of film thickness measurement.

In another aspect, in the above machine learning device, the machine learning model includes a plurality of pixels associated with each pixel of the film image, and the machine learning model includes a plurality of pixels associated with each pixel of the film image, and a film portion in the polishing mark and a member part in the polishing mark. Identification result images displayed in mutually different manners are output. Preferably, in the above-mentioned machine learning device, the machine learning model uses a first correct probability of a film portion in the polishing mark and a second correct probability of a member part in the polishing mark as a label, as the identification result. By determining each of a plurality of pixels associated with each pixel of an image, determining the maximum label at the pixel for each of the plurality of pixels, and displaying the pixel in a manner corresponding to the determined maximum label. , generate and output the identification result image.

Such a machine learning device displays the film part in the polishing mark and the member part in the polishing mark in different ways, so by referring to the identification result image, it is possible to visually distinguish between the film part and the member part. can be identified.

In another aspect, in the above-described machine learning device, the machine learning model further identifies a peeled portion of the film peeled from the surface of the member, and the training data further identifies a peeled portion of the film peeled from the surface of the member. The machine learning model outputs an identification result image in which the peeled part of the film peeled off from the surface of the member is displayed in a further different manner. Preferably, in the above-mentioned machine learning device, the machine learning model includes, as identification results, a first correct probability of a film portion in the polishing marks, a second correct probability of a member part in the polishing marks, and peeling from the surface of the member. The third correct probability of the peeled part of the film is obtained as a label for each of the plurality of pixels associated with each pixel of the film image, and for each of the plurality of pixels, the maximum label at the pixel is obtained, By displaying the pixel in a manner corresponding to the obtained maximum label, the identification result image is generated and output. Preferably, in the machine learning device described above, the machine learning model further identifies a surface portion of the membrane, the training data further includes a fourth answer indicating that the surface portion of the membrane The learning model outputs identification result images that display the surface portion of the membrane in further different ways. Preferably, in the above machine learning device, the machine learning model includes, as identification results, a first correct probability of a film portion in the polishing mark, a second correct probability of a member part in the polishing mark, and a separation probability from the surface of the member. A third correct probability for the peeled part of the film and a fourth correct probability for the surface part of the film are obtained as labels for each of the plurality of pixels associated with each pixel of the film image, and for each of the plurality of pixels. , the maximum label for the pixel is determined, and the pixel is displayed in a manner corresponding to the determined maximum label, thereby generating and outputting the identification result image. Preferably, in the above-mentioned machine learning device, the machine learning model divides the peeled part of the film peeled from the surface of the member into an outer isolated part formed outside the polishing mark and an outer isolation part formed inside the polishing mark. and the third answer includes a 3A answer representing the outer peeled portion and a 3B answer representing the inner peeled portion, and the machine learning model further identifies the inner isolated portion. When displaying the peeled portion of the film peeled off from the surface of the member in a further different manner, an identification result image in which the outer peeled portion and the inner peeled portion are displayed in a further different manner is output. Preferably, in the above machine learning device, the machine learning model includes, as identification results, a first correct probability of a film portion in the polishing mark, a second correct probability of a member part in the polishing mark, and a separation probability from the surface of the member. 3A correct answer probability of the outer peeled part in the peeled part of the film peeled off from the surface of the member, 3B correct answer probability of the inner peeled part in the peeled part of the film peeled from the surface of the member, and 4th correct answer probability of the surface part of the film. is obtained as a label for each of a plurality of pixels associated with each pixel of the film image, and for each of the plurality of pixels, the maximum label at that pixel is obtained, and in a manner corresponding to the obtained maximum label. By displaying the pixels, the identification result image is generated and output.

In such a machine learning device, the machine learning model further identifies the peeled portion of the film that has peeled off from the surface of the member, so that the film portion in the polishing trace and the member portion in the polishing trace can be more appropriately identified. .

A machine learning method according to another aspect of the present invention includes a learning data acquisition step of acquiring learning data, and a machine learning implementation step of machine learning a machine learning model based on the learning data acquired in the learning data acquisition step. The machine learning model identifies a film part in the polishing mark and a member part in the polishing mark after the polishing mark is formed on the surface of a film formed on the surface of a predetermined member by a spherical polishing method. The learning data includes a film image obtained by capturing the film to include the polishing marks, and teacher data associated with each pixel of the film image, and the training data includes a film image in the polishing marks. It includes a first answer indicating that it is a part, and a second answer indicating that it is a part of the member in the polishing mark.

Such a machine learning method generates a machine learning model that identifies a film part in a polishing mark and a member part in the polishing mark. Therefore, by including this machine learning method in the film thickness measurement method, it is possible to use a machine learning model machine learned by the machine learning device from a target image taken of the film to be measured so as to include polishing marks. Accordingly, the film portion in the polishing mark and the member part in the polishing mark can be automatically identified, and based on the identification result, the film thickness can be calculated by, for example, the so-called Calotest film thickness calculation method. Therefore, the machine learning method described above enables automation of film thickness measurements.

A film thickness measuring device according to another aspect of the present invention is a film thickness measuring device that measures a film thickness and includes one of the above-mentioned machine learning devices, and is configured to an image acquisition unit that acquires a target image obtained by imaging the membrane of and a film thickness processing section that calculates the film thickness based on the identification result image. Preferably, in the above-mentioned film thickness measuring device, the film thickness processing section calculates the film thickness using a Calotest film thickness calculation method.

According to this, it is possible to provide a film thickness measuring device equipped with any of the above-mentioned machine learning devices. Such a film thickness measuring device is equipped with a machine learning model that identifies a film portion in a polishing mark and a member part in the polishing mark, and therefore can automatically measure the film thickness.

In another aspect, in the above-mentioned film thickness measuring device, the film thickness processing unit determines an abnormality degree representing the degree of abnormality regarding the shape of the polishing mark based on the determined identification result image, and When the film thickness satisfies a predetermined condition, the film thickness is determined based on the determined identification result image. Preferably, in the above-mentioned film thickness measuring device, the film thickness processing section calculates the degree of circularity representing the degree of perfect circularity with respect to polishing marks in the film to be measured based on the determined identification result image. Find it as. Preferably, the film thickness processing unit determines the roundness based on the outer peripheral contour of the polishing mark (outer peripheral contour of the film portion) and the position of the center of gravity of the polishing mark in the determined identification result image.

Such a film thickness measuring device calculates the film thickness based on the identification result image when the degree of abnormality representing the degree of abnormality regarding the shape of the polishing mark satisfies a predetermined condition. Since the film thickness is measured with respect to the polishing marks, a reliable measurement value can be obtained.

In another aspect, in the above-mentioned film thickness measuring device, the film thickness processing unit further determines the film thickness based on the obtained identification result image when the obtained degree of abnormality does not satisfy a predetermined condition. inform the outside that it is not possible to request

Such a film thickness measuring device notifies the outside that the film thickness cannot be determined if the degree of abnormality does not meet a predetermined condition. Therefore, reliable measurement values can be obtained.

A film thickness measuring device according to another aspect of the present invention is a film thickness measuring device that measures a film thickness and includes the above-mentioned machine learning device, and which images a film to be measured so as to include the polishing marks. an image acquisition unit that acquires the target image obtained by the image acquisition unit, and a recognition result image for the target image acquired by the image acquisition unit, which is obtained by using a machine learning model machine learned by the machine learning implementation unit, and a film thickness processing unit that calculates the film thickness based on the determined identification result image, and the film thickness processing unit calculates a peeling degree representing the degree of peeling in the film to be measured based on the determined identification result image, and When the obtained peeling degree satisfies a predetermined second condition, the film thickness is obtained based on the obtained identification result image. Preferably, in the film thickness measuring device described above, the film thickness processing unit determines the degree of peeling based on one or more of the area, height, and spread of the peeled portion in the determined identification result image. Preferably, the film thickness processing unit calculates the area based on the number of pixels of the peeled portion in the determined identification result image. Preferably, the film thickness processing section calculates a length from the center of gravity of the polishing mark in the obtained identification result image to the position of a pixel farthest from the center of gravity of the peeled part in the obtained identification result image. The height is determined based on the above. Preferably, the film thickness processing unit generates a first line connecting the position of a pixel at one end in the circumferential direction of the peeled portion of the determined identification result image and the position of the center of gravity of the polishing mark in the determined identification result image. and a second line segment connecting the position of the other end pixel in the circumferential direction of the peeled portion of the identified identification result image and the position of the center of gravity is determined as the spread.

Such a film thickness measuring device calculates the film thickness based on the identification result image when the degree of peeling, which represents the degree of peeling in the film to be measured, satisfies a predetermined second condition. Since the film thickness is measured on polishing marks that were previously performed or on a film that has been appropriately formed on the surface of the member, reliable measured values can be obtained.

A film thickness measuring method according to another aspect of the present invention is a film measuring method for measuring film thickness, comprising the above-mentioned machine learning method, wherein the machine learning model is configured to correlate each pixel of the film image. outputting an identification result image in which the film part in the polishing mark and the member part in the polishing mark are displayed in mutually different manners, and the film to be measured is imaged so as to include the polishing mark. an image acquisition step of acquiring a target image; and a classification result image for the target image acquired in the image acquisition step, which is obtained using a machine learning model machine learned in the machine learning implementation step, and based on the determined classification result image. and a film thickness processing step of determining the film thickness.

According to this, a film thickness measuring method including the above-described machine learning method can be provided. Such a film thickness measurement method includes a machine learning model that identifies a film part in a polishing mark and a member part in the polishing mark, so that the film thickness can be automatically measured.

The machine learning device and machine learning method according to the present invention enable automation of film thickness measurement. According to the present invention, it is possible to provide a film thickness measuring device and a film thickness measuring method for measuring film thickness, which are equipped with a machine learning device and a machine learning method.

FIG. 1 is a block diagram showing the configuration of a film thickness measuring device including a machine learning device in an embodiment. As an example, it is a diagram showing an image of polishing marks. FIG. 3 is a diagram for explaining learning data. It is a flowchart which shows the operation of the film thickness measuring device regarding machine learning. 2 is a flowchart showing the operation of the film thickness measuring device regarding film thickness measurement. It is a figure which shows the measurement result by the said film thickness measuring device as an example. It is a flowchart which shows the operation|movement of the film thickness measuring device of a 1st modification regarding film thickness measurement. It is a figure for explaining the calculation method of the abnormality degree in the said 1st modification. It is a figure for explaining the calculation method of the degree of peeling in the film thickness measuring device of the 2nd modification. It is a figure for demonstrating the generation method of the identification result image in the film thickness measuring device of a 3rd modification.

Hereinafter, one or more embodiments of the present invention will be described with reference to the drawings. However, the scope of the invention is not limited to the disclosed embodiments. It should be noted that structures with the same reference numerals in each figure indicate the same structure, and the description thereof will be omitted as appropriate. In this specification, when referring to a general term, a reference numeral without a subscript is used, and when referring to an individual configuration, a reference numeral with a suffix is used.

A machine learning device in an embodiment includes a learning data acquisition unit that acquires learning data, and a machine learning implementation unit that performs machine learning on a machine learning model using the learning data acquired by the learning data acquisition unit. The machine learning model identifies a film part in the polishing mark and a member part in the polishing mark after the polishing mark is formed on the surface of a film formed on the surface of a predetermined member by a spherical polishing method. The learning data includes a film image obtained by capturing the film so as to include the polishing marks, and teacher data associated with each pixel of the film image. The teacher data includes a first answer indicating that the polishing mark is a film portion, and a second answer indicating that the polishing mark is a member portion. A film thickness measurement device that measures film thickness and includes such a machine learning device includes an image acquisition unit that acquires a target image of a film to be measured so as to include the polishing marks, and A film thickness processing unit that obtains a classification result image for the obtained target image using a machine learning model machine-learned by the machine learning implementation unit and obtains the film thickness based on the obtained classification result image. Hereinafter, embodiments will be described in more detail using a film thickness measuring device including a machine learning device as an example.

FIG. 1 is a block diagram showing the configuration of a film thickness measuring device including a machine learning device in an embodiment. FIG. 2 is a diagram showing an image of polishing marks as an example. FIG. 3 is a diagram for explaining learning data. FIG. 3A shows an example of a film image of a film without peeling, and FIG. 3B shows a painted film image as teacher data associated with the film image of FIG. 3A. FIG. 3C shows an example of a film image of a film with scratches caused by abrasives, etc., and FIG. 3D shows a painted film image as teacher data associated with the film image of FIG. 3C. FIG. 3E shows an example of a film image of a film with peeling, and FIG. 3F shows a painted film image as teacher data associated with the film image of FIG. 3E.

The film thickness measuring device in the embodiment is a device that measures the film thickness of a film formed on the surface of a predetermined member, and forms polishing marks on the surface of the film by a spherical polishing method, and measures the thickness of the film in the polishing marks. part and the component part in the polishing trace, and based on the radius of the outer circumferential contour of the membrane part (radius of the outer circle) and the radius of the inner circumferential contour of the membrane part (radius of the inner circle), for example, the so-called Calotest The film thickness is determined by a film thickness calculation method. In this embodiment, in order to automatically identify a film portion in a polishing mark and a member part in the polishing mark, this film thickness measuring device uses a film image to identify a film part in the polishing mark and a member part in the polishing mark from a film image. It is equipped with a machine learning device that performs machine learning to create a machine learning model that identifies. The predetermined member (base material) is arbitrary, and the film is also arbitrary.

For example, as shown in FIG. 1, a film thickness measuring device D including such a machine learning device includes a learning data acquisition section 1, an image acquisition section 2, a control processing section 3, an input section 4, and an output section. 5, an interface section (IF section) 6, and a storage section 7.

The learning data acquisition unit 1 is a device that is connected to the control processing unit 3 and acquires learning data under the control of the control processing unit 3. For example, the learning data acquisition unit 1 is an interface circuit that inputs and outputs data with an external device, and the external device is a storage medium that stores learning data. The storage medium is, for example, a USB (Universal Serial Bus) memory, an SD card (registered trademark), or the like. Alternatively, for example, the learning data acquisition unit 1 is a drive device that reads data from a recording medium on which learning data is recorded. Examples of the recording medium include a CD-ROM (Compact Disc Read Only Memory), a CD-R (Compact Disc Recordable), and a DVD-ROM (Digital Versatile Disc Read Only Memory). ) and DVD-R (Digital Versatile Disc Recordable), etc. . Alternatively, for example, the learning data acquisition unit 1 is a communication interface circuit that transmits and receives communication signals to and from an external device, and the external device is connected to a network (WAN (Wide Area Network), including a public communication network) or LAN (Local). This is a server device that is connected to the communication interface circuit via an area network (such as an area network) and manages learning data.

The learning data includes a film image obtained by capturing a film so as to include polishing marks, and teacher data associated with each pixel of the film image. The learning data is created by the creator observing the film image and adding training data.

An example of the membrane image is shown in FIG. 2. In FIG. 2, 48 film images are illustrated. In this example, a diamond-like carbon (DLC) film was applied to the surface of the steel material. Polishing marks are formed on the surface of the film by polishing until the member (base material) is exposed using a so-called spherical polishing method using a steel ball with a diameter of 30 [mm] coated with diamond paste as an abrasive. Ta.

The teacher data includes a first answer indicating that the polishing mark is a film part, and a second answer indicating that the polishing mark is a member part. In the present embodiment, in order to cope with a case where the film peels off from the surface of the member, and to more appropriately identify the film part in the polishing trace and the member part in the polishing trace, the teacher data is It further includes a third answer representing the peeled portion of the film peeled off from the surface. In this embodiment, in order to more appropriately identify the outer circumferential contour of the membrane portion and the inner circumferential contour of the membrane portion in the polishing trace, the third answer is an outer peeled portion formed outside the polishing trace. There are a 3rd A answer indicating that there is, and a 3rd B answer indicating that the inner peeled portion is formed within the polishing trace. In the present embodiment, in order to clarify the outer peripheral contour and the contour of the outer peeled portion, the teacher data is a fourth answer representing a surface portion of the film where no polishing marks are formed. further including. In this embodiment, such teacher data is represented by a colored image that is colored differently depending on the type of correct answer. As mentioned above, the types of correct answers are as follows: in this embodiment, the first answer (film part in the polishing mark), the second answer (member part in the polishing mark), the third answer (peeling part), and the third answer A (external peeling part). ), Answer 3B (inner peeled part)), and Answer 4 (surface part of the membrane). The mutually different aspects are realized, for example, in mutually different colors. Alternatively, for example, the mutually different aspects are realized by mutually different hatching. Alternatively, for example, the mutually different aspects are realized with mutually different brightnesses.

For example, the film image IMa shown in FIG. 3A is a film image without peeling, and from the film image IMa shown in FIG. A pixel value of a color (for example, green) is assigned, and a pixel value of a second color (for example, blue) is assigned to the pixel determined by the creator to be the second answer (part of the member in the polishing mark). By assigning a pixel value of a fourth color (for example, black) to the pixels determined to be Answer 4 (surface portion of the film), the image area AR1a of the first answer and the image area AR2a of the second answer shown in FIG. 3B are created. A colored image PCa having the image area AR4a of the fourth answer is generated. In addition, in the film image IMb of a film with scratches caused by abrasives, etc., shown in FIG. 3C, the type of correct answer is determined while ignoring the scratches. For example, from the film image IMb shown in FIG. , a colored image PCb having an image area AR1b of the first answer, an image area AR2b of the second answer, and an image area AR4b of the fourth answer shown in FIG. 3D is generated. For example, the film image IMc shown in FIG. 3E is a film image with peeling, and from the film image IMc shown in FIG. A pixel value of one color (for example, green) is assigned, and a pixel value of a second color (for example, blue) is assigned to the pixel determined by the creator to be the second answer (part of the member in the polishing trace). A pixel value of the 3rd A color (for example, red) is given to the pixel determined to be the 3rd A answer (outer peeled part), and a 3rd B color (for example, red) is assigned to the pixel determined by the creator to be the 3rd B answer (inner peeled part). For example, by assigning a pixel value of yellow) and assigning a pixel value of a fourth color (for example, black) to the pixel determined by the creator to be the fourth answer (surface portion of the membrane), the fourth answer shown in FIG. A colored image PCc is generated having an image area AR1c for the first answer, an image area AR2c for the second answer, an image area AR3Aa for the third A answer, an image area AR3Bc for the third B answer, and an image area AR4c for the fourth answer. Note that a known annotation tool such as Labelme developed at MIT may be used for the color separation. Furthermore, for convenience of illustration, in FIG. 3, the first color (e.g., green) is represented by hatching consisting of a plurality of straight lines descending to the left at relatively narrow intervals in plan view, and the second color (e.g., blue) is expressed by hatching consisting of a plurality of straight lines sloping downward to the right at relatively medium intervals in plan view, and the 3rd A color (for example, red) is expressed by a plurality of straight lines sloping downward to the right at relatively narrow intervals in plan view. The third B color (e.g., yellow) is represented by a hatching consisting of a plurality of straight lines descending to the left at relatively intermediate intervals in plan view, and the fourth color (e.g., black) is expressed in plan view by It is expressed by hatching consisting of multiple straight lines slanting to the left at relatively wide intervals.

A plurality of sets of such film images IM and coloring images PC as teacher data are created and used as learning data.

The film image serving as the learning data may be only the original image generated by capturing the polishing marks formed by performing the spherical polishing method with an imaging device, but in this embodiment, the original image and the original image are used. and a processed image generated by performing predetermined image processing on the image. The image processing may include, for example, a rotation process in which a numerical value in a range from 0 to 360° is randomly generated, and the original image is rotated by the rotation angle using the generated random numerical value as a rotation angle, for example. A zoom process (enlargement/reduction process) that randomly generates a numerical value in the range from 0.9 to 1.1 and uses the generated random numerical value as a magnification factor to zoom the original image by the magnification factor, and noise that follows a normal distribution for each pixel. and noise processing that generates and adds it to the original image (noise processing may be performed multiple times by considering the processed image after noise processing as the original image), and randomly generates a numerical value in a predetermined range and adds this This is a brightness change process that changes the brightness value of the original image using a generated random value. Note that each of these processes may be performed alone or in combination. Then, the film images (original images and processed images) of the learning data are changed to an image size according to the machine learning model described below by enlarging processing or reducing processing. In this embodiment, the film image of the learning data is changed to an image size of 496×372 by reduction processing. In this way, since the film image of the learning data includes not only the original image but also the processed image, the amount of the learning data is increased.

Returning to FIG. 1, the image acquisition unit 2 is connected to the control processing unit 3, and is a device that acquires a target image of the film to be measured so as to include the polishing marks under the control of the control processing unit 3. . The image acquisition unit 2 is, for example, a device that captures an image of a subject and generates an image of the subject, and includes, for example, an imaging optical system that forms an optical image of the subject on a predetermined imaging plane, and an imaging optical system that forms an optical image of the subject on a predetermined imaging plane. an image sensor arranged with a light-receiving surface aligned with the object, and converting an optical image of the object into an electrical signal; and an image sensor that processes the output of the image sensor to generate image data representing the image of the object. This is a digital camera equipped with an image processing unit and the like to generate images. Alternatively, for example, the image acquisition unit 2 is an interface circuit that inputs and outputs data with an external device, and the external device is a storage medium that stores a target image. Alternatively, for example, the image acquisition unit 2 is a drive device that reads data from a recording medium on which a target image is recorded. Alternatively, for example, the image acquisition unit 2 is a communication interface circuit that transmits and receives communication signals to and from an external device, and the external device is a server device that is connected to the communication interface circuit via a network and manages target images. It is.

Note that when the image acquisition section 2 is an interface circuit and the learning data acquisition section 1 is an interface circuit, the image acquisition section 2 and the learning data acquisition section 1 may share one interface circuit. Similarly, when the image acquisition section 2 is a communication interface circuit and the learning data acquisition section 1 is a communication interface circuit, the image acquisition section 2 and the learning data acquisition section 1 share one communication interface circuit. It's fine.

The input unit 4 is connected to the control processing unit 3, and inputs various commands such as a command to start machine learning, a command to start measuring film thickness, and a film thickness information such as the name of the film to be measured. It is a device for inputting various data necessary for operating the measuring device D into the film thickness measuring device D, and is, for example, a plurality of input switches, a keyboard, a mouse, etc. to which predetermined functions are assigned. The output unit 5 is a device that is connected to the control processing unit 3 and outputs commands and data input from the input unit 4, identification result images, measurement results, etc. under the control of the control processing unit 3, and is, for example, a CRT display, These include display devices such as liquid crystal displays and organic EL displays, and printing devices such as printers.

Note that the input section 4 and the output section 5 may constitute a so-called touch panel. When configuring this touch panel, the input section 4 is a position input device that detects and inputs an operating position, such as a resistive film type or a capacitive type, and the output section 5 is a display device. In this touch panel, the position input device is provided on the display surface of the display device, one or more input content candidates that can be input are displayed on the display device, and a display displaying the input content that the user wants to input is displayed. When a position is touched, the position is detected by the position input device, and the display content displayed at the detected position is input to the film thickness measuring device D as the user's operation input content. With such a touch panel, it is easy for the user to intuitively understand the input operations, so a film thickness measuring device D that is easy for the user to handle is provided.

The IF section 6 is a circuit that is connected to the control processing section 3 and performs input/output of data with an external device under the control of the control processing section 3, and is, for example, an interface circuit for RS-232C, which is a serial communication method. , an interface circuit using the Bluetooth (registered trademark) standard, an interface circuit that performs infrared communication such as the IrDA (Infrared Data Association) standard, and an interface circuit using the USB (Universal Serial Bus) standard. Further, the IF section 6 is a circuit that performs communication with an external device, and may be, for example, a data communication card, a communication interface circuit according to the IEEE802.11 standard, or the like.

Note that when the learning data acquisition section 1 is an interface circuit or a communication interface circuit, the IF section 6 may also be used as the learning data acquisition section 1. Similarly, when the image acquisition section 2 is an interface circuit or a communication interface circuit, the IF section 6 may also be used as the image acquisition section 2.

The storage section 7 is a circuit that is connected to the control processing section 3 and stores various predetermined programs and various predetermined data under the control of the control processing section 3. The various predetermined programs include, for example, a control processing program, and the control processing program controls each section 1, 2, 4 to 7 of the film thickness measuring device D according to the function of each section. A control program, a machine learning implementation program that performs machine learning on a machine learning model using the learning data acquired by the learning data acquisition unit 1, and an identification result image (described later) for a target image acquired by the image acquisition unit 2, are used in the machine learning implementation program. The film thickness processing program includes a film thickness processing program that calculates the film thickness using a machine learning model that is machine learned in , and calculates the film thickness based on the determined identification result image. The various predetermined data include data necessary for executing each of these programs, such as the radius of the sphere used to form the polishing mark and the actual length reflected in one pixel. Such a storage unit 7 includes, for example, a ROM (Read Only Memory) that is a nonvolatile storage element, an EEPROM (Electrically Erasable Programmable Read Only Memory) that is a rewritable nonvolatile storage element, and the like. The storage unit 7 includes a RAM (Random Access Memory), which serves as a so-called working memory of the control processing unit 3 that stores data generated during execution of the predetermined program. Note that the storage unit 7 may include a hard disk device capable of storing a large capacity in order to store a relatively large capacity of learning data.

The control processing section 3 is a circuit for controlling each section 1, 2, 4 to 7 of the film thickness measuring device D according to the function of each section, and measuring the film thickness. The control processing unit 3 includes, for example, a CPU (Central Processing Unit) and its peripheral circuits. The control processing section 3 functionally includes a control section 31, a machine learning implementation section 32, and a film thickness processing section 33 by executing the control processing program.

The control section 31 controls each section 1, 2, 4 to 7 of the film thickness measuring device D according to the function of each section, and controls the entire film thickness measuring device D.

The machine learning implementation unit 32 performs machine learning on a machine learning model using the learning data acquired by the learning data acquisition unit 1. The machine learning model is a model that identifies a film part in the polishing mark and a member part in the polishing mark after the polishing mark is formed on the surface of a film formed on the surface of a predetermined member by a spherical polishing method. be. More specifically, the machine learning model includes a plurality of pixels associated with each pixel of the film image, and the machine learning model includes an identification method that displays the film part in the polishing mark and the member part in the polishing mark in mutually different manners. Output the result image. More specifically, the machine learning model corresponds to each pixel of the film image using a first correct probability of a film portion in the polishing mark and a second correct probability of a member part in the polishing mark as a label as the identification result. The identification result image is obtained by determining each of the plurality of attached pixels, determining the maximum label at the pixel for each of the plurality of pixels, and displaying the pixel in a manner corresponding to the determined maximum label. Generate and output.

In this embodiment, the machine learning model further identifies a peeled portion of the film that has peeled off from the surface of the member, and the training data is the peeled portion of the film that has peeled off from the surface of the member, as described above. The machine learning model outputs an identification result image in which the peeled part of the film peeled off from the surface of the member is displayed in a further different manner. More specifically, the machine learning model has, as identification results, a first correct probability of a film portion in the polishing marks, a second correct probability of a member part in the polishing marks, and a peeled part of the film separated from the surface of the member. The third correct probability is determined as a label for each of the plurality of pixels associated with each pixel of the film image, and for each of the plurality of pixels, the maximum label at the pixel is determined, and the maximum label obtained is The identification result image is generated and output by displaying the pixels in a corresponding manner.

In this embodiment, the machine learning model further identifies a surface portion of the membrane, the training data further includes a fourth answer representing the surface portion of the membrane, as described above, and the machine learning model further identifies a surface portion of the membrane; The learning model outputs identification result images that display the surface portion of the membrane in further different ways. More specifically, the machine learning model has, as identification results, a first correct probability of a film portion in the polishing mark, a second correct probability of a member part in the polishing mark, and a peeled part of the film separated from the surface of the member. The third correct probability and the fourth correct probability of the surface portion of the membrane are obtained as labels for each of the plurality of pixels associated with each pixel of the membrane image, and for each of the plurality of pixels, the maximum The identification result image is generated and output by determining a label and displaying the pixel in a manner corresponding to the determined maximum label.

In the present embodiment, the machine learning model divides the peeled portion of the film peeled from the surface of the member into an outer isolation portion formed outside the polishing marks and an inner isolation portion formed within the polishing marks. and the third answer comprises a third A answer representing the outer peeled portion and a third B answer representing the inner peeled portion, as described above, and the machine learning model When displaying the peeled portion of the film peeled off from the surface of the member in a further different manner, an identification result image is output in which the outer peeled portion and the inner peeled portion are displayed in a further different manner.

That is, in the present embodiment, the machine learning model determines, as identification results, a first correct probability of a film portion in the polishing mark, a second correct probability of a member part in the polishing mark, and a probability of a film peeled off from the surface of the member. Label the 3rd A correct answer probability of the outer peeled part in the peeled part, the 3B correct answer probability of the inner peeled part in the peeled part of the film peeled from the surface of the member, and the 4th correct answer probability of the surface part of the film. , for each of the plurality of pixels associated with each pixel of the film image, and for each of the plurality of pixels, find the maximum label for the pixel, and label the pixel in a manner corresponding to the maximum label found above. By displaying, the identification result image is generated and output.

The machine learning model is a neural network used for deep learning, such as a convolutional neural network (CNN). More specifically, in this embodiment, PSPNet (Pyramid Scene Parsing Network) was used as a machine learning model.

The film thickness processing section 33 obtains a classification result image for the target image obtained by the image obtaining section 2 using a machine learning model machine learned by the machine learning implementation section 32, and determines the film thickness based on the obtained classification result image. This is what we seek. The film thickness processing unit 33 calculates the film thickness using, for example, the so-called Calotest film thickness calculation method. More specifically, first, the film thickness processing unit 33 determines the positions of the centers of gravity (geometric centers of gravity) of the film identification region identified as a film portion and the member identification region identified as a member portion in the identification result image. Next, the film thickness processing unit 33 calculates each peripheral pixel (pixel at the boundary between the film part and the surface part in the film part) forming the peripheral contour (outer circle) of the film identification area, and determines the position of the center of gravity. Each distance from to each position of each outer circumferential pixel is determined, and the average value of each distance is determined as the radius RO of the outer circle. By multiplying the number of pixels from the center of gravity to the position of the outer peripheral pixel by the actual length reflected in one pixel, the actual length of the radius RO of the outer circle is determined. Note that instead of the average value, the minimum value, maximum value, median value, etc. of each distance may be used as the radius RO of the outer circle. Next, the film thickness processing unit 33 processes each inner peripheral pixel (pixel at the boundary between the film parts in the member part, or in the film part) forming the outer peripheral outline (inner circle) of the member identification area. A pixel at the boundary between the membrane portion and the member portion) is determined, each distance from the center of gravity to each position of each inner peripheral pixel is determined, and the average value of each distance is determined as the radius RI of the inner circle. The actual length of the radius RI of the inner circle is determined by multiplying the number of pixels from the center of gravity to the inner circumferential pixel position by the actual length reflected in one pixel. Note that instead of the average value, the minimum value, maximum value, median value, etc. of each distance may be used as the radius RI of the inner circle. Next, the film thickness processing unit 33 calculates the difference between the outer circle radius RO and the inner circle radius RI as the radial width X (=RO-RI) of the film portion, and calculates the difference between the outer circle radius RO and the inner circle radius RI. The sum of the radius RI of the inner circle and the radius RI of the inner circle is determined as the length Y (=X+RI). Then, the film thickness processing section 33 calculates the film thickness t based on the width X, the length Y, and the radius R of the steel ball used to form the polishing marks, using the calculation formula 1 for calculating the film thickness of the so-called Calo test.
Film thickness calculation formula 1; t=(X×Y)/(2×R)

These control processing section 3, input section 4, output section 5, IF section 6, and storage section 7 can be configured by, for example, a desktop, notebook, or tablet computer. When the learning data acquisition section 1 is an interface circuit or a communication interface circuit, the IF section 6 can also be used as the learning data acquisition section 1. Therefore, the film thickness measuring device D including the learning data acquisition section 1 can be connected to a computer. configurable by Similarly, when the image acquisition unit 2 is an interface circuit or a communication interface circuit, the IF unit 6 can also be used as the image acquisition unit 2, so the film thickness measuring device D including the image acquisition unit 2 can be Configurable.

Next, the operation of this embodiment will be explained. FIG. 4 is a flowchart showing the operation of the film thickness measuring device regarding machine learning. FIG. 5 is a flowchart showing the operation of the film thickness measuring device regarding film thickness measurement. FIG. 6 is a diagram showing, as an example, measurement results by the film thickness measuring device. FIG. 6A shows the measurement results of the radius RI of the inner circle, FIG. 6B shows the measurement results of the outer circle radius RO, and FIG. 6C shows the measurement results of the film thickness. Each horizontal axis in FIG. 6 is an actual measurement value obtained from the film image by the creator, and each vertical axis is the measurement result obtained by the film thickness measuring device D.

In machine learning of the machine learning model, as shown in FIG. S11).

Then, in the film thickness measuring device D, the machine learning implementation unit 32 of the control processing unit 3 performs machine learning on the machine learning model using the learning data acquired by the learning data acquisition unit 1, and ends this process (S12). This enables the film thickness processing unit 33 to use the machine learning model to measure the film thickness for the target image.

In measuring the film thickness for a target image, in FIG. S21).

Next, the film thickness measuring device D uses the film thickness processing unit 33 of the control processing unit 3 to use the classification result image for the target image acquired by the image acquisition unit 2 in process S21 to perform machine learning using the machine learning execution unit 32. It is determined and generated using a learning model and stored in the storage unit 7 (S22).

Next, the film thickness measuring device D uses the film thickness processing unit 33 to determine the radius RO of the outer circle and the radius RI of the inner circle based on the identification result image, and stores them in the storage unit 7 (S23 ).

Next, the film thickness measuring device D calculates the width X (=RO-RI) and length Y (=X+RI) based on the radius RO of the outer circle and the radius RI of the inner circle by the film thickness processing unit 33. Based on the determined width X and length Y and the radius R of the steel ball used to form the polishing marks, the film thickness t (=(X×Y)/( 2×R)) is obtained and stored in the storage unit 7 (S23).

Next, in the film thickness measuring device D, the control unit 31 of the control processing unit 3 outputs the film thickness t of the measurement result to the output unit 5 (S25). In addition to the film thickness t, the radius RO of the outer circle and the radius RI of the inner circle may also be output to the output section 5. Further, if necessary, the control section 31 may output the measurement result, such as the film thickness t, from the IF section 6 to an external device.

Then, the control processing unit 3 of the film thickness measuring device D determines whether or not the measurement is finished (S26). If the result of this determination is that the measurement is finished (Yes), for example by turning off the power or inputting the end of measurement, the film thickness measuring device D terminates this process; If the process has not ended (No), the film thickness measuring device D returns the process to process S21.

Through such an operation, the film thickness measuring device D measures the film thickness of the film with respect to the polishing marks reflected in the target image.

An example of the measurement results by such a film thickness measuring device D is shown in FIG. In FIG. 6, the radius RO of the outer circle and the radius RI of the inner circle are determined not by the average value but by the maximum value of each distance as described above. Since the measurement results for the inner circle radius RI, outer circle radius RO, and film thickness t shown in FIGS. 6A, 6B, and 6C are approximately distributed on a 45-degree straight line, the machine learning model is By machine learning using learning data, the membrane portion in the polishing trace and the member portion in the polishing trace can be successfully identified. If we exclude the measurement results in the upper 7.5% of the degree of abnormality, which will be described later, which indicates the degree of abnormality regarding the shape of the polishing marks, in the example of the radius RI of the inner circle shown in Fig. 6A, the absolute value of the error is averaged. The mean absolute error (MAE) was 1.39, the root mean squared error (RMSE) was 1.87, and the coefficient of determination of the regression equation was 0.992. there were. In the example of the radius RO of the outer circle shown in FIG. 6B, the average absolute error was 1.83, the root mean square error was 2.26, and the coefficient of determination of the regression equation was 0.992. In the example of film thickness t shown in FIG. 6C, the average absolute error was 0.095, the root mean square error was 0.126, and the coefficient of determination of the regression equation was 0.991. Therefore, this film thickness measuring device D can appropriately measure the film thickness t.

As explained above, the machine learning device included in the film thickness measuring device D in this embodiment and the machine learning method implemented therein are machine learning methods for identifying film parts in polishing marks and member parts in the polishing marks. Generate the model. For this reason, by including the machine learning device and method, the film thickness measuring device and method can be used to determine whether the film thickness measuring device and method includes the machine learning device and the method. By using the trained machine learning model, it is possible to automatically identify the film part in the polishing mark and the component part in the polishing mark, and based on this identification result, the film thickness can be calculated using, for example, the so-called Calotest film thickness calculation method. can. Therefore, the machine learning device and method described above enable automation of film thickness measurement.

The machine learning device and the machine learning method display the film part in the polishing mark and the member part in the polishing mark in different ways. The parts can be identified. In this embodiment, the surface portion, the outer peeled portion, and the inner peeled portion of the membrane are further displayed in different manners, so that the machine learning device and method can visually , the membrane portion, the member portion, the surface portion of the membrane, the outer peeled portion, and the inner peeled portion can be distinguished.

In the above machine learning device D and machine learning method, the machine learning model further identifies the peeled part of the film that has peeled off from the surface of the member, so that the film part in the polishing mark and the member part in the polishing mark can be more appropriately identified. It becomes like this.

The film thickness measuring device D and the film thickness measuring method implemented therein in this embodiment are equipped with a machine learning model that identifies the film part in the film polishing mark and the member part in the polishing mark, so that the film thickness can be measured automatically. It can be implemented.

In the above-described embodiment, the film thickness processing unit 33 determines the degree of abnormality representing the degree of abnormality regarding the shape of the polishing mark based on the identification result image determined as described above, and determines whether the determined degree of abnormality is a predetermined degree. If the condition (first condition) is satisfied, the film thickness may be determined based on the identification result image. Then, the film thickness processing unit 33 further notifies the outside that the film thickness cannot be determined based on the determined identification result image when the determined degree of abnormality does not satisfy a predetermined condition. (first variant). For example, as shown in FIG. 6, some of the measurement results are distributed deviating from a 45 degree straight line. One reason for this is thought to be that the radius RO of the outer circle and the radius RI of the inner circle could not be appropriately determined as a result of the distortion of the polishing marks. The film thickness measuring device D of the first modified form calculates the film thickness based on the identification result image when the degree of abnormality representing the degree of abnormality regarding the shape of the polishing mark satisfies the predetermined first condition. Since the film thickness is measured on polishing marks that have been properly performed to a certain degree, reliable measurement values can be obtained. The film thickness measuring device D of the first modification notifies the outside that the film thickness cannot be determined when the degree of abnormality does not satisfy a predetermined first condition, so for example, the film thickness is determined manually. Reliable measurement values can be obtained by re-creating the polishing marks.

FIG. 7 is a flowchart showing the operation of the film thickness measuring device of the first modification regarding film thickness measurement. FIG. 8 is a diagram for explaining a method of calculating the degree of abnormality in the first modification. FIG. 8A is a schematic diagram showing a substantially perfect circular polishing mark, and FIG. 8B is a schematic diagram showing the polishing mark shown in FIG. It is a graph showing the distance from the position to the position of the center of gravity of the polishing mark (the position of the center of gravity of the membrane portion and the member portion) for each position up to the degree. FIG. 8C is a schematic diagram showing elliptical polishing marks, and FIG. 8D shows distances for each position in the polishing marks shown in FIG. 8C, similar to FIG. 8B for the polishing marks shown in FIG. 8A. It is a graph. FIG. 8E is a schematic diagram showing an irregularly shaped polishing mark, and FIG. 8F shows each distance at each position in the polishing mark shown in FIG. 8E, similar to FIG. 8B for the polishing mark shown in FIG. 8A. This is a graph showing.

In measuring the film thickness for the target image, in FIG. 7, the film thickness measuring device D of the first modified form first executes the process S21 of acquiring and storing the target image in the same way as described above, and then, A process S22 of generating and storing an identification result image is executed.

Next, in the film thickness measuring device D of the first modification, the film thickness processing section 33 of the control processing section 3 determines the degree of abnormality representing the degree of abnormality regarding the shape of the polishing mark, and stores it in the storage section 7 (S31 ).

More specifically, the film thickness processing unit 33 obtains the degree of circularity representing the degree of perfect circularity with respect to the polishing marks in the film to be measured as the degree of abnormality, based on the identification result image obtained in step S22. More specifically, the film thickness processing unit 33 determines the roundness based on the outer peripheral contour of the polishing mark (outer peripheral contour of the film portion) in the identification result image and the position of the center of gravity of the polishing mark. As shown in FIGS. 8A to 8D, when the polishing mark is a perfect circle or close to a perfect circle (ellipse in FIG. 8C), the outer circumferential contour of the polishing mark (the outer circumferential contour of the membrane portion) has a radius of 0 to 360 degrees in the circumferential direction. For each position up to a degree, the distance between the position and the center of gravity is approximately constant. On the other hand, if the polishing mark deviates from a perfect circle and becomes irregularly shaped, for example, as shown in FIGS. For each position from 360 degrees to 360 degrees, the distance between the position and the center of gravity changes. For this reason, for example, the film thickness processing unit 33 calculates the slope, which is a distance change for a unit angle change, for each position from 0 degrees to 360 degrees in the circumferential direction on the outer circumferential contour of the polishing mark (the outer circumferential contour of the film portion). (Differential value of distance with respect to angle) is determined, and its maximum value is determined as the degree of circularity (degree of abnormality). Alternatively, for example, the film thickness processing unit 33 determines the difference between the maximum value and the minimum value at each distance at each position as the roundness (the abnormality degree).

Next, in the film thickness measuring device D of the first modification, the film thickness processing unit 33 determines whether the degree of abnormality obtained in step S31 satisfies the first condition (S32). The first condition is a condition in which the degree of abnormality is low and a reliable film thickness can be determined based on the identification result image, for example, the degree of abnormality (roundness) is a predetermined condition. It is less than (or less than) a threshold value (anomaly determination threshold value). The abnormality determination threshold is appropriately set, for example, from a plurality of samples. As a result of this determination, if the degree of abnormality (roundness) is less than the abnormality determination threshold, it is determined that the degree of abnormality (roundness) satisfies the first condition (Yes), and the first modified form As described above, the film thickness measuring device D performs a process S23 of determining and storing each of the radius RO of the outer circle and the radius RI of the inner circle, a process S24 of determining and storing the film thickness t, and a process S24 of determining and storing the film thickness t. The process S25 for outputting the thickness t and the process S26 for determining the end are sequentially executed. On the other hand, as a result of the determination, if the degree of abnormality (roundness) is equal to or greater than the abnormality determination threshold, it is determined that the degree of abnormality (roundness) does not satisfy the first condition (No), and The film thickness measuring device D of the first modified form uses the film thickness processing unit 33 to notify the outside that the film thickness t cannot be determined based on the identification result image obtained in step S22 (S33), and determines the end. Processing S26 is executed. The notification is performed, for example, by displaying a message, outputting a voice message, or sounding an alarm.

Further, in the first modification described above, the degree of abnormality representing the degree of abnormality regarding the shape of the polishing mark was used, but instead of the degree of abnormality, the degree of peeling representing the degree of peeling in the film to be measured is used. (Second variant). In the second modification of the film thickness measuring device D, the film thickness processing section 33 calculates the peeling degree representing the degree of peeling in the film to be measured based on the identification result image calculated as described above, and calculates the degree of peeling in the film to be measured. When the peeling degree obtained satisfies a predetermined condition (second condition), the film thickness is determined based on the identification result image. Then, the film thickness processing unit 33 further notifies the outside that the film thickness cannot be determined based on the determined identification result image when the determined peeling degree does not satisfy a predetermined second condition. You may.

FIG. 9 is a diagram for explaining a method of calculating the peeling degree in the film thickness measuring device of the second modification. The degree of peeling is determined based on, for example, one or more of the area, height, and spread of the peeled portion in the identification result image. The area as the degree of peeling is determined based on the number of pixels of the peeled portion in the identification result image. More specifically, the area as the degree of peeling is determined by summing up the number of pixels in the image area of the peeled portion based on the identification result image DP. The area as the degree of peeling may be expressed by the number of pixels itself, or may be expressed by the actual area by multiplying the total number of pixels by the actual area reflected in one pixel. The height as the peeling degree is determined based on the length from the position of the center of gravity of the polishing mark in the identification result image to the position of the pixel farthest from the position of the center of gravity of the peeled portion in the identification result image. More specifically, as shown in FIG. 9, the height as the degree of peeling is determined based on the identification result image DP, for example, as shown in FIG. The length (distance) CPdHP between the position HP of the farthest pixel (farthest pixel) of the peeled part and the position CPd of the center of gravity with respect to the position CPd of the center of gravity of the area DA2 is determined, and this determined length is determined. The film thickness t is determined by regarding CPdHP as the radius RO of the outer circle, and regarding the film thickness t as the height of the peeling degree. For example, as shown in FIG. 9, the spread as the degree of peeling is based on the identification result image DP, and the position EP1 of a pixel at one end in the circumferential direction of the peeled part and the center of gravity of the polishing mark in the identification result image DP. The angle θ formed by the first line segment CPdEP1 that connects the position CPd of the peeled portion and the second line segment CPdEP2 that connects the position EP2 of the other end pixel in the circumferential direction of the peeled portion and the position CPd of the center of gravity is It is sought by seeking it as an expanse. A weighted average of a plurality of combinations of these may be used as the peeling degree.

The film thickness measuring device D of the second modification uses the peeling degree instead of the abnormality degree (S31a) and uses the second condition instead of the first condition (S32a) in measuring the film thickness of the target image. Except for this, the device operates in the same manner as the film thickness measuring device D of the above-described first modification shown in FIG. 7, so a description thereof will be omitted. The second condition is a condition in which the peeling degree is low and a reliable film thickness can be obtained based on the identification result image, for example, the peeling degree is set to a predetermined threshold (peeling determination threshold ) is less than (or less than). The peeling determination threshold is appropriately set, for example, from a plurality of samples.

Furthermore, in the above-described embodiment, if the machine learning model determines that the peeled portion is small, the machine learning model may be modified (third modification). To determine whether the machine learning model has determined that the peeled part is small, create a painted image of the target image, and compare the peeled part in the classification result image of the machine learning model for the target image with the peeled part in the created painted image. It can be determined by comparing.

FIG. 10 is a diagram for explaining a method of generating an identification result image in the film thickness measuring device of the third modification. FIG. 10A shows the identification result image before correction, and FIG. 10B shows the identification result image after correction.

In this embodiment, as described above, the machine learning model (machine learning model before modification) is based on the target image acquired by the image acquisition unit 2 using a model learned by the machine learning execution unit 32 (for example, PSPNet, etc.). , as the identification results, a first correct probability of the film part in the polishing mark, a second correct probability of the member part in the polishing mark, and a third A correct answer of the outer peeled part in the peeled part of the film peeled from the surface of the member. The probability, the 3rd B correct probability of the inner peeled part in the peeled part of the film peeled from the surface of the member, and the 4th correct answer probability of the surface part of the film are associated as labels for each pixel of the film image. The identification result image is generated by determining the maximum label for each of the plurality of pixels, and displaying the pixel in a manner corresponding to the determined maximum label. and output it.

In modifying the machine learning model, before determining the maximum label, the third correct probability obtained by the model is multiplied by a predetermined multiplication factor (>1.0), and the third correct probability of the multiplication result is It is compared with the correct probability, the second correct probability, and the fourth correct probability, and the maximum label is determined. More specifically, before determining the maximum label, the 3rd A correct answer probability determined by the model is multiplied by a predetermined Ath multiplier (>1.0, for example 1.2 or 1.3), The 3B correct answer probability calculated using the model is multiplied by a predetermined B multiplier (>1.0, for example 1.2 or 1.3, etc.), and the 3A correct answer probability and 3B correct answer probability of the multiplication results are The first correct probability, the second correct probability, and the fourth correct probability are compared, and the maximum label is determined. As a result, for example, the outer peeled portion DA3b1 and the inner peeled portion DA3b2 of the identification result image DPb shown in FIG. parts are corrected.

In order to express the present invention, the present invention has been adequately and fully described through embodiments with reference to the drawings in the above description, but those skilled in the art will easily be able to modify and/or improve the embodiments described above. It should be recognized that this is possible. Therefore, unless the modification or improvement made by a person skilled in the art does not depart from the scope of the claims stated in the claims, such modifications or improvements do not fall outside the scope of the claims. It is interpreted as encompassing.

D Film thickness measurement device (film thickness measurement device equipped with machine learning device)
1 Learning data acquisition section 2 Image acquisition section 3 Control processing section 4 Input section 5 Output section 6 Interface section (IF section)
7 Storage unit 31 Control unit 32 Machine learning implementation unit 33 Film thickness processing unit

Claims (9)

a learning data acquisition unit that acquires learning data;
a machine learning implementation unit that performs machine learning on a machine learning model based on the learning data acquired by the learning data acquisition unit,
The machine learning model identifies a film part in the polishing mark and a member part in the polishing mark after the polishing mark is formed on the surface of a film formed on the surface of a predetermined member by a spherical polishing method,
The learning data includes a film image obtained by capturing the film so as to include the polishing marks, and teacher data associated with each pixel of the film image,
The teacher data includes a first answer indicating that the polishing mark is a film part, and a second answer indicating that the polishing mark is a member part.
Machine learning device. The machine learning model includes a plurality of pixels associated with each pixel of the film image, and outputs an identification result image in which the film part in the polishing mark and the member part in the polishing mark are displayed in mutually different manners.
The machine learning device according to claim 1. The machine learning model further identifies a peeled portion of the film that has peeled off from the surface of the member,
The teacher data further includes a third answer indicating that the part is a peeled part of a film peeled off from the surface of the member,
The machine learning model outputs an identification result image in which the peeled part of the film peeled off from the surface of the member is displayed in a further different manner.
The machine learning device according to claim 2. a learning data acquisition step of acquiring learning data;
a machine learning implementation step of performing machine learning on a machine learning model based on the learning data acquired in the learning data acquisition step,
The machine learning model identifies a film part in the polishing mark and a member part in the polishing mark after the polishing mark is formed on the surface of a film formed on the surface of a predetermined member by a spherical polishing method,
The learning data includes a film image obtained by capturing the film so as to include the polishing marks, and teacher data associated with each pixel of the film image,
The teacher data includes a first answer indicating that the polishing mark is a film part, and a second answer indicating that the polishing mark is a member part.
Machine learning methods. A film thickness measuring device for measuring film thickness, comprising the machine learning device according to claim 2 or 3,
an image acquisition unit that acquires a target image of the film to be measured so as to include the polishing marks;
A film thickness processing unit that obtains an identification result image for the target image obtained by the image acquisition unit using a machine learning model machine learned by the machine learning implementation unit, and obtains the film thickness based on the obtained identification result image. and
Film thickness measuring device. The film thickness processing section calculates a degree of abnormality representing the degree of abnormality regarding the shape of the polishing mark based on the determined identification result image, and when the determined degree of abnormality satisfies a predetermined condition, the determined degree of abnormality is determined based on the determined identification result image. determining the film thickness based on the result image;
The film thickness measuring device according to claim 5. The film thickness processing unit further notifies an external party that the film thickness cannot be determined based on the determined identification result image when the determined degree of abnormality does not satisfy a predetermined condition.
The film thickness measuring device according to claim 6. A film thickness measuring device for measuring film thickness, comprising the machine learning device according to claim 3,
an image acquisition unit that acquires a target image of the film to be measured so as to include the polishing marks;
A film thickness processing unit that obtains an identification result image for the target image obtained by the image acquisition unit using a machine learning model machine learned by the machine learning implementation unit, and obtains the film thickness based on the obtained identification result image. and
The film thickness processing section calculates a peeling degree indicating the degree of peeling in the film to be measured based on the determined identification result image, and when the calculated peeling degree satisfies a predetermined second condition, determining the film thickness based on the determined identification result image;
Film thickness measuring device. A film measuring method for measuring film thickness, comprising the machine learning method according to claim 4,
The machine learning model includes a plurality of pixels associated with each pixel of the film image, and outputs an identification result image in which the film part in the polishing mark and the member part in the polishing mark are displayed in mutually different manners,
an image acquisition step of acquiring a target image of the film to be measured so as to include the polishing marks;
A film thickness processing step in which an identification result image for the target image obtained in the image acquisition step is obtained using a machine learning model learned by machine learning in the machine learning implementation step, and the film thickness is obtained based on the obtained identification result image. and
Film thickness measurement method.

Images