Measurement system, method of generating a learning model, and storage medium

Abstract

The present application provides a technology capable of performing a measurement process without referring to a design drawing in a case where parameter adjustment of image processing requiring skills or design of a case where it is difficult to obtain is not available. A measurement system of the present disclosure refers to a learning model generated based on supervised data generated from a sample image of a semiconductor and a sample image, generates a region segmentation image from an input image (measurement target) of a semiconductor having a predetermined structure, and performs image measurement using the region segmentation image. Here, the supervised data is an image in which a label including a structure of a semiconductor in a sample image is assigned to each pixel of an image, and the learning model includes a parameter for estimating the supervised data from a sample image (see FIG. 1).

Description

TECHNICAL FIELD

[0001] The present disclosure relates to a measurement system, a method of generating a learning model used when performing image measurement of a semiconductor including a predetermined structure, and a storage medium storing a program for causing a computer to execute a process of generating a learning model used when performing image measurement of a semiconductor including a predetermined structure. BACKGROUND

[0002] With the miniaturization of patterns manufactured through semiconductor processes in recent years, there is a demand for an increase in precision of overlay, that is, overlay, of patterns between multiple layers throughout an exposure device. In addition, it is expected that the importance of measuring overlay with high precision and feeding back to an exposure device will become higher in the future.

[0003] Regarding overlay measurement, for example, a technique is disclosed in Patent Literature 1 in which a plurality of brightness regions divided by a brightness boundary on an input image are extracted by image processing, and overlay measurement is performed based on positional relationships of centers of gravity of the brightness regions. In addition, a technique is disclosed in Patent Literature 2 in which a design image such as a CAD image or a prediction image of an input image inferred from a design image is referred to, and the input image is regionally divided in units of pixels, and overlay measurement is performed based on positional relationships of centers of gravity of regions after the regional division. Furthermore, a technique is disclosed in Patent Literature 3 in which a sample of an image of an overlay measurement target is collected in advance, a machine learning model that infers an overlay amount (an amount of shift of a position of a structure of a semiconductor that is a target of overlay measurement) from an image is learned, and the overlay amount is measured from an input image with reference to the machine learning model.

[0004] PRIOR ART DOCUMENTS

[0005] PATENT LITERATURE

[0006] Patent Literature 1: WO2017 / 130365

[0007] Patent Literature 2: Japanese Patent Application Laid-Open No. 2018-522238

[0008] Patent Literature 3: Japanese Patent Application Laid-Open No. 2010-538474 SUMMARY

[0009] PROBLEMS TO BE SOLVED BY THE INVENTION

[0010] However, in the technique disclosed in Patent Literature 1, it is necessary to manually adjust parameters in image processing according to an input image, and skill is also required in the adjustment of the parameters, and thus there is a problem in that an operator who performs overlay measurement is limited to an expert.

[0011] In addition, in the technology disclosed in Patent Literature 2, there is a problem that the technology cannot be used in a case where the design image cannot be obtained due to a reason such as non-disclosure.

[0012] In addition, in the technology disclosed in Patent Literature 3, it is difficult to perform cause analysis in a case where an unintended amount of overlap is measured from the input image because the process of measuring the amount of overlap from the input image cannot be visually confirmed.

[0013] The present disclosure was completed in view of such a situation, and proposes a technology capable of performing a measurement process without reference to a design drawing in a case where the design drawing is difficult to obtain, and without parameter adjustment of image processing requiring skills.

[0014] Means for solving the problem

[0015] To achieve the above-described problem, an embodiment of the present disclosure proposes a method of generating a learning model used when performing image measurement of a semiconductor including a predetermined structure, the method including: generating, by at least one processor, supervised data by assigning a label of a structure including at least one measurement target to a region segmentation image obtained from a sample image of the semiconductor; and generating, by at least one processor, a learning model based on a network structure composed of a plurality of layers, using the region segmentation image of the sample image and the supervised data, the learning model including a parameter for estimating the supervised data from the sample image.

[0016] Further features related to the present disclosure will become apparent from the description of the specification, the drawings, and the appended claims. In addition, the mode of the present disclosure is realized by elements, combinations of various elements, and the following detailed description and the appended claims.

[0017] The description of the specification is merely a typical example, and does not limit the scope of the claims or the application of the present disclosure in any way.

[0018] Effects of the Invention

[0019] According to the present disclosure, it is possible to perform a measurement process without reference to parameter adjustment of image processing requiring skills or a design drawing that is sometimes difficult to obtain. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a diagram showing an example of a functional structure from the production of supervised data to overlap measurement in Embodiment 1.

[0021] Figure 2 is a diagram showing an example of a sample image.

[0022] Figure 3 is a diagram showing an example of a region segmentation image.

[0023] Figure 4 is a diagram showing a structural example of a user interface for producing supervision data 14 from a sample image 13 provided by the supervision production section 1.

[0024] Figure 5 is a diagram showing an example of a deep neural network structure 179 in the learning model 11.

[0025] Figure 6 is a diagram for supplementarily explaining the geometric relationship of the neural network structure 179 and an image.

[0026] Figure 7 is a diagram showing an example of an image 50 as an input image 12.

[0027] Figure 8 is a diagram showing an output example of the grouping section 4.

[0028] Figure 9 is a flowchart for explaining the detailed contents of the grouping processing performed by the grouping section 4.

[0029] Figure 10 is a diagram (table) showing label examples of the first measurement object and the second measurement object of the overlay measurement.

[0030] Figure 11 is a flowchart for explaining the detailed contents of the overlay measurement processing performed by the overlay measurement section 5.

[0031] Figure 12 is a diagram showing a structural example of the template data 85.

[0032] Figure 13 is a diagram (table) showing other examples of labels of the first measurement object and the second measurement object of the overlay measurement.

[0033] Figure 14 is a diagram showing a functional structural example from the supervision data production to the overlay measurement of Embodiment 2.

[0034] Figure 15 is a diagram showing examples of the groups 30a and 30b of the sample image and the region segmentation images (labeled supervision data) 40a and 40b.

[0035] Figure 16 is a diagram showing a functional structural example from the supervision data production to the overlay measurement of Embodiment 3.

[0036] Figure 17 is a diagram showing a structural example of the sample image 213.

[0037] Figure 18 is a diagram showing a structural example of the supervision data 214.

[0038] Figure 19 is a flowchart for explaining the supervised production processing by the supervised data production section 201.

[0039] Figure 20 is a graph for explaining an example of the correction based on the statistical processing of step S204.

[0040] Figure 21 is a graph for explaining another example of the correction based on the statistical processing of step S204.

[0041] Figure 22 is a graph showing an example of the functional structure from the supervised data production to the overlay measurement of Embodiment 4.

[0042] Figure 23 is a graph showing an example of the position information image 340.

[0043] Figure 24 is a graph showing an example of the position information image 360.

[0044] Figure 25 is a flowchart for explaining the overlay measurement processing by the overlay measurement section 305.

[0045] Figure 26 is a graph showing an example of the functional structure from the supervised data production to the overlay measurement of Embodiment 5.

[0046] Figure 27 is a graph showing an example of the image 430 estimated from the region division image 440.

[0047] Figure 28 is a graph showing an example in a case where the layout of the small region 71q is changed from the region division image 440.

[0048] Figure 29 is a graph showing an example in a case where the layout of the small region 71q is changed from the region division image 440 and the occlusion is generated (an example in a case where it is determined in the order of the labels 43, 41, and 42 that it is located on the front side).

[0049] Figure 30 is a graph showing an example of the functional structure from the supervised data production to the image measurement check of Embodiment 6.

[0050] Figure 31 is a graph showing an example of the outline structure of the measurement system 310 of the present embodiment (common to each embodiment). DETAILED DESCRIPTION

[0051] The present embodiment and each example relate to a measurement system that performs image measurement in a semiconductor having a predetermined structure (for example, a multilayer structure), and more specifically, to a measurement system that measures an amount of shift between layers in the case where the semiconductor is a multilayer structure. However, the technology of the present disclosure is not limited to application to overlay measurement, but can be widely applied to general image measurement.

[0052] Hereinafter, the present embodiment and examples of the present disclosure will be described with reference to the accompanying drawings. In the drawings, the same elements are sometimes denoted by the same reference numerals. Further, the drawings show specific embodiments and examples that follow the principles of the present disclosure, but these are for understanding the present disclosure, and are by no means to be interpreted limitatively.

[0053] In the present embodiment, the present disclosure has been described in sufficient detail by those skilled in the art for implementation of the present disclosure, but it is to be understood that other installation methods are possible, as long as the constitution, structural changes, and replacement of various elements are made without departing from the scope and spirit of the technical idea of the present disclosure. Therefore, the following description cannot be interpreted limitatively.

[0054] Also, as described later, the present embodiment of the present disclosure can be installed by software running on a general-purpose computer, or by dedicated hardware or a combination of software and hardware.

[0055] Further, in the following description, each information of the present disclosure is described by a "table" form, but the information can not necessarily be represented by a data structure based on a table, but can be represented by a data structure such as a list, a DB, a queue, or the like, or other than the above. Therefore, in order to represent independence from the data structure, "table", "list", "DB", "queue", and the like are sometimes simply referred to as "information".

[0056] In addition, in describing the content of each information, "number", "identification information", "identifier", "name", "name", "ID", and the like can be used, and these can be replaced with each other.

[0057] (1) Embodiment

[0058] This embodiment (and Embodiments 1 to 6) relates to, for example, a measurement system that performs image measurement of a semiconductor including a predetermined structure (e.g., a multilayer structure). The measurement system generates a region segmentation image from an input image (measurement target) of a semiconductor having a predetermined structure, with reference to a learning model generated based on supervised data generated from a sample image of a semiconductor and the sample image, using the region segmentation image to perform image measurement. Here, the supervised data is an image in which each pixel of the image is assigned a label including a structure of a semiconductor in the sample image, and the learning model includes parameters for estimating the supervised data from the sample image. By using the learning model, estimation from the sample image to the supervised data is applied to the input image, so that the measurement process can be performed without using design data of the input image.

[0059] Figure 31 is a diagram showing an example of a schematic configuration of a measurement system 310 of this embodiment (common to each embodiment). The measurement system 310 corresponds to, for example, an overlay measurement system that performs overlay measurement, a dimension measurement system that measures the dimensions of a profile extraction, a hole shape, and the like in a semiconductor image, a defect pattern detection system that detects a defect pattern and the like, a pattern matching system that searches for a matching position of an estimated design pattern and an actual design pattern of a design pattern, and the like.

[0060] The measurement system 310 includes, for example, a host computer 191 including a host processor 190, an input / output device 192 that inputs instructions and data to the host computer 191 and outputs an operation result, an electron microscope or a server computer 193 that stores an image of the electron microscope (hereinafter referred to as an electron microscope or the like) that supplies an image of a measurement target, a first sub-computer 191a including a first sub-processor 190a, and a second sub-computer 191b including a second sub-processor 190b, each of which is connected through a network (e.g., LAN or the like). In addition, in the measurement system 310, the first sub-computer 191a and the second sub-computer 191b are provided, but all operations can be performed in the host computer, and one or more sub-computers for assisting the host computer 191 can be provided. Figure 31

[0061] The host computer 191 performs the following Figure 1 14 ​​, 16, 22, 26, and 30, and the learning model creation process (processes corresponding to the supervision creation sections and the learning sections in the respective drawings). In addition, the estimation process and the measurement process (processes corresponding to the region division sections, the grouping sections, and the measurement sections (overlap measurement sections) in the respective drawings) can also be processed by the main computer 191 or the first and second sub-computers 191a and 191b in a distributed manner. Furthermore, in the first and second sub-computers 191a and 191b, it is possible to be configured to execute only the estimation process and the measurement process, without executing the supervision creation process and the learning model creation process. In addition, in the case where a plurality of sub-computers are provided (for example, the first and second sub-computers 191a and 191b), the estimation process and the measurement process can also be distributed among the sub-computers.

[0062] The electron microscope 193 or the like, for example, acquires (photographs) a semiconductor pattern image formed on a wafer, and provides it to the main computer 191, the sub-computers 191a and 191b. In the case where the electron microscope 193 or the like is a server computer, the server computer stores a semiconductor pattern image photographed by the electron microscope in a storage device (for example, a hard disk drive (HDD)), and in response to an instruction from the main computer 191, provides the main computer 191 or the like with an image corresponding to the instruction.

[0063] Hereinafter, each embodiment of the process executed in the measurement system 310 will be described in detail. In addition, in Embodiments 1 to 5, as a measurement example, an overlap measurement is exemplified. In Embodiment 6, it is explicitly shown that the technology of the present disclosure can be applied to all measurement processes.

[0064] (2) Embodiment 1

[0065] Figure 1 is a diagram showing an example of the functional structure from the supervision data creation to the overlap measurement of Embodiment 1. The functions of the supervision creation section 1 and the learning section 2 are realized by the main processor 190 of the main computer 191 reading in the corresponding respective programs from a storage section not shown. In addition, the functions of the region division section 3, the grouping section 4, and the overlap measurement section 5 are realized by the main processor 190 of the main computer 191 or the sub-processors 190a and 190b of the sub-computers 191a and 191b reading in the corresponding respective programs from a storage section not shown.

[0066] <Outline of Functional Structure>

[0067] First, the functions of the supervision creation section 1 and the learning section 2 will be described. Figure 1The illustrated outline of the functional configuration from the supervised data production to the overlay measurement is explained. The sample images 13 are images of a sample collected in advance as images of the sample which are the object of the overlay measurement. The supervised data 14 is data in which a label of a structure in a semiconductor which is the object of the measurement of the overlay measurement is assigned to each pixel in the image for each of the sample images 13, and a region segmentation image is obtained.

[0068] The supervised production section 1 produces the supervised data 14 from the sample images 13, and also provides a user interface for producing the supervised data 14. The learning model 11 is a parameter such as a coefficient of a machine learning model which obtains a region segmentation image from an image such as a sample image. The learning section 2 calculates the learning model 11 which estimates a region segmentation image closest to the supervised data 14 as possible when the sample images 13 are input.

[0069] The input image 12 is an image which is the object of the measurement at the time of the overlay measurement. The region segmentation section 3 refers to the learning model 11, and estimates a region segmentation image from the input image 12. The grouping section 4 groups the object of the measurement of the overlay measurement in the region segmentation image in units of small regions. The overlay measurement section 5 performs the overlay measurement from the position of the small region after the grouping by the grouping section 4.

[0070] Each of the above functions of the supervised production section 1, the learning section 2, the region segmentation section 3, the grouping section 4, and the overlay measurement section 5 can be realized by signal processing on an arbitrary computer.

[0071] <Details of each functional configuration>

[0072] The details of each functional configuration of the embodiment 1 are explained below. The sample images 13 are images photographed before the overlay measurement is applied, and are sample images of the appearance of the image being close to a specimen of a semiconductor which is the object of the measurement or a specimen of a semiconductor which is the object of the measurement. The sample images 13 can be collected by an electron microscope to which the overlay measurement is applied or an electron microscope which photographs an image close to the quality of the image.

[0073] (i) Figure 2 is a drawing showing an example of a sample image. Figure 2 The image 30 in (i) shows, for example, a part of a sample image. The structure in a semiconductor which is the object of the overlay is included in the image 30. The sample image is constituted by one or more images which are the same as the image 30.

[0074] (ii) The supervised data 14 is constituted by region segmentation images obtained from each of the images 30 in the sample images 13. Figure 3 is a drawing showing an example of a region segmentation image. In Figure 3In the region division image 40, in each pixel in the image 30, the label 41, the label 42, and the label 43 indicating the first measurement object, the second measurement object, and the background are assigned. The first measurement object and the second measurement object corresponding to the labels 41 and 42 are overlapping measurement objects, for example, corresponding to a via, a trench, and other structures in a semiconductor. How to determine the structures in the semiconductor is determined in advance according to the application of the overlapping measurement. In addition, an additional label other than the labels 41, 42, and 43 can be assigned to the region division image 40. For example, in the case of a semiconductor wafer, a label indicating a wafer edge can be assigned to the region division image 40. Figure 3 In the region division image 40, a label 49 corresponding to an invalid region excluded by the learning unit 2 is assigned.

[0075] (iii) Figure 4 is a diagram indicating a structure example of a user interface provided by the supervision production unit 1 for producing supervision data 14 from a sample image 13. In the Figure 4 In the user interface, for example, a main screen 90, an input screen 91, an input selection region 92, and an input pen 93 are included. In the input screen 91, in a state where the operator has selected any one item from the single selection buttons of the input selection 92, by operating the input pen 93 on the image 30 (for example, smearing the region designated by the label), a label is assigned to the region operated by the input pen. The input selection 92 sequentially selects the labels 41, 42, and 43 according to the selection of the item of the single selection button of the first measurement object, the second measurement object, and the background. For the position input by the input pen 93, a predetermined color or tone is assigned according to the label selected by the input selection 92.

[0076] The input selection 92 can also include an item capable of selecting an additional label such as the invalid region 49. In addition, the user interface in the input selection 92 is one example, and the string added to the single selection button, the user interface other than the single selection button can be changed. The selection region 94 indicates an example in which the labels 41, 42, and 43 are assigned to a part of the image 30. Then, the region division image 40 is produced by assigning the same label to the entire region of the image 30.

[0077] By using the above-described supervision production unit 1, the operator can produce the supervision data 14 from the sample image 13 with a simple operation without performing the skill-required parameter adjustment, the reference of the design drawing.

[0078] (iv) Figure 5 is a diagram indicating an example of a deep neural network structure 179 in the learning model 11. The deep neural network structure 179 can be constituted by, for example, an input layer 170, an output layer 171, a plurality of intermediate layers 172, 173, and 174.

[0079] The image 30 is stored in the input layer 170. From the input layer 170 to the intermediate layer 172, and from the intermediate layer 172 to the intermediate layer 173, the data within the layers is aggregated by image reduction through a convolution operation based on a predetermined coefficient filter. On the other hand, from the intermediate layer 173 to the intermediate layer 174, and from the intermediate layer 174 to the output layer 171, the data within the layers is spread by image expansion through a convolution operation based on a predetermined coefficient filter. In general, such a network structure is called a convolutional neural network. The data within the output layer (final layer) 171 represents the likelihood of each label in the region segmentation image 40. By assigning the label with the greatest likelihood to each pixel, the region segmentation image 40 can be obtained. The learning model 11 corresponds to the coefficients of the filters in the intermediate layers. In addition, the deep neural network structure 179 is an example of the network structure in the learning model 11, and the number of intermediate layers such as the intermediate layers 172 and 174 is not limited to three. In addition, additional structures such as a bypass structure as connecting the intermediate layers 172 and 174, or additional operations in addition to the filter operation can be taken. In the case where additional parameters are added as the additional structures, the additional operations, and the like are added to the network structure in the learning model 11, the additional parameters are also added to the learning model 11. Figure 5

[0080] (v) Figure 6 is a diagram for supplementarily explaining the geometric relationship between the neural network structure 179 and the image. In Figure 6 , attention is paid to the pixel 175 in the image 30 and the pixel 177 in the region segmentation image 40 at the same coordinates as the pixel 175. At this time, in the input layer 170, the vicinity of a predetermined range of the pixel 175 referred to as a receptive field 176 (the receptive field refers to the range of the image related to the determination of the label, and the size thereof is determined in accordance with the neural network structure 179). Then, the data of the intermediate layers 172, 173, and 174 are obtained by a convolution operation, image expansion, and image reduction from the data of the receptive field 176 in the input layer 170, and the label of the pixel 177 is determined in the output layer 171. In addition, in fact, the labels of the pixels in the region segmentation image 40 are efficiently obtained in a parallel operation manner in accordance with each pixel in the image 30.

[0081] ​The learning unit 2 calculates the parameters of the learning model 11 for estimating the region segmentation image 40 when the image 30 is provided, with reference to the sample images 13 and the supervision data 14. Specifically, the learning unit 2 estimates the region segmentation image 40 from the image 30 in the sample images 13, for example, compares the estimated region segmentation image 40 with the region segmentation image 40 corresponding to the image 30 in the supervision data 14, and calculates the learning model 11 in which the difference between the region segmentation images 40 is optimal and minimal. For example, the difference between the region segmentation images 40 is set to the number of pixels in which the labels are different in all the pixels, and the partial differential coefficients of each element in the learning model 11 (neural network structure 179) with respect to the number of pixels are calculated. Also, there is a method in which the value obtained by adding the negative of the predetermined coefficient to the partial differential coefficient with respect to each element in the learning model 11 is sequentially calculated in each image 30 in the sample images 13 and is updated (updated little by little in such a way as to reduce the number of pixels). However, the method is not limited to this. Here, in the case where the corresponding region segmentation image 40 in the supervision data 14 includes the label 49 (invalid region), this part is excluded from the total of the number of pixels in which the labels are different.

[0082] Also, the learning unit 2 can apply random noise to the image 30 in the sample images 13 or increase the number of synthetic images in which the synthetic processing such as geometric transformation of enlargement, reduction, and left-right reversal, up-down reversal, and the like is performed when generating the learning model 11. By increasing the number of synthetic images, the learning unit 2 can calculate the learning model 11 from a more abundant number of images 30.

[0083] (vi) The input image 12 is an image photographed at the time of overlay measurement. Figure 7 is a diagram showing an image 50 as an example of the input image 12. Figure 7 The region 59 in the image 50 is a region of the same size as the image 30 in the sample images 13, and includes the structure in the semiconductor that is the object of overlay measurement. Outside the region 59 of the image 50, the structure in the semiconductor that is the object of overlay measurement is periodically reflected as in the region 59.

[0084] The region segmentation unit 3 estimates the region segmentation image 40 from the input image 12 with reference to the learning model 11 Figure 7The region segmentation image 60 is shown in the image. Region 69 in region segmentation image 60 is a region of the same extent as region 59. Here, estimation means inputting data into the network structure of learning model 11, performing calculations such as convolution operations on each layer of the network structure (input layer 170, output layer 171, intermediate layers 172, 173, and 174), and obtaining the calculation results. For each pixel in region 69 in region segmentation image 60, a label of any one of 41, 42, and 43 is assigned, similar to that in region segmentation image 40. If the image of region 59 is similar to any of the images 30 in sample image 13, learning model 11 has the characteristic of estimating the region segmentation image 40 corresponding to image 30 during the process of learning model 11 being obtained in learning unit 2. Therefore, the correct label can be assigned to region 69. The same applies to the remaining regions in region segmentation image 60 other than region 69. Here, the scale (benchmark) of "similarity" means that the images inside are similar within a unit of receptive field 176 (a small region in image 30). In particular, in overlapping measurements, the structure in the semiconductor being measured is periodic. Therefore, even if the size of the image in sample image 13 is smaller than that of region segmentation image 60, it is possible to expect that image 30, which satisfies similar conditions to region segmentation image 60, exists in sample image 13 and is assigned the correct label.

[0085] (vii) Group 4 executes Figure 9 The process shown in the flowchart groups overlapping measurement objects in the region segmentation image 60 into small regions. Details of the grouping process will be described later.

[0086] (viii) Overlap measurement unit 5 performs... Figure 11 The process shown in the flowchart involves overlapping measurements based on the grouped images 70. Details of the overlapping measurement process will be described later.

[0087] <Details on group processing>

[0088] Figure 9 This is a flowchart illustrating the detailed process of grouping performed by grouping unit 4. The grouping process is carried out according to... Figure 10 The items of the measurement objects specified in Table 80 are grouped according to each associated small region based on the labels assigned to the predetermined regions of the region segmentation image 60.

[0089] In steps S1 to S4, grouping unit 4 targets... Figure 10 For each item of the measurement object specified in Table 80 (first measurement object and second measurement object), the processing of steps S2 and S3 is performed repeatedly.

[0090] In step S2, the grouping section 4 obtains a 2-value image in which the pixels of the object label in the region division image 60 are set to 1 and the others are set to 0. According to Figure 10 For the first measurement object, the object label is label 41, and for the second measurement object, the object label is label 41 and label 42. A plurality of labels are assigned to the second measurement object because the structure corresponding to label 41 in the image 50 is located on the near side compared to the structure corresponding to label 42, and therefore a part of label 42 is occluded by label 41 in the region division image 60.

[0091] In step S3, the grouping section 4 groups in small regions based on the 2-value image obtained in step S2. As a grouping method in small regions, a method called labeling that groups in units of connected regions constituted by the pixels of value 1 in the 2-value image can be applied. However, labeling is not limited, and other methods that can group in small regions can be applied.

[0092] The grouping processing is executed in the above-described manner. Here, Figure 8 is a diagram that shows an example of the grouping image output by the grouping section 4. In Figure 8 , the image 70 shows an example of the grouping image obtained by the grouping processing Figure 9 ). The small regions 71a, 71b, 71c, and 71d are small regions corresponding to the first measurement object in Table 80. In addition, the small regions 72a, 72b, 72c, and 72d are small regions corresponding to the second measurement object in Table 80. The region 79 in the grouping image 70 is a region of the same range as the region 59 (refer to Figure 7 ). For example, in the case of a semiconductor pattern, in addition to the region 59 in the region division image 60 (refer to Figure 7 ), the same small region as the region 59 repeatedly appears. That is, in the case of a semiconductor pattern, similar patterns repeatedly appear, and therefore it is also possible to not generate the supervision data for all the pixels in the image. For example, if the supervision data of the receptive field 176 (refer to Figure 6 ) as a small region is obtained, it is possible to construct a learning model that can estimate the image as a whole of a region larger than the receptive field 176, and it is possible to save the time of the worker (operator).

[0093] <Details of the overlay measurement processing>

[0094] Figure 11 is a flowchart for explaining the details of the overlay measurement processing executed by the overlay measurement section 5.

[0095] (i) Step S11

[0096] The superimposition measurement section 5 performs alignment of the template with respect to the first measurement object. Here, the template is the X coordinate and Y coordinate of each element of the first measurement object, and is data prepared in advance by use of the superimposition measurement. Figure 12 is a diagram showing an example of the structure of the template data 85. In Figure 12 , the template data 85 is composed of the X coordinate and Y coordinate of each element from the first to the Nth. The template data 85 is calculated from the X coordinate and Y coordinate of the center of gravity of the small regions 71a, 71b, 71c, and 71d, etc. in the group image 70 that becomes a representative example. Alternatively, the template data 85 can be calculated from a design drawing of a semiconductor, etc.

[0097] As a reference for alignment, there can be mentioned a method in which the center of gravity of all points in the template is aligned with the center of gravity of the small region 71a, etc. of the first measurement object, but the method is not limited to this.

[0098] (ii) Step S12

[0099] The superimposition measurement section 5 selects the small region 71a, etc. corresponding to each element in the template data 85 after alignment. The reference for selection can be set to the element whose center of gravity is closest to the template data 85 in the small region 71a, etc., but is not limited to this.

[0100] (iii) Steps S13 to S18

[0101] The superimposition measurement section 5 repeatedly performs the processing of steps S14 to S17 for each small region selected in step S12. In the following description, a case in which the small region 71a is the object will be described.

[0102] (iii-1) Step S14

[0103] The superimposition measurement section 5 calculates the position 1 that is a representative position of the first measurement object. The position 1 is composed of two elements of the X coordinate X1 and the Y coordinate Y1. The position 1 is calculated from the X coordinate and Y coordinate of the center of gravity position of the small region 71a, etc.

[0104] (iii-2) Step S15

[0105] The superimposition measurement section 5 selects the small region 71a and the small region for measuring the amount of superimposition in the small region of the second measurement object. The reference for selection can apply the reference for selecting the small region whose position of the center of gravity is closest. In the case of Figure 8 , for example, the small region 72a is selected.

[0106] (iii-3) Step S16

[0107] The overlap measurement section 5 calculates the position 2 of the representative position of the small region (for example, the small region 72a) selected as the second measurement target in step S15 in the same order as in step S14. The position 2 is composed of two elements of an X coordinate X2 and a Y coordinate Y2.

[0108] (iii-4) Step S17

[0109] The overlap measurement section 5 calculates the displacement amounts Dx and Dy of the X coordinate and the Y coordinate from the position 2 and the position 1 by the following equations 1 and 2.

[0110] Dx = X2 - X1... (Equation 1)

[0111] Dy = Y2 - Y1... (Equation 2)

[0112] (iv) Step S19

[0113] The overlap measurement section 5 calculates a statistical amount of the displacement amounts Dx and Dy calculated based on the equations 1 and 2. In calculating the statistical amount, an additive average can be applied, but is not limited thereto, and can be a multiplicative average, a median value. The overlap measurement section 5 sets the statistical amount of the displacement amounts calculated in step 19 as the overlap amount of the image 50.

[0114] Technical Effects of Embodiment 1

[0115] According to Embodiment 1, the process in which the learning section 2 uses the supervision data 14 produced in the supervision production section 1 based on the sample image 13 to calculate the learning model 11 is set in advance. Then, the region segmentation section 3 uses the region segmentation image 60 calculated based on the input image 12 by referring to the learning model 11, and the overlap amount can be measured by the grouping section 4 and the overlap measurement section 5. Thus, unlike Patent Literature 1, there is no need to adjust parameters requiring skills, and unlike Patent Literature 2, there is no need to input design data of the input image 12, and correct overlap measurement can be performed by estimating the region segmentation image 60. In addition, in order to be able to visualize intermediate processing data such as the region segmentation image 60 and the grouping image 70, unlike Patent Literature 3, in the case where an undesirable overlap amount is measured, by displaying the intermediate processing data on a screen, the cause can be grasped. That is, since the input image 12 is region-segmented by referring to the learning model 11 in the region segmentation section 3, the above-described first and second problems are solved. In addition, since the region segmentation image is data that can be visualized, the operator can easily confirm. Thus, the third problem is also solved.

[0116] In addition, according to Embodiment 1, the learning model 11 estimates the region segmentation image 40 in units of the receptive field 176. In addition, in the overlap measurement processing, generally, the input image 12 (the image 60: refer to FIG. 6) is displayed on a screen, and the region segmentation image 40 is displayed on the same screen. Thus, the operator can easily confirm the region segmentation image 40. Figure 7This reflects the periodic structure in the semiconductor. Therefore, in the supervision production unit 1, the user interface of the main screen 90 can be used to assign the size of image 30 in sample image 13 of supervision data 14 to be smaller than that of image 60 in input image 12. This reduces the time spent by the operator when assigning supervision data 14.

[0117] <Modification of Example 1>

[0118] In Embodiment 1 described above, the constituent elements can be changed. For example, in addition to the neural network structure described above, the learning model 11 can also apply any machine learning model that infers region segmentation of image 40 based on image 30 in units of receptive field 176. For example, it could also be a linear discriminator that determines the label of pixel 177 based on all pixels of receptive field 176 of image 30.

[0119] In step S16, the X2 and Y2 coordinates of position 2 can also be calculated based on the small regions of different labels. For example, when referring to... Figure 13 In the case shown in Table 81, in step S16, the overlap measurement unit 5 calculates the coordinate X2 based on the small areas of label 41 and label 42, and calculates the coordinate X2 based on label 44 (in Figure 3 In the example of region segmentation image 40, the coordinates Y2 are calculated from the small regions (labels assigned to predetermined structures in semiconductors not shown). Because the longitudinal contour (contrast) of the structure in the semiconductor corresponding to label 42 is not sharp, it is sometimes impossible to calculate the correct Y2 from the small regions corresponding to labels 41 and 42. In such cases, by appending the coordinates assigned to the small regions in the semiconductor... Figure 3 Label 44 with a structure exhibiting strong vertical contrast (horizontal stripes) (e.g., Example 2) Figure 15 The correct coordinate Y2 can be obtained by using label 42b). That is, when the longitudinal contrast of the structure in the semiconductor corresponding to label 42 is not sharp, the coordinate Y2 can be correctly obtained by changing the overlapping measurement object to the newly assigned label 44.

[0120] Furthermore, in step S16, referring to Table 81, compared to referring to Table 80 in step S16, the supervision and production unit 1 adds label 44 to the label allocation object. Then, the region segmentation unit 3 adds label 44 to the estimated object, and the grouping unit 4 adds label 44 to the grouping object. Similarly, in step S14, the X1 and Y1 coordinates of position 1 can also be calculated based on the small regions of different labels.

[0121] The processing of step Sll is generally performed as an overlay measurement, but can also be excluded. In this case, all of the small regions 71a and the like in the group image 70 are selected in step S12. Alternatively, in step S12, additional selection of small regions 71a and the like in which the area is small to the extent of noise and the like can also be performed.

[0122] (3) Embodiment 2

[0123] <Functional Configuration Example>

[0124] Figure 14 is a diagram showing a functional configuration example of the production of supervision data to overlay measurement in Embodiment 2. In Figure 14 , the sample images 113 are, for example, a group of images obtained by changing the imaging conditions and imaging the same site in a semiconductor wafer multiple times. Here, the imaging conditions refer to the acceleration voltage of the electron microscope, the imaging of the reflected electron image, the imaging of the secondary electron image, the synthesis ratio when synthesizing both to obtain a composite image, and the like, but are not limited thereto.

[0125] Figure 15 is a diagram showing an example of the group of sample images 30a and 30b and the group of region segmentation images (supervision data to which labels are assigned) 40a and 40b. The image 30a has a horizontally long structure within the region 31a, but the upper and lower profiles (vertical contrast) are not sharp. On the other hand, the image 30b contains an image of the region 31b which is the same site as the region 31a. In the image 30b, the upper and lower profiles (vertical contrast) become a sharp structure.

[0126] In the overlay measurement processing, a multilayer semiconductor is taken as an imaging object, and from the electron microscope observation, the imaging conditions in which each layer of the multilayer is most sharp at a different depth are different. Therefore, there are cases where a sharp image can be obtained from all structures in the image by changing the imaging conditions and imaging multiple times. The supervision production unit 1 assigns labels to the group of images in the sample images 113, produces supervision data 114, and provides a user interface for this. Figure 15 The supervision data 40a and 40b in are an example of the supervision data 114, and are a group of region segmentation images to which the labels 41a, 42a, and 43a and the labels 42b and 42c are assigned in the image 30a and the image 30b, respectively.

[0127] The learning unit 102 calculates the learning model 111 based on the sample image 113, the supervision data 114, the groups 30a and 30b of images, and the groups 40a and 40b of region segmentation images. The learning model 111 includes a neural network structure for estimating the region segmentation image 40a from the image 30a and a neural network structure for estimating the region segmentation image 40b from the image 30b. These neural network structures can be two independent neural network structures, or can be neural network structures that share (share) a part of an intermediate layer 173 and the like.

[0128] The input image 112 is a group of images captured under the same or similar imaging conditions as the sample image 113. The region segmentation unit 103 outputs a group of region segmentation images having the same structure as the groups 40a and 40b of region segmentation images from the input image 112. The grouping unit 104 performs an operation of extracting small regions to which three types of labels, the label 41a, the labels 41a and 42a, and the label 42b, are respectively assigned, from the group of region segmentation images output by the region segmentation unit 103.

[0129] The overlap measurement unit 105 measures the amount of overlap by calculating the X coordinate (X1) of the position 1 in step S14 shown in FIG. 8, the X coordinate (X2) of the position 2 in step S16 calculated from the labels 41a and 42a, and the Y coordinate (Y2) calculated from the label 42b. Figure 11 The overlap measurement unit 105 measures the amount of overlap by calculating the X coordinate (X1) of the position 1 in step S14 shown in FIG. 8, the X coordinate (X2) of the position 2 in step S16 calculated from the labels 41a and 42a, and the Y coordinate (Y2) calculated from the label 42b.

[0130] <Effects of Embodiment 2>

[0131] In Embodiment 2, according to the structure described above, even in a case where correct overlap measurement is difficult under a single imaging condition, a group of images is used for the same part under a plurality of imaging conditions. Thus, the labels estimated by the region segmentation unit 103 become correct, and the overlap measurement unit 105 can perform correct overlap measurement. Also, since the supervision preparation unit 101 assigns labels to distinct parts in the sample image 113, the supervision preparation unit 101 can correctly assign labels to small regions to prepare the supervision data 114.

[0132] (4) Embodiment 3

[0133] Embodiment 3 discloses a technique in which the object to which labels are assigned is reduced to a part of a sample image, and the amount of work of a worker to reduce the label assignment work of supervision data is reduced.

[0134] <Outline of Functional Structure>

[0135] Figure 16 is a diagram showing an example of a functional structure from supervision data preparation to overlap measurement in Embodiment 3. First, an outline of the example of the functional structure in Embodiment 3 is described.

[0136] The supervised production section 201 learns an intermediate learning model from the region segmentation image group 241 in which the labels are assigned to the partial set of the sample images 213, that is, the image group 231 (in which the labels are assigned to each of the images 30 in the image group 231).

[0137] Next, the supervised production section 201 refers to the intermediate learning model, estimates the region segmentation image group 243 from the remaining sample images 233, and performs a process of correcting the region segmentation image group 243 or provides a user interface for correction. The functions of the learning section 2, the region segmentation section 3, the grouping section 4, and the overlap measurement section 5, the learning model 11, and the data of the input images 12 are the same as in Embodiment 1, and thus the description thereof is omitted.

[0138] <Details of the supervised data production process>

[0139] Reference Figures 17 to 19 The details of the supervised data production process will be described. Figure 17 is a diagram showing an example of the structure of the sample images 213. Figure 18 is a diagram showing an example of the structure of the supervised data 214. Figure 19 is a flowchart for explaining the supervised data production process of the supervised production section 201.

[0140] (i) Step S201

[0141] As shown in Figure 17 , the sample images 213 are divided in advance into the partial set of the image group 231 and the image group 233.

[0142] In step S201, the supervised production section 201 provides the user interface (main screen 90: see Figure 18 ) for producing the region segmentation image group 241 shown in Figure 4 to the operator, and assigns the labels to the image group 231 in response to the input of the operator.

[0143] (ii) Step S202

[0144] The supervised production section 201 obtains the intermediate learning model (the learning model for generating the region segmentation image group 241 from the image group 231) from the image group 231 and the region segmentation image group 241 in the same step as the learning section 2.

[0145] (iii) Step S203

[0146] The supervision production section 201, similarly to the region division section 3, refers to the intermediate learning model calculated in step S202, and performs estimation in the region division image group 243 from the image group 233 (to be exact, region division images 40 are respectively estimated from the images 30 in the image group 233, and thus the region division image group 243 is calculated).

[0147] (iv) Step S204

[0148] The image group 231, which is a partial set of sample images, is difficult to completely encompass all the images contained in the image group 233, and thus in most cases contains erroneous labels in the region division image group 243. Therefore, the supervision production section 201 performs correction of the labels in the region division image group 243 by statistical processing. As the correction based on statistical processing, for example, correction of the most frequent value of the labels within the partial set in the region division image group 243 obtained by repeatedly photographing the same photographing site within the semiconductor chip can be performed.

[0149] Here, Figure 17 The images 32a, 32b to 32m (m is an arbitrary number) in the image group 232, which is a partial set within the image group 233, are images obtained by repeatedly photographing the same site within the semiconductor wafer. At this time, in step S204, the region division images 42a, 42b to 42m within the region division image group 242 estimated from the image group 232 are all assigned the most frequent value of the labels by the correction of taking the most frequent value. In addition, when the same coordinates of the pixels 43a, 43b to 43m within the region division images 42a, 42b to 42m are focused on, the label that appears the highest frequency in each pixel before correction is assigned after correction by the correction of taking the most frequent value. By the correction of taking the most frequent value, it is possible to correct cases where the labels of the region division images 40 change due to the quality of the images 30 or overlapping noise when the same site is repeatedly photographed.

[0150] In addition, the alignment can be performed in advance in a case where the photographing positions of the images 32a, 32b to 32m in the image 232 are slightly shifted. The alignment can be performed by calculating the displacement amount between the images with respect to each of the images 32b to 32m with respect to the image 32a, and moving the images 32b to 32m in parallel by the displacement amount, but the alignment can be performed by a method other than this. In the case of the image 32b and the image 32a, the displacement amount between the images can be calculated under a condition that the sum of the differences in the brightness of each pixel of the images 32b and 32a is minimized when the image 32b is moved in parallel according to the displacement amount. In addition, the displacement amount can be calculated under a condition that the amount of pixels in which the labels are not consistent between the regionally divided images 42b and 42a is minimized when the regionally divided image 42b is moved in parallel according to the displacement amount. The object to be calculated for the displacement amount between the images can be the image 32a other than the initial image.

[0151] (v) Step S205

[0152] The supervising production section 201 provides a user interface for the operator to confirm whether the labels assigned to the regionally divided images in the regionally divided image group 243 are correct. The user interface displays each of the regionally divided images 40 in the regionally divided image group 243. At this time, in order to easily determine whether the label assigned to the regionally divided image 40 is appropriate, the images 30 in the image group 233 can be arranged and displayed on the user interface in step S205, or a mixed image in which the regionally divided image 40 is transparent can be additionally displayed on the image 30. In the user interface provided in step S205, a function of correcting the labels of the regionally divided images 40 in the regionally divided image group 243 can be provided. The function of correcting the labels is to display the regionally divided images 40 in the regionally divided image group 243 or the mixed image in the input screen 91 (refer to FIG. 9), and the labels in the display can be corrected by the operation of the input pen 93 in the input screen 91. Figure 4 ) in which the regionally divided images 40 in the regionally divided image group 243 or the mixed image are displayed, and the labels in the display can be corrected by the operation of the input pen 93 in the input screen 91.

[0153] (vi) Step S206

[0154] The supervising production section 201 outputs the regionally divided image group 241 and the regionally divided image group 243 as the supervising data 214.

[0155] <Effects of Embodiment 3>

[0156] According to Embodiment 3, in sample image 213 (i.e., image group 233), the object of the label assignment task can be narrowed down to image group 231 as a partial set using the user interface (main screen 90) provided by the supervision production unit 201, and labels can be assigned to all images 30 within sample image 213 to obtain supervision data 214. In Embodiment 1, regarding the learning model 11, the more numbers of the parent data in sample image 13, the more accurate the estimation result. On the other hand, there is a trade-off where the workload of assigning labels to supervision data 14 increases. However, according to Embodiment 3, by reducing the workload of label assignment, the above-mentioned trade-off can be eliminated. In particular, in overlap measurement, structures that repeatedly appear in semiconductor images are usually taken as the objects of overlap measurement. Therefore, it is expected that even if the number of parent data in image group 231 is significantly reduced, the accuracy of the estimation result in step S203 will not decrease to the point where it is difficult to correct by statistical processing in step S204 or by the user interface in step S205. Therefore, according to Embodiment 3, the reduction in workload is considered effective.

[0157] Furthermore, by performing the correction in the statistical processing of step S204, the reproducibility of the overlap measurement can be improved. Here, the reproducibility of the overlap measurement refers to the degree of deviation of the overlap amount shown in Equations 1 and 2 above when repeatedly photographing the same area in the semiconductor wafer. Typically, three times the standard deviation σ, i.e., 3σ, is used as the indicator. The effect of improving the reproducibility of the overlap measurement is that when the correction in the statistical processing of step S204 is performed, the labels of the region segmentation image group 242, etc., obtained by repeatedly photographing the same area are the same. As a result, when multiple images of the same area as image group 231 in the semiconductor wafer become input images 12, the labels of the region segmentation images 60 estimated by the region segmentation unit 3 based on the multiple images are close. Therefore, the overlap amount calculated by the overlap measurement unit 5 in S19 is close to uniform, and the reproducibility of the overlap measurement is improved (reduced). In addition, due to this effect, the correction based on statistical processing in step S204 can also be performed on the region segmentation image group 241 generated using the user interface provided in step S201.

[0158] <Modified Example 3>

[0159] (i) in Figure 19 In the flowchart shown, steps S204 or S205 can also be deleted. This is because step S204 alone also has the effect of correcting the labels of the region segmentation image group 243, and step S205 alone also has the effect of confirming and correcting the labels of the region segmentation image group 243.

[0160] (ii) In the correction based on statistical processing in step S204, a statistical amount other than the most frequent value can be applied, or additional processing other than taking the most frequent value can be performed. For example, in a case where the frequency of the most frequent value in the labels of the pixels of the same coordinates such as 32a, 32b to 32m is low, the labels of the pixels can be set to be large in deviation, and the labels of the invalid region 49 can be assigned.

[0161] In the correction based on statistical processing in step S204, instead of using the images in the image group 232 obtained by repeatedly capturing the same part of the semiconductor wafer, a plurality of similar partial regions can be extracted in the image group 232. Here, the partial region is a partial region in the image 30 such as the region 94 in Figure 4 In addition, the plurality of similar partial regions means a plurality of partial regions of the same size extracted from one image such as 32a or a plurality of images such as 32a to 32m in the image group 232 with high similarity as a condition. Here, the similarity can be determined from the correlation value of the luminance of the pixels of two images 30 in the partial region, or the similarity can be determined from the ratio of the pixels with the same label among the pixels in the two region division images 40 in the partial region. Since the semiconductor image is a repeating pattern, the plurality of similar partial regions can be easily found.

[0162] (iii) The correction based on statistical processing in step S204 can be set to a correction of the labels in the region division image group 243 in units of small regions. Here, the correction based on statistical processing is performed using Figure 20 The correction based on statistical processing is described. Figure 20 is a diagram for explaining an example of the correction based on statistical processing in step S204. Figure 20 An example of the grouped image 270 before the correction and an example of the grouped image 270' after the correction, which are obtained by the grouping unit 4 from an arbitrary one of the region division images 40 in the region division image group 243, are shown.

[0163] In Figure 20In this process, through correction based on this statistical processing, small region 71h is moved parallel to small region 71h'. The amount of parallel movement of small region 71h can be determined in a manner consistent with the average displacement of the centroids of small regions 71h and 72h (the displacement obtained by the method in step S17) and the average displacement of the centroids of other small regions 71i and 72i, 71j and 72j, 71k and 72k. In addition to correction on a small region basis, all small regions, including small region 71h, can also be moved parallel to make the statistical value of the displacement obtained in step S19 the target value. For example, within the region segmentation image group 243, in the set of parts where the grouped image 270 is the same as the shooting location in the semiconductor wafer, the average value of the statistical value of the displacement obtained in step S19 when calculating the statistical value of the displacement based on each element in the set of parts can be used as the target value. Alternatively, if the semiconductor wafers for which the images in sample image 213 were taken were manufactured with an artificial overlap amount (manufactured by artificially offsetting predetermined layers in a multilayer semiconductor), and the artificial overlap amount is known, a subset of images in sample image 213 that are identical to image 270 in terms of grouping and artificial overlap amount can be determined. In this specification, the artificial overlap amount is sometimes referred to as the design value of the overlap amount.

[0164] (iv) Furthermore, given the design value of the overlap amount of image group 233 known in advance, a target value can be determined to improve the sensitivity characteristics of the overlap amount. Figure 21 This is another diagram illustrating the statistical processing-based correction in step S204. Based on Figure 21 This can explain the reference (used to determine the amount of movement) used when generating the corrected grouped image 270' from the uncorrected grouped image 270.

[0165] exist Figure 21 In this context, consider a curve 290 where the X component of the design value of the overlap amount is set as the U-axis 295, and the X component of the overlap amount (Dx in Equation 1 above) is set as the V-axis 296. In this case, the straight line connecting the ideal value of the V-axis 296 with respect to the U-axis 295 is set as 297 (a straight line with a slope of 1 and an intercept of 0, a straight line with a slope of 1 passing through the centroids of points 292a, 293b, and 293c, or a straight line with a slope of 1 and a predetermined slope, etc.). Furthermore, the point 292a representing the X component of the overlap amount of the region segmentation image 270 can be lowered to point 294a after the straight line 297 as the target value's X component. That is, the vector 293a can be set as the parallel-shifted Xb. Furthermore, in... Figure 21In this way, the X component of the amount of overlap in the set of region division images 243 falls on the straight line 297 after the parallel movement. This corresponds to the condition in which the sensitivity characteristic of the overlap measurement is optimal in the set of region division images 243. That is, the slope of the regression straight line when the regression straight line is calculated from the points 294a, 294b, and 294c, and the like, approaches the straight line 297, or the regression residual of the regression straight line (which corresponds to the square distance of the average of the vectors 293a, 293b, and 293c, and the like) is minimized. For the Y component of the amount of overlap (Dy shown in the above-described Equation 2), the target value in which the sensitivity characteristic of the overlap measurement is optimal can also be calculated from the same curve as the curve 290.

[0166] For example, in the case where there are a plurality of image groups A, B, and C, it is desirable that all of the values be plotted on the straight line 297. However, in reality, it is difficult to obtain such an ideal state. Therefore, in the case where there is a displacement amount, if the displacement amount is corrected in units of image groups, at least the supervision data is plotted on the straight line 297 (within the range of the learning model, can be correctly plotted on the straight line 297). Since the displacement amount of the supervision data can be identified as the habit of the learning model and the habit of the supervision data, it is possible to suppress the deviation of the displacement amount in each image. That is, by performing correction based on statistical processing, it is possible to improve the sensitivity characteristic of the overlap measurement.

[0167] (v) In the case where the correction based on statistical processing of step S204 is performed in units of small regions, the above-described several corrections can also be combined and performed. In addition, in the correction in units of small regions, in addition to the parallel movement, geometric deformation of the center of gravity of the small region 71h, and the like, can also be performed. For example, as the geometric deformation, it is possible to cite cutting the right half of the small region 71h. Thereby, the center of gravity of the small region 71h is moved to the left side.

[0168] (vi) The supervision creation section 201 can display the grouped images 270 and 270' before and after the correction, and the images 30 and the region division images 40 corresponding to these, so that the correction of the statistical processing of step S204 can be confirmed on the main screen 90, and the like.

[0169] (vii) In the supervision production unit 201, the intermediate learning model shown in step S202 can be obtained in several stages. For example, a first intermediate learning model is obtained based on image group 231 and region segmentation image group 241, and a second intermediate learning model is obtained based on image group 232 and region segmentation image group 242 whose labels are obtained by referring to the estimation result of the first intermediate learning model, in addition to image group 231 and region segmentation image group 241. Furthermore, labels can also be assigned to all region segmentation images 40 in the supervision data 214 by referring to the second intermediate learning model and estimating other than image group 232 in image group 233.

[0170] (5) Example 4

[0171] use Figures 22 to 25 Example 4 will be described. Figure 22 This is a diagram illustrating the functional structure example of Example 4, from supervisory data creation to overlay measurement.

[0172] <Overview of Functional Structure>

[0173] First, a brief overview of the functional structure will be provided. In addition to the functions of the production supervision unit 1, the production supervision unit 301 also has the function of maintaining the position of the main screen 90 (see reference). Figure 4 Each pixel in the region segmentation image 40 is mapped to a small region 71a within the region segmentation image 40 (see reference). Figure 8 The function of the position information image representing the displacement of the position is as follows: Additionally, the supervision production unit 301 produces supervision data 314 by adding the position information image to the supervision data 14. The learning unit 302 calculates a learning model 311 based on image 30 in the sample image 13, which can estimate the region segmentation image 40 and the position information image in the supervision data 314 as accurately as possible.

[0174] The region segmentation unit 303, referring to the learning model 311, estimates the region segmentation image 60 and the location information image based on the input image 12. The grouping unit 304 generates and outputs the grouped image 70 based on the region segmentation image 60 in the same order as the grouping unit 3. The overlap measurement unit 305 performs overlap measurement using the location information of the location information image included in the small region 71a, etc., in the grouped image output from the grouping unit 304.

[0175] <Detailed description of functional structure>

[0176] Hereinafter, except for the grouping section 304 which is not specifically described in the summary, the details of each functional structure in Embodiment 4 will be described.

[0177] (i) The supervision production section 301 produces the supervision data 14 from the sample image 13 in response to the operator's operation on the main screen 90, and adds the position information image 340 described below to the region segmentation image 40 in the supervision data to output the supervision data 314. Referring to Figure 23 The position information image 340 will be described.

[0178] Figure 23 is a diagram showing an example of the position information image 340. In Figure 23 , the image 370 is a grouping image obtained by the grouping section 4 from the region segmentation image 40 in the sample image 13. The grouping image 370 is obtained from the label of the first measurement object contained in Table 80 (refer to Figure 10 ). The position information image 340 is an image in which position information is given to each pixel within the small regions 371m, 371n, 371o, 371p corresponding to the small regions 71m, 71n, 71o, 71p in the grouping image 370. The range of the small region 371m is the same range as the small region 71m, or a range in the vicinity of the small region 71m obtained by performing dilation processing on the small region 71m, or by obtaining an outer rectangle of the small region 71m. The ranges of the other small regions 371n and the like are also the same or in the vicinity of the small regions 71n and the like. If the coordinates of a certain pixel 342m within the small region 371m are set as (Xp, Yp), and the X coordinate of the representative position 341m is set as (Xc, Yc), the displacement amount (Rx, Ry) obtained by the following Equations 3 and 4 is assigned to the pixel 342m. This displacement amount (Rx, Ry) corresponds to the displacement amount from the pixel 342m to the representative position 341m. Here, the representative position 341m is set as the coordinates of the center of gravity of the small region 371m. The displacement amount up to the representative position 341m is assigned to each pixel within the small region 371m in the same manner as the pixel 342m.

[0179] Rx= Xc- Xp (Equation 3)

[0180] Ry= Yc- Yp (Equation 4)

[0181] In the same manner, the displacement amount from each pixel to the representative position 341n, 341o, 341p of each small region is assigned to each pixel within the small regions 371n, 371o, 371p in the position information image 340. In the position information image 340, the attributes of the invalid regions excluded in the calculation of obtaining the learning model 311 in the learning section 302 are assigned to regions other than the small regions 371m, 371n, 371o, 371p.

[0182] Further, the supervision production section 301 also gives the same position information image as the position information image 340 to the label of the second measurement object in Table 80.

[0183] (ii) The learning unit 302 calculates the learning model 311 capable of estimating the region segmentation image 40 and the position information image 340 in the supervision data 314 as accurately as possible from the image 30 in the sample image 13. The learning model for estimating the region segmentation image 40 from the image 30 and the learning model for estimating the position information image 340 from the image 30 can be respectively assigned independent neural network structures 179. Alternatively, all or a part of the layers of the neural network structure 179 can be shared between the 2 learning models. For example, when the image 30 is input to the input layer 170, the region segmentation image 40 can be output from a part of the output layer 171, and the position information image 340 can be output from other layers such as the intermediate layer 174 or a part other than the part of the output layer 171.

[0184] As with the learning unit 2, the learning unit 302 optimizes the parameters in the learning model 311 so that the difference when comparing the region segmentation image 40 and the position information image 340 estimated from the image 30 in the sample image 13 with the corresponding region segmentation image 40 and the position information image 340 in the supervision data 314 becomes small. For example, the parameters in the learning model 311 are initialized by random numbers. Thereafter, the error (error 1) of the 2 region segmentation images 40 is set to the number of pixels in which the labels are not consistent, and the error (error 2) of the 2 position information images 340 is set to the sum of the absolute values of the differences in the displacement amounts in the respective pixels. And, each parameter in the learning model 311 is sequentially added to the value obtained by multiplying the partial differential coefficient with respect to the error 1 and the error 2 by a predetermined coefficient in the negative. By repeating such processing for the image 30 in the sample image 13, optimization can be performed. However, the method is not limited to this.

[0185] (iii) Figure 24 is a diagram showing an example of the position information image 360. The region segmentation unit 303 refers to the learning model 311 to perform estimation of the region segmentation image 60 and the position information image 360 from the input image 12 (i.e., the image 50). Figure 24 In the case of correct estimation, Figure 24 the region 369 in the image 50 becomes a region of the same range as the region 59 in the position information image 360. In the region 369, the pixel 342a within the small region 371a contains information of the estimated value (Rix, Riy) of the displacement amount up to the representative position 341a of the small region 371a. If the coordinates of the pixel 342a are set to (Xip, Yip), the estimated value (Xic, Yic) of the position of the representative position 341a can be obtained by the following (Equation 5) and (Equation 6).

[0186] Xic = Rix + Xip ··· (Equation 5)

[0187] Yic = Riy + Yip ... (Equation 6)

[0188] Here, for the small regions 71a, 71b, 71c, and 71d obtained by the grouping unit 4 from the labels of the first measurement object in Table 80, small regions 371a, 371b, 371c, and 371d become small regions with the same range as small region 71a, etc. Based on each pixel in small regions 371b, 371c, and 371d, the positions representing positions 341b, 341c, and 341d can also be estimated using Equations 5 and 6 (estimated values ​​can be calculated). In the position information image 360, the portion outside region 369 also stores the same position information as pixel 342a.

[0189] In addition, the region segmentation unit 303 also outputs the same location information image as the location information image 360 ​​from the label of the second measurement object in Table 80.

[0190] (iv) Overlap measurement unit 305 according to Figure 25 The flowchart shown illustrates how to perform overlap measurement processing.

[0191] exist Figure 25 In the flowchart, excluding steps S314 and S316, the process performed by the overlap measurement unit 5 during overlap measurement processing is as follows ( Figure 11 Since the same applies, the explanation is omitted. Hereinafter, the case where the loop of steps S13 to S18 is applied to the small region 371a will be described.

[0192] In step S314, the overlap measurement unit 305 calculates the estimated value representing position 341a based on each pixel (pixel 342a, etc.) in the small region 371a of the position information image 360 ​​using Equations 5 and 6 above. Next, the overlap measurement unit 305 calculates the estimated value representing position 341a based on each pixel and their statistical measure, setting this statistical measure as position 1. The statistical measure can be calculated as a median value, but other statistical measures such as summation or averages can also be used.

[0193] In step S316, the overlap measurement unit 305 also obtains position 2 (a statistical measure representing the estimated value of the position) by performing the same processing as in step S314 on the second measurement object shown in Table 80, based on the position information image of the small area obtained from the label of the second measurement object.

[0194] <Technical Effects of Example 4>

[0195] According to Embodiment 4, the amount of overlap can be measured using the position information image 360. Sometimes the quality of the input image 12 is degraded by superimposing random noise, or reducing the contrast, etc. In such a case, the labels in the region segmentation image 60 become incorrect, the range of the small region 71a, etc. in the grouping image 70 becomes incorrect, but if the technique disclosed in Embodiment 4 is used, correct overlap measurement can be performed even in such a case. For example, in the case where the right half of the small region 71a is missing, the center of gravity of the small region 71a shifts to the left from the original position, as a result of which the correct position 1 cannot be found in step S14. In contrast to this, in step S314, the representative position 341a can be calculated from any pixel of the small region 371a in the position information image 370 by using the equations 5 and 6. Therefore, even in the case where the right half of the small region 71a (i.e., the small region 371a) is missing, the correct position 1 can be calculated in step S314.

[0196] <Embodiment 4>

[0197] (i) The supervision production section 301 can also correct the region segmentation image 40 in the supervision data 14 by the statistical processing of step S204. By this, the center of gravity of the small region 71a, etc. in the region segmentation image 40 can be corrected in a manner that improves the reproducibility or sensitivity characteristic of the overlap measurement in the overlap measurement section 5. Therefore, each pixel (342m, etc.) in the position information image 370 in the supervision data 314 determined from the center of gravity of the small region (71a, etc.) in the region segmentation image 40 can also be corrected in a manner that improves the reproducibility or sensitivity characteristic of the overlap measurement.

[0198] (ii) The grouping section 304 can also refer to the position information image 360 when finding the grouping image 70 from the region segmentation image 60. For example, sometimes two small regions (e.g., the small regions 71m and 71n) are connected to form one small region by the two small regions (e.g., the small regions 71m and 71n) being inflated beyond the original range in the grouping image 70. In this case, in the position information image 360, the representative positions found by the equations 5 and 6 within one small region are sometimes separated (in the position information image 360, the representative positions are separated into 371m and 371n corresponding to 71m and 71n). Therefore, in the case where the representative positions are separated, the one small region can also be divided by referring to the separation of the representative positions. This division of the small region is useful, for example, in the case where the small regions in the grouping image 70 become inaccurate due to the quality of the input image 12 being less sharp than the sample image 30.

[0199] (6) Embodiment 5

[0200] Example 5 discloses a technique of generating a region segmentation image from a limited amount of sample images (corresponding to supervised data), moving small regions within the region segmentation image in parallel to perform layout change of the region segmentation image and the sample images (compositing the region segmentation image and the sample images), thereby increasing the supervised data and the sample images.

[0201] <Outline of functional configuration>

[0202] Figures 27 to 29 is a diagram showing an example of a functional configuration from supervised data production to overlay measurement of Example 5. The supervised production section 401 generates supervised data 14 from a sample image 13 and a user interface provided therefor, similarly to the supervised production section 1. In addition, the supervised production section 401 generates and outputs supervised data 414 in which the layout of a region segmentation image 40 in the supervised data 14 is changed. Also, the supervised production section 401 has a function of estimating an image 30 from the region segmentation image 40 (hereinafter, this function will be referred to as an image estimation function of the supervised production section 401), and outputs a sample image 413 by estimating each of the corresponding images 30 from the region segmentation image 40 in the supervised data 414. Further, each of the functions of the learning section 2, the region segmentation section 3, the grouping section 4, and the overlay measurement section 5 is the same as that of Example 1, and thus the description thereof is omitted. That is, the learning section 2 in Example 5 regards the supervised data 14 and the supervised data 414 and the sample image 13 and the sample image 413 as homogeneous data, and calculates the learning model 11 in the same steps as those of Example 1.

[0203] <Detail of functional configuration>

[0204] Hereinafter, the processing when the supervised production section 401 outputs the supervised data 414 and the sample image 413 will be described in detail with reference to Figure 27

[0205] (i) Figure 27 is a diagram showing an example of an image 430 estimated from a region segmentation image 440. In Figure 28 ​In the present embodiment, the image 440 represents an example of a region segmentation image produced by the operator using the user interface of the supervision production section 1 provided by the supervision production section 401. The image 440 is composed of labels 41, 42, and 43. In addition, the small region 71q represents an example of a small region of the label 41 within the region segmentation image produced by the grouping section 4. Furthermore, the image 430 represents an image estimated by the supervision production section 401 from the region segmentation image 440 using the image estimation function. In this case, the image estimation function of the supervision production section 401 can be implemented by (a) collecting samples of an arbitrary image 30 and a group of region segmentation images 40 in advance, like the sample image 13 and the supervision data 14, as with the learning section 2, and (b) performing learning to find parameters in a structure similar to the neural network 179 that minimizes the error between an image 30 estimated from one region segmentation image 40 and an image included in the arbitrary image 30 and region segmentation image 40 group and corresponding to the one region segmentation image 40. In this case, the error between the image corresponding to the region segmentation image 40 and the estimated image 30 can be found by summing the absolute values of the differences in luminance of each pixel, but is not limited to this method. In addition, the image estimation function of the supervision production section 401 can be implemented by determining the parameters in a network structure similar to the neural network structure 179 using a machine learning algorithm called a generative adversarial network, which is excellent at generating images from random numbers or symbols, in addition to the above-described method.

[0206] In addition, in the network structure similar to the neural network structure 179, estimation can be performed in units of the receptive field 176, as with the network structure 179. That is, for each pixel of the region segmentation image 40, the luminance of the pixel at the same coordinates in the image 30 is determined from the range of the receptive field 176 around the pixel in the region segmentation image 40.

[0207] (ii) Figure 28 is a diagram representing an example of a case in which the small region 71q is layout-changed from the region segmentation image 440. In this case, the image 430 represents an image estimated by the supervision production section 401 from the region segmentation image 440 using the image estimation function. In this case, the image estimation function of the supervision production section 401 can be implemented by (a) collecting samples of an arbitrary image 30 and a group of region segmentation images 40 in advance, like the sample image 13 and the supervision data 14, as with the learning section 2, and (b) performing learning to find parameters in a structure similar to the neural network 179 that minimizes the error between an image 30 estimated from one region segmentation image 40 and an image included in the arbitrary image 30 and region segmentation image 40 group and corresponding to the one region segmentation image 40. In this case, the error between the image corresponding to the region segmentation image 40 and the estimated image 30 can be found by summing the absolute values of the differences in luminance of each pixel, but is not limited to this method. In addition, the image estimation function of the supervision production section 401 can be implemented by determining the parameters in a network structure similar to the neural network structure 179 using a machine learning algorithm called a generative adversarial network, which is excellent at generating images from random numbers or symbols, in addition to the above-described method. Figure 29In this case, the region-divided image 440r is an image in which the layout of the small region 71q in the region-divided image 440 is changed so as to be parallelly moved to the small region 71r. According to the effect of the estimation in the unit of the receptive field 176, if the image 430r estimated from the region-divided image 440r is compared with the image 430 corresponding to the region-divided image 440, it is known that the image within the small region 431q corresponding to the small region 71q is moved to the small region 431r within the small region 71r. If the small region 431q and the small region 431r are focused on, it is equivalent to the image in which the amount of overlap is changed by the amount of parallel movement from the small region 71q to the small region 71r. Therefore, in the region-divided image 440r, if all the small regions 71r and the like are parallelly moved by a uniform amount, the image 430r becomes an image in which the amount of overlap is changed by a uniform amount from the image 430.

[0208] <Effects of Embodiment 5>

[0209] According to Embodiment 5, by the layout change, it is possible to increase the supervised data 414 in the supervised data 14. In addition, by using the layout change and the image estimation function of the supervised production section 401, it is possible to increase the sample image 413 in the sample image 13. For example, even in the case where the sample image 13 is constituted by the image 30 in which the amount of overlap is uniform, by the layout change, it is possible to increase the region-divided image 40 and the image 30 of various amounts of overlap in the supervised data 14 and the sample image 13. Therefore, the operator does not need to prepare a plurality of sample images having various layouts in advance, and it is possible to save the effort for the overlay measurement.

[0210] <Modifications of Embodiment 5>

[0211] (i) In the layout change, in addition to the above-mentioned parallel movement, it is possible to apply any geometric deformation such as enlargement, reduction, and the like, which accompanies a change in the amount of overlap.

[0212] (ii) It is also possible to decide the depth information of the labels 41, 42, and 43 in the region-divided image 440 in advance. Thereby, the supervised production section 401 can estimate an image in which the occlusion between the labels is taken into consideration. Figure 29 is a diagram showing an example in which the occlusion occurs when the layout of the small region 71q in the region-divided image 440 is changed. As shown in Figure 30 , the range overlapped with the label 43 when the small region 71q in the region-divided image is parallelly moved to the small region 71s is deleted. As a result, in the image 430s estimated from the region-divided image 440s, with respect to the small region 431s corresponding to the small region 71s, an image in which the occlusion is taken into consideration is estimated. Thereby, it is possible to synthesize a region-divided image 430 which is closer to reality by the occlusion between the labels.

[0213] (iii) In addition to being implemented individually, the above-described embodiments 1 to 5 can also be implemented by combining a part or all of the embodiments.

[0214] (7) Example 6

[0215] Example 6 describes an example of applying the supervision data creation and processing, learning model creation and processing, and region segmentation processing from Examples 1 to 5 to measurement processing other than overlapping measurement.

[0216] Figure 30 This is a diagram illustrating the functional structure example of Example 6, from supervisory data creation to image measurement and inspection. Figure 30 In this context, the supervision production unit 501 corresponds to any one of the supervision production units 1, 101, 201, 301, and 401 in Examples 1 to 5. Similarly, the supervision data 514, the learning unit 502, the learning model 511, and the region segmentation unit 503 correspond to the supervision data, learning unit, learning model, and region segmentation unit in any one of Examples 1 to 5, respectively.

[0217] The image measurement and inspection unit 505 uses the region segmentation unit 503 to estimate the region segmentation image 60 based on the input image 12 to perform image inspection and measurement, not limited to overlap measurement. Examples of image inspection and measurement in the image measurement and inspection unit 505 include contour extraction from semiconductor images, size measurement such as hole shapes, detection of defect patterns such as short-circuit defects, and pattern matching to find the comparison position between the estimated design drawing and the actual design drawing based on the image. However, it is not limited to these applications and can be used for any purpose of image measurement using the region segmentation image 60 estimated by the region segmentation unit 503. Furthermore, in the image measurement and inspection unit 505, the input image 12 can be used as an auxiliary reference for purposes such as correcting the region segmentation image 60. Figure 19 Other data not shown. Additionally, in Embodiment 6, the sample image 13 and the input image 12 may also be images obtained by capturing images other than those of a semiconductor.

[0218] (i) The case where the functions of the production supervision unit 501, the learning unit 502, and the area division unit 503 are respectively equipped with the functions of Embodiment 1.

[0219] For example, in order to learn model 511 and make inferences in units of receptive field 176, if the case reflects a periodic pattern within the input image 12, then the size of image 30 in sample image 13 when assigning labels to supervision data 14 in supervision production unit 501 can be smaller than the size of image 60 in input image 12. This reduces the time spent by the operator assigning labels to supervision data 14.

[0220] (ii) a case where each function of the supervision production section 501, the learning section 502, and the region division section 503 is provided with the functions of Embodiment 2

[0221] For example, by using a plurality of image groups in which the photographing conditions are changed, it is possible to accurately produce the supervision data 114 using an image in which a structure that becomes a target in the images 30a and 30b and the like is clearly expressed, and to accurately perform the estimation of the region division section 503.

[0222] (iii) a case where each function of the supervision production section 501, the learning section 502, and the region division section 503 is provided with the functions of Embodiment 3

[0223] For example, by reducing the target to which the supervision data is assigned to the image group 231 using the main screen 90 in the sample image 213, it is possible to reduce the working hours of the operator. In addition, by performing the steps S202 and S203 of Embodiment 3, it is possible to acquire the supervision data 514 (corresponding to the supervision data 214) with respect to the remaining total number of the sample images 13 (corresponding to the sample image 213). Furthermore, by performing the steps S204 and S205, it is also possible to correct the supervision data 514 (corresponding to the supervision data 214). ​

[0224] (iv) a case where each function of the supervision production section 501, the learning section 502, and the region division section 503 is provided with the functions of Embodiment 4

[0225] Image measurement using the region division image 60 and the position information image 360 obtained on the basis of the input image 12 can be performed.

[0226] (v) a case where each function of the supervision production section 501, the learning section 502, and the region division section 503 is provided with the functions of Embodiment 5

[0227] For example, by using the image estimation function in the supervision production section 501 to change the layout in the region division image 40, the sample image 413 and the supervision data 414 (combining the supervision data 14 and the supervision data 414 into the supervision data 514) are added to the sample image 13 and the supervision data 14.

[0228] <Effects of Embodiment 6>

[0229] According to Embodiment 6, it is shown that not only the techniques disclosed in Embodiments 1 to 5 can be applied to the overlay measurement, but also to all systems in which image measurement and image inspection are performed using a region division image.

[0230] (8) Other Embodiments

[0231] ​Each of the embodiments can also be realized by program codes of software. In this case, a storage medium having the program codes recorded therein is supplied to a system or an apparatus, and a computer (or CPU, MPU) of the system or the apparatus reads out the program codes stored in the storage medium. In this case, the program codes read out from the storage medium realize the functions of the above-described embodiments by themselves, and the program codes themselves and the storage medium storing the program codes constitute the present disclosure. As the storage medium for supplying such program codes, for example, a floppy® disk, a CD-ROM, a DVD-ROM, a hard disk, an optical disk, a magneto-optical disk, a CD-R, a magnetic tape, a non-volatile memory card, a ROM, and the like are used.

[0232] In addition, a part or all of the actual processing can be performed by an OS (operating system) or the like running on a computer based on the instructions of the program codes, and the functions of the above-described embodiments can be realized by the processing. Further, after the program codes read from the storage medium are written in the memory of the computer, a CPU or the like of the computer can execute a part or all of the actual processing based on the instructions of the program codes, and the functions of the above-described embodiments can be realized by the processing.

[0233] Further, the program codes of the software realizing the functions of the embodiments can be distributed via a network, and saved in a storage unit or a storage medium such as a hard disk or a memory of a system or an apparatus, or a CD-RW, a CD-R, and the like, and in use, a computer (or CPU or MPU) of the system or the apparatus reads out and executes the program codes saved in the storage unit or the storage medium.

[0234] Symbol Explanation

[0235] 1, 101, 201, 301, 401, 501 Supervision production section

[0236] 2, 102, 302, 502 Learning section

[0237] 3, 103, 303, 503 Region division section

[0238] 4, 104, 304 Grouping section

[0239] 5, 105, 305 Overlapping measurement section

[0240] 11, 111, 311, 511 Learning model

[0241] 12, 112 Input image

[0242] 13, 113, 213, 413 Sample image

[0243] 14, 114, 214, 314, 414, 514 Supervision data

[0244] 190 main processor

[0245] 190a first sub processor

[0246] 190b second sub processor

[0247] 191 main computer

[0248] 191a first sub computer

[0249] 191b second sub computer

[0250] 192 input output device

[0251] 193 electron microscope or the like

[0252] 505 image measurement checking section

Claims

1. A measurement system for performing image measurements of semiconductors containing periodic structures, characterized in that, The measurement system includes: At least one processor performs various processes related to the image measurement; and An output device that outputs the results of the image measurement. The at least one processor performs the following processing: Processing to generate supervisory data based on semiconductor sample images; Processing to generate a learning model based on the sample images and the supervised data; Based on the learning model, a process is performed to generate a region segmentation image from an input image related to the semiconductor; Measurement processing involves performing image measurements using the region segmentation image; as well as The result of the measurement processing is output to the output device for further processing. The supervisory data is an image obtained by assigning labels containing the structure of the semiconductor in the sample image to each pixel of the image. The learning model includes parameters for inferring the supervision data or the region segmentation image based on the sample image or the input image. The sample image and the supervision data contain image regions that are smaller in size than the input image and correspond to the periodic structure. In the process of generating the learning model, the at least one processor generates the parameters of the learning model based on the sample image and the supervision data.

2. The measurement system according to claim 1, characterized in that, The learning model is a machine learning model that refers to the neighborhood of each pixel in the input image when determining the label to be assigned to each pixel.

3. The measurement system according to claim 1, characterized in that, The learning model is a convolutional neural network.

4. The measurement system according to claim 1, characterized in that, The image region is an image region corresponding to at least one periodic structure in the periodic structure.

5. The measurement system according to claim 1, characterized in that, The measurement process is the semiconductor's overlap measurement process, size measurement process, defect pattern detection process, or pattern matching process.

6. A measurement system for performing image measurements of a semiconductor comprising a predetermined structure, characterized in that, The measurement system includes: At least one processor performs various processes related to the image measurement; and An output device that outputs the results of the image measurement. The at least one processor performs the following processing: Processing to generate supervisory data based on semiconductor sample images; Processing to generate a learning model based on the sample images and the supervised data; Based on the learning model, a process is performed to generate a region segmentation image from an input image related to the semiconductor; Measurement processing involves performing image measurements using the region segmentation image; as well as The result of the measurement processing is output to the output device for further processing. The supervisory data is an image obtained by assigning labels containing the structure of the semiconductor in the sample image to each pixel of the image. The learning model includes parameters for inferring the supervision data or the region segmentation image based on the sample image or the input image. The at least one processor further performs the following processing: further dividing the region segmentation image into smaller regions corresponding to the label, and grouping them according to each category of the smaller region. As part of the measurement process, the at least one processor performs an overlap measurement starting from the centroid of each of the grouped subregions.

7. The measurement system according to claim 6, characterized in that, The supervisory data includes location information images, which represent the displacement from each pixel to a representative position in a small region assigned the label. The at least one processor generates the region segmentation image and the location information image of the input image based on the learning model containing the location information image, and uses the location information image in the grouped small regions to perform the overlap measurement.

8. The measurement system according to claim 7, characterized in that, The location information image represents the displacement calculated using the supervised data after correction based on statistical processing.

9. A measurement system for performing image measurement of a semiconductor comprising a predetermined structure, characterized in that, The measurement system includes: At least one processor performs various processes related to the image measurement; and An output device that outputs the results of the image measurement. The at least one processor performs the following processing: Processing to generate supervisory data based on semiconductor sample images; Processing to generate a learning model based on the sample images and the supervised data; Based on the learning model, a process is performed to generate a region segmentation image from an input image related to the semiconductor; Measurement processing involves performing image measurements using the region segmentation image; as well as The result of the measurement processing is output to the output device for further processing. The supervisory data is an image obtained by assigning labels containing the structure of the semiconductor in the sample image to each pixel of the image. The learning model includes parameters for inferring the supervision data or the region segmentation image based on the sample image or the input image. The sample images comprise a group of images obtained by taking multiple images of the same location in the semiconductor under different shooting conditions. The at least one processor generates the supervision data corresponding to the sample image and the shooting conditions, and generates the learning model based on the supervision data generated corresponding to the shooting conditions and the sample image.

10. The measurement system according to claim 9, characterized in that, Taking pictures under the different shooting conditions includes at least one of changing the accelerating voltage to take pictures, taking pictures of different types of electronic images, or changing the synthesis ratio when generating composite images of different types of electronic images.

11. A measurement system for performing image measurements of a semiconductor comprising a predetermined structure, characterized in that, The measurement system includes: At least one processor performs various processes related to the image measurement; and An output device that outputs the results of the image measurement. The at least one processor performs the following processing: Processing to generate supervisory data based on semiconductor sample images; Processing to generate a learning model based on the sample images and the supervised data; Based on the learning model, a process is performed to generate a region segmentation image from an input image related to the semiconductor; Measurement processing involves performing image measurements using the region segmentation image; as well as The result of the measurement processing is output to the output device for further processing. The supervisory data is an image obtained by assigning labels containing the structure of the semiconductor in the sample image to each pixel of the image. The learning model includes parameters for inferring the supervision data or the region segmentation image based on the sample image or the input image. The at least one processor segments the sample image into two or more sample image groups, generates first supervision data by assigning labels to the images contained in the first sample image group, generates an intermediate learning model based on the images in the first sample image group and the first supervision data, appends supervision data generated by inferring the images contained in image groups other than the first sample image group based on the intermediate learning model to the first supervision data to generate second supervision data, and generates the learning model for application to the input image based on the sample image and the second supervision data.

12. The measurement system according to claim 11, characterized in that, The at least one processor performs statistically-based correction on the supervisory data generated by estimating the images contained in an image group other than the first sample image group.

13. The measurement system according to claim 12, characterized in that, The at least one processor performs correction based on the statistical processing on multiple images obtained by repeatedly photographing the same part of the semiconductor.

14. The measurement system according to claim 12, characterized in that, The at least one processor extracts a region with high similarity from the sample image and performs correction on the extracted region based on the statistical processing.

15. The measurement system according to claim 12, characterized in that, The correction based on the statistical processing is to perform parallel shifts or geometric deformations on a unit basis, using small regions in the second supervised data that are assigned the labels.

16. A measurement system for performing image measurements of a semiconductor comprising a predetermined structure, characterized in that, The measurement system includes: At least one processor performs various processes related to the image measurement; and An output device that outputs the results of the image measurement. The at least one processor performs the following processing: Processing to generate supervisory data based on semiconductor sample images; Processing to generate a learning model based on the sample images and the supervised data; Based on the learning model, a process is performed to generate a region segmentation image from an input image related to the semiconductor; Measurement processing involves performing image measurements using the region segmentation image; as well as The result of the measurement processing is output to the output device for further processing. The supervisory data is an image obtained by assigning labels containing the structure of the semiconductor in the sample image to each pixel of the image. The learning model includes parameters for inferring the supervision data or the region segmentation image based on the sample image or the input image. The at least one processor further performs the following processing: changing the layout of the supervision data to generate modified supervision data, and appending the modified supervision data to the supervision data before the layout change as a process of updating supervision data; And the process of appending the image inferred based on the change monitoring data to the sample image as an updated sample image. The at least one processor generates the learning model based on the updated supervision data and the updated sample image.

17. The measurement system according to claim 16, characterized in that, The at least one processor considers the occlusion between labels contained in the supervision data to change the layout of the supervision data.

18. A method for generating a learning model, which is used when performing image measurements of a semiconductor containing a predetermined structure, characterized in that, The method includes: At least one processor generates supervisory data by assigning labels to a region segmentation image obtained from a sample image of a semiconductor, each label containing a structure of at least one measurement object; and The at least one processor generates the learning model based on a multi-layered network structure, using the region segmentation image of the sample image and the supervision data. The learning model includes parameters for estimating the supervision data or the region segmentation image based on the sample image or an input image related to the semiconductor. The at least one processor further performs the following processing: further dividing the region segmentation image into smaller regions corresponding to the label, and grouping them according to each category of the smaller region. The at least one processor performs overlap measurements from the centroid of each of the grouped subregions.

19. A storage medium storing a program for causing a computer to perform processing to generate a learning model, said learning model being used in image measurement of a semiconductor comprising a predetermined structure, characterized in that, The program causes the computer to perform the following processing: The process of generating supervisory data involves assigning labels to structures containing at least one measurement object to region-segmented images obtained from semiconductor sample images; and Based on a multi-layered network structure, the learning model is generated using the region segmentation images of the sample images and the supervision data. The learning model includes parameters for estimating the supervision data or the region segmentation image based on the sample image or an input image related to the semiconductor. The program also causes the computer to perform the following processes: Corresponding to the labels, the region segmentation image is further divided into smaller regions with smaller image sizes, and grouped according to each category of the smaller region; and The process of performing overlap measurements is performed starting from the centroid of each of the grouped small regions.

Images