Pattern matching device, pattern measurement system, pattern matching program

Abstract

The present disclosure provides a pattern matching device, and in particular, enables matching processing characterized by a learning function even in a semiconductor pattern including a repeating pattern. The pattern matching device of the present disclosure includes a learner that estimates a first correlation image having a numerical value representing the correlation between a first image and a second image as a pixel value. The pattern matching device calculates a second correlation image having a numerical value representing the correlation between a derived image generated from the first image and the first image as a pixel value. The learner learns in such a way as to reduce the difference between the first correlation image and the second correlation image (see FIG. 1).

Description

Technical Field

[0001] The present disclosure relates to techniques for performing pattern matching between images. Background Art

[0002] Devices that measure and inspect patterns formed on semiconductor wafers often use template matching technology to align the inspection device's field of view with the desired measurement or measurement location. Template matching is the process of finding the area in the image of the target that most closely matches a pre-registered template image. Patent Document 1 describes an example of such template matching.

[0003] Patent Document 2 describes a method for creating a template for template matching based on semiconductor device design data. Creating a template based on design data has the advantage of eliminating the need to acquire images using an inspection device specifically for creating the template.

[0004] Non-Patent Document 1 describes a learnable model that takes two images as input and outputs transformation parameters (e.g., matching offsets) within the images. Using supervised data for learning enables highly accurate model learning, with the advantage of eliminating the need for manual input image processing.

[0005] Prior art literature

[0006] Patent Literature

[0007] Patent Document 1: Japanese Patent No. 4218171 (corresponding to U.S. Patent No. 6627888)

[0008] Patent Document 2: Japanese Patent No. 4199939 (corresponding to U.S. Patent No. 7235782)

[0009] Non-patent literature

[0010] Non-patent literature 1: I. Rocco, R. Arandjelovic and J. Sivic. Convolutional neural network architecture for geometric matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017 Summary of the Invention

[0011] Problems to be solved by the invention

[0012] In template matching, when there is a large deviation in the appearance of the image between the template image and the image to be searched, the matching sometimes fails. As reasons for the increase in the deviation of the appearance of the image between the template image and the image to be searched, for example, the following cases can be cited: (a) the difference in the shooting conditions of the inspection device when registering the template and the shooting conditions of the inspection device when shooting the image to be searched becomes large; (b) the difference in the final quality of the semiconductor pattern captured when registering the template and the final quality of the semiconductor pattern when shooting the image to be searched becomes large; (c) in a multi-layer pattern, the difference in the deviation of the upper and lower layers of the semiconductor multi-layer pattern captured when registering the template and the deviation of the upper and lower layers of the semiconductor multi-layer pattern when shooting the image to be searched becomes large; (d) the semiconductor process of the semiconductor pattern when registering the template is different from the manufacturing process of the semiconductor pattern when shooting the image to be searched (this case is sometimes implemented to reduce the template registration man-hour by reusing the template image in other processes).

[0013] Patent Document 1 does not disclose how to handle the matching process when there is a deviation between the template image and the image to be searched.

[0014] Patent Document 2 discloses performing a smoothing process on the graphic data generated based on the design data and rounding each part of the pattern to produce a template of a pattern close to the actual image. However, it is difficult to cope with all deviations between the template image and the image to be searched.

[0015] In addition, in the conventional pattern matching as in Patent Documents 1 and 2, when the template matching fails due to such a deviation, it takes time to improve the algorithm and it cannot be simply improved in a short time.

[0016] Non-Patent Document 1 can absorb the deviation between two images through short-term learning and achieve the learning of a high-precision model. However, in the characteristic repetitive patterns (patterns with multiple matching offset amounts at the matching correct position) in semiconductor patterns, it is difficult to learn the model. This is because the repetitive patterns have multiple identical shapes within a close range (for example, within the field of view), so there are multiple candidates for the true correct position, and it is difficult to determine the true correct position.

[0017] The present disclosure has been completed to solve such problems, and in particular, a pattern matching device is proposed that can achieve a matching process characterized by a learning function even in a semiconductor pattern including repetitive patterns.

[0018] Means for Solving the Problems

[0019] The pattern matching device of the present disclosure includes a learner that estimates a first correlation image having a numerical value representing the correlation between a first image and a second image as a pixel value. The pattern matching device calculates a second correlation image having a numerical value representing the correlation between a derived image generated from the first image and the first image as a pixel value. The learner learns in such a way as to reduce the difference between the first correlation image and the second correlation image.

[0020] Advantages of the Invention

[0021] According to the pattern matching device of the present disclosure, in particular, even in a semiconductor pattern including a repeating pattern, it is possible to implement matching processing characterized by a learning function. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Fig. shows a structural example of the pattern matching device 100 according to Embodiment 1.

[0023] Figure 2 Fig. shows a structural example of the correlation image calculation unit 1202.

[0024] Figure 3 Fig. is a block diagram showing a structural example of the learning model 301 included in the correlation image estimation unit 1201.

[0025] Figure 4 Fig. is an example of a GUI for a user to input a matching offset.

[0026] Figure 5 Fig. is a flowchart for explaining the operation of the pattern matching device 100.

[0027] Figure 6 Fig. shows a structural example of a pattern measurement system including the pattern matching device 100 and the SEM 600.

[0028] Figure 7 Fig. shows Figure 6 Another structural example of the pattern measurement system.

[0029] Figure 8 Fig. shows a structural example of the pattern matching device 100 according to Embodiment 2.

[0030] Figure 9 Fig. shows a structural example of the learning data generation unit 8206.

[0031] Figure 10 Fig. is a flowchart for explaining the operation of the pattern matching device 100 according to Embodiment 2.

[0032] Figure 11 Fig. is a flowchart for explaining the operation of the pattern matching device 100 according to Embodiment 3. DETAILED DESCRIPTION OF THE INVENTION

[0033] Hereinafter, a pattern matching device, a pattern measurement system, and a pattern matching program of the present disclosure will be described with reference to the accompanying drawings. In the drawings, elements having the same function may sometimes be denoted by the same reference numerals or corresponding reference numerals. The drawings show embodiments and installation examples based on the principles of the present disclosure, but these drawings are for understanding the present disclosure and are not used to limitatively interpret the present disclosure. The description in this specification is merely a typical example and does not limit the claimed technical solutions or application examples of the present disclosure in any sense.

[0034] <Embodiment 1>

[0035] Figure 1 FIG. 9 shows a structural example of a pattern matching device 100 according to Embodiment 1 of the present disclosure. The pattern matching device 100 can be configured as an arithmetic device that executes pattern matching processing. The pattern matching device 100 can be composed of a storage medium 110, a pattern matching system 120, and an input device 130.

[0036] Figure 1 FIG. 13 shows a structural example in which pattern matching processing is particularly executed by an arithmetic processing device. The pattern matching processing includes, for example, a process of obtaining a matching offset amount for position alignment between a template image (SEM image 1104 obtained by a measuring device) stored in the storage medium 1103 and a search image (design drawing 1102 obtained from design data) stored in the storage medium 1101. The template image is not limited to an SEM image and may be set to other types of images such as a design drawing. In addition, the search image is not limited to a design drawing and may be set to other types of images such as an SEM image.

[0037] In the present Embodiment 1, as an example of the measuring device, a scanning electron microscope (SEM) is used. The SEM is used, for example, to measure the dimensions of patterns of semiconductor devices formed on a semiconductor wafer. A specific structural example of the SEM will be described later. Figure 6 It will be described later.

[0038] The pattern matching system 120 is composed of one or more computer subsystems including one or more CPUs (Central Processing Unit) or GPUs (Graphics Processing Unit). The pattern matching system 120 includes one or more components executed by the one or more computer subsystems. The one or more computer systems can also use one or more processors to implement the following processing in software, and can also implement a part or all of the processing by hardware such as an electronic circuit.

[0039] The pattern matching system 120 performs a pattern matching process between the design drawing 1102 stored in the storage medium 1101 and the SEM image 1104 stored in the storage medium 1103. As a result of the pattern matching process, the estimated matching offset 1301 is output. The estimated matching offset 1301 represents the positional offset or positional difference between the design drawing 1102 and the SEM image 1104.

[0040] The estimated matching offset 1301 can be represented by two-dimensional scale values, for example, (offset in the X direction, offset in the Y direction).

[0041] As an example, the pattern matching system 120 includes a correlation image estimation unit 1201, a correlation image calculation unit 1202, an estimation error calculation unit 1203, an estimation parameter update unit 1204, and a matching offset calculation unit 1205. The pattern matching system 120 is configured to receive input of various information from the input device 130.

[0042] The correlation image estimation unit 1201 estimates the estimated correlation image 1211 between the design drawing 1102 and the SEM image 1104. The matching offset calculation unit 1205 calculates the estimated matching offset 1301 using the estimated correlation image 1211.

[0043] A correlation image is an image in which correlation values between a template image and a search image are arranged over the entire search image. More specifically, an image of the same size as the template image is intercepted from the search image, and the correlation value between the intercepted image and the template image is obtained. The search image is raster scanned to intercept (sliding window) the intercepted image. An image in which the obtained correlation values are arranged as pixel values according to the displacement (sliding) in the X direction and the displacement (sliding) in the Y direction of the intercepted image is called a correlation image. If the template image and the intercepted image are more consistent, the correlation value is calculated higher (there are also cases designed to be calculated lower). Therefore, at positions where the correlation between the search image and the intercepted image is high, the pixel value becomes larger (high brightness value), and at low positions, the pixel value becomes smaller (low brightness value).

[0044] The correlation image estimation unit 1201 is composed of a learner that takes the design drawing 1102 and the SEM image 1104 as inputs and estimates the estimated correlation image 1211 between the two input images. The correlation image estimation unit 1201 also takes a model with updated estimation parameters 1212 as an input and uses this learning model to perform the estimation of the estimated correlation image 1211. The estimation parameters 1212 are appropriately updated by the estimation parameter update unit 1204 and are provided to the correlation image estimation unit 1201. The learner can be composed of, for example, the Figure 3 neural network structure described later.

[0045] The correlation image calculation unit 1202 takes the design drawing 1102 and the input matching offset 1302 as inputs, and uses these inputs to calculate the correlation image 1213. The input matching offset 1302 can be input from the input device 130. The correlation image 1213 can be used as supervised data when the learner performs learning.

[0046] The estimation error calculation unit 1203 calculates the estimation error 1214 between the estimated correlation image 1211 and the calculated correlation image 1213. Specifically, the sum of the differences between the pixel values of the estimated correlation image 1211 estimated by the correlation image estimation unit 1201 and the pixel values of the calculated correlation image 1213 serving as supervised data is obtained through an error function. For example, the error is calculated by the mean square error, the mean absolute error, etc. The error function is not limited to these, as long as it can calculate the difference between images.

[0047] The estimated parameter update unit 1204 adjusts the parameters of the learner in the correlation image estimation unit 1201 in such a way that the estimation error of each pixel value in the estimated correlation image 1211 becomes smaller, and provides it to the correlation image estimation unit 1201.

[0048] By repeating the above-mentioned estimation and backpropagation one or more times, a learning model that makes the estimated correlation image 1211 close to the calculated correlation image 1213 can be learned. Even for an image having multiple matching correct positions such as a repetitive pattern characteristic of a semiconductor pattern, there is only one correlation image, so learning can be appropriately performed.

[0049] Using the learned model provided in the correlation image estimation unit 1201, the estimated correlation image 1211 between the design drawing 1102 to be matched and the SEM image 1104 is estimated. The matching offset calculation unit 1205 calculates the estimated matching offset 1301 from the estimated correlation image 1211. As a method for calculating the matching offset, for example, the following steps can be used.

[0050] (Step 1 of calculating the matching offset)

[0051] Determine the highest pixel value (highest brightness pixel) among the pixel values of the estimated correlation image 1211. The position of the highest brightness pixel is the position where the design drawing 1102 and the SEM image 1104 match the most, and thus represents the matching offset. However, the following origin correction is required. The pixel values of the estimated correlation image 1211 are set, for example, by taking the upper left end of the estimated correlation image 1211 as the origin and calculating the correlation values at each offset position while offsetting the coordinates. On the other hand, the design drawing 1102 or the SEM image 1104 sometimes has the center position as the origin. Thus, when the origin positions are different between the estimated correlation image 1211 and the design drawing 1102, or between the estimated correlation image 1211 and the SEM image 1104, correction to make the origin positions consistent is required. Based on the position of the highest brightness pixel after the origin correction, the matching offset between the design drawing 1102 and the SEM image 1104 can be obtained.

[0052] (Step 2 of calculating the matching offset)

[0053] Instead of the pixel with the highest brightness, it is also possible to determine all pixel values above the threshold, and use the pixel value closest to the specified position (such as the image center) within the image to obtain the matching offset. In the case where origin correction is required, it is implemented in the same way as the above item 1.

[0054] As described above, the correlation image estimation unit 1201 (learned model) estimates the correlation image, and the matching offset calculation unit 1205 estimates the matching offset from this correlation image. Thus, even in a semiconductor pattern including a repeating pattern, a learning model for obtaining the matching offset can be realized.

[0055] In the learning of the correlation image, it is also possible to separate the information of the horizontal edge and the vertical edge of the pattern, and learn the correlation image (emphasis correlation image) that emphasizes the information in each direction. In this case, the learner is configured to: (a) estimate the vertical correlation image, which has the correlation value indicating the degree of consistency between the design drawing 1102 and the SEM image 1104 in the vertical direction as the pixel value; (b) estimate the horizontal correlation image, which has the correlation value indicating the degree of consistency between the design drawing 1102 and the SEM image 1104 in the horizontal direction as the pixel value. The pattern matching system 120 can generate the estimated correlation image 1211 by synthesizing the vertical correlation image and the horizontal correlation image. By separating the horizontal correlation and the vertical correlation and learning them separately, it is possible to emphasize the information in the direction with less information volume for learning. Thus, matching can be performed stably. The separation of the pattern information is not limited to the horizontal edge and the vertical edge, as long as it is the edge in the direction that can be stably learned.

[0056] Figure 2Shows a structural example of the related image calculation unit 1202. Image 211 is an example of the design drawing 1102, and image 212 is an example of the SEM image 1104. The input matching offset 1302 is the matching offset between image 211 and image 212 input by the user.

[0057] The image cropping unit 201 crops image 202 from image 211. Image 202 is an image obtained by cropping a region of the same size as image 212 from image 211 at the position specified by the input matching offset 1302. The related operation unit 203 calculates the related image 204 between image 202 and image 211. The related operation unit 203 calculates the related image 204 using, for example, methods such as normalized cross-correlation. It is also possible to preprocess image 211 and image 202 in a way that makes the related image easier to learn. The method for calculating the related image 204 is not limited to these methods, as long as the correlation value at the matching correct position is calculated to be the highest (or lowest).

[0058] Figure 3 Is a block diagram showing a structural example of the learning model 301 included in the related image estimation unit 1201. The learning model 301 can be constituted by, for example, a convolutional neural network. The learning model 301 adjusts parameters (coupling weights between neurons, biases, etc.) to reduce the difference between the calculated related image 1213 and the estimated related image 1211. The learning process is implemented through this parameter adjustment. Learning can be implemented, for example, by sequentially updating the parameters using the error backpropagation method.

[0059] Specifically, the estimated parameter update unit 1204 calculates how the error between the output data (estimated related image 1211) and the supervised data (calculated related image 1213) changes with respect to each parameter (for example, the gradient). The estimated parameter update unit 1204 updates the parameters little by little based on this change amount to adjust the parameters to be the best output.

[0060] As an example, the learning model 301 is composed of an input layer 311, an output layer 316, and multiple intermediate layers 312, 313, 314, 315.

[0061] The design drawing 1102 (the image to be searched) and the SEM image 1104 (the template image) as the input images are input to the input layer 311. From the input layer 311 to the intermediate layer 312, and from the intermediate layer 312 to the intermediate layer 313, the data within the layer is pooled through convolution operations and image reduction based on a predetermined coefficient filter. The intermediate layer 313 stores the data in which the design drawing 1102 and the SEM image 1104 are pooled.

[0062] From the intermediate layer 313 to the intermediate layer 314, the correlation data between the two is calculated by calculating the correlation value between the aggregated data of the design drawing 1102 and the aggregated data of the SEM image 1104. This correlation data is stored in the intermediate layer 314.

[0063] In the intermediate layer 314 to the intermediate layer 315 and the intermediate layer 315 to the output layer 316, the data within the layer is expanded through convolution operations and image magnification based on a predetermined coefficient filter. The data within the output layer 316 is the estimated correlation image 1211 between the design drawing 1102 and the SEM image 1104.

[0064] The estimated error calculation unit 1203 calculates the error between the estimated correlation image 1211 and the calculated correlation image 1213. The estimated parameter update unit 1204 uses this error to update the parameters (weights, biases) of each layer by the error backpropagation method.

[0065] In this way, a model that takes two images as matching objects as input and outputs a correlation image can be learned end-to-end (learning in which the input-output relationship is directly learned for the learning task). By learning end-to-end, in the intermediate layer 315, data that brings the template image closer to the searched image can be automatically aggregated. Thus, it has the effect of absorbing the deviation between the input images. In addition, the intermediate layer 315 can automatically select the data required to generate the estimated correlation image 1211 based on the correlation data stored in the intermediate layer 314. Even if there are remaining patterns in the SEM image (patterns in the SEM image although there is no design drawing), it has the effect of stably estimating the estimated correlation image 1211 that is close to the calculated correlation image 1213.

[0066] To easily absorb the deviation between the input images, multiple images (multi-channels) can be input into the template image and the searched image respectively. Multi-channel images are, for example, (a) images taken by changing the shooting conditions for the same shooting object, (b) images taken by different detectors as described later in Figure 6 and (c) images taken by a detector and images taken by an optical camera, etc. The multi-channel images are not limited to these images, and can also be images that easily absorb the deviation between the input images.

[0067] Auxiliary information can also be input to the learning model via one channel of the input image, for example, to increase the contrast of the estimated correlation image 1211 (the difference between the correlation value at the correct matching position and the correlation values at other positions). An example of auxiliary information is a weighted image that increases the brightness of the region of interest. By inputting this information, the data in the region of interest is emphasized, while the importance of data in other areas is reduced, resulting in a higher-contrast correlation image. The auxiliary information is not limited to this; any information that facilitates the estimation of a high-contrast correlation image can be used.

[0068] Figure 4 This is a diagram showing an example of a GUI (Graphical User Interface) for a user to input a matching offset. Figure 4 The illustrated GUI 401 can be displayed, for example, on Figure 1 The user can input the matching offset 1302 required for learning by the correlation image estimation unit 1201 through the GUI 401. The GUI 401 includes an image display area 402, an image transparency setting area 403, a matching offset input area 404, and a setting button 405.

[0069] Image display area 402 displays a design drawing 421, an SEM image 422, and a cursor 423. Design drawing 421 and SEM image 422 are displayed as an overlay. The user can use cursor 423 to move SEM image 422 to a position that matches design drawing 421. The amount of movement of SEM image 422 corresponds to the matching offset. The matching offset, which changes as cursor 423 is moved, is reflected in real time in matching offset setting box 441.

[0070] The user can also directly input the matching offset through the matching offset setting box 441 in the matching offset input area 404. The input matching offset is reflected by the relative movement of the SEM image 422 with respect to the design drawing 421.

[0071] The user can input the transparency (intensity) of the design drawing 421 and the SEM image 422 to the box 431 in the image transparency setting area 403 , so that the matching result can be easily checked.

[0072] When the user presses the set button 405 , the matching offset is provided as an input matching offset 1302 to the correlation image calculation unit 1202 .

[0073] use Figure 4 The method of inputting the matching offset has been described, but the method of inputting the matching offset is not limited to the method described, and any method that can input the matching offset may be used.

[0074] Figure 5 This is a flowchart illustrating the operation of the pattern matching device 100. The following describes Figure 5 each step of

[0075] ( Figure 5 : Steps S501 to S502)

[0076] The pattern matching system 120 acquires learning data (design drawing 1102, SEM image 1104) stored in a storage medium (S501). The user inputs the matching offset between the two by manually matching the design drawing 1102 and the SEM image 1104 through the Figure 4 illustrated GUI401 (S502).

[0077] ( Figure 5 : Step S503)

[0078] The correlation image calculation unit 1202 receives the design drawing 1102 and the input matching offset 1302, and calculates the calculated correlation image 1213.

[0079] ( Figure 5 : Step S504)

[0080] The correlation image estimation unit 1201 receives the design drawing 1102 and the SEM image 1104, and generates the estimated correlation image 1211. The estimation error calculation unit 1203 uses an error function to calculate the difference between the estimated correlation image 1211 and the calculated correlation image 1213, that is, the estimation error 1214 of the correlation image estimation unit 1201. The estimation parameter update unit 1204 calculates the changes in the weights and biases of the neural network by backpropagating the estimation error 1214, and updates their values. Learning is performed by repeating the estimation and backpropagation more than once in this way.

[0081] ( Figure 5 : Steps S505 to S506)

[0082] The correlation image estimation unit 1201 uses the learned model to estimate the estimated correlation image 1211 (S505). The matching offset calculation unit 1205 calculates the estimated matching offset 1301 from the estimated correlation image 1211 (S506).

[0083] Figure 6 This shows a structural example of a pattern measurement system including the pattern matching device 100 and the SEM 600. The SEM 600 measures, for example, the pattern size of semiconductor devices formed on a semiconductor wafer 603. The arithmetic processing device or computer system in the pattern measurement system can be configured as, for example, the control unit 614.

[0084] The control unit 614 includes an arithmetic unit (e.g., CPU / GPU 616) and a storage unit (e.g., a memory including an image memory 615). Information can be stored in the storage unit, such as a program related to pattern matching processing.

[0085] The CPU / GPU 616 executes this program to perform Figure 1 the pattern matching processing shown. That is, the control unit 614 functions as the pattern matching device 100. In other words, this program causes the computer system to function as the arithmetic processing device included in the pattern matching device 100 and perform Figure 1 the pattern matching processing shown.

[0086] The SEM 600 generates an electron beam from the electron gun 601. The deflector 604 and the objective lens 605 are controlled so that the electron beam is focused and irradiated at an arbitrary position on the semiconductor wafer 603, which is a sample placed on the workbench 602.

[0087] Secondary electrons are released from the semiconductor wafer 603 irradiated with the electron beam and detected by the secondary electron detector 606. The detected secondary electrons are converted into digital signals by the A / D converter 607. The image represented by the digital signals is stored in the image memory 615 within the control unit 614.

[0088] This image is used, for example, as the SEM image 1104. Based on this image, the control unit 614 or the CPU / GPU 616 performs Figure 1 the learning processing shown in the pattern matching processing shown.

[0089] The setting processing required for these processes and the display of the processing results can be performed through the input device 130.

[0090] In the alignment using an optical camera with a lower magnification than the SEM, the optical camera 611 can also be used. The signal obtained by photographing the semiconductor wafer 603 with the optical camera 611 is converted into a digital signal by the A / D converter 612 (if the signal from the optical camera 611 is a digital signal, the A / D converter 612 is not required). The image represented by the digital signal is stored in the image memory 615 within the control unit 614, and the CPU / GPU 616 performs image processing corresponding to the purpose.

[0091] The SEM600 may also be equipped with a reflected electron detector 608. When equipped with the reflected electron detector 608, the reflected electrons released from the semiconductor wafer 603 are detected by the reflected electron detector 608, and the detected reflected electrons are converted into digital signals by the A / D converter 609 or 610. The image represented by the digital signal is stored in the image memory 615 within the control unit 614, and the CPU / GPU 616 performs image processing corresponding to the purpose.

[0092] A storage unit 621 may also be provided separately from the image memory 615. The control unit 614 may control the stage 602 via the stage controller 630, and may also control the objective lens 605 etc. via the deflection control unit 631.

[0093] In Figure 6 the example of, the SEM600 is shown as an example of an inspection device used together with the pattern matching device 100, but the devices that can be used together with the pattern matching device 100 are not limited to this. Any device (measurement device, inspection device, etc.) that can acquire an image and perform pattern matching processing can be used together with the pattern matching device 100.

[0094] Figure 7 Denote Figure 6 Another structural example of the pattern measurement system of. Figure 7 The structural example of can also be understood as other expressions relative to the same structure as Figure 6 . The pattern measurement system includes: an SEM main body 701, a control device 702 that controls the SEM main body 701, an arithmetic processing device 704 that performs Figure 1 the pattern matching processing shown, a design data storage medium 705 that stores design data, and an input device 130 that inputs necessary information to the arithmetic processing device 704.

[0095] The arithmetic processing device 704 has an arithmetic unit (for example, an arithmetic processing unit 707) and a storage unit (for example, a memory 708). Information can be stored in the storage unit, for example, a program related to the pattern matching processing is stored.

[0096] The arithmetic processing unit 707 performs the pattern matching processing shown by executing this program. That is, the arithmetic processing device 704 functions as the pattern matching device 100. In other words, this program causes the computer system to function as the arithmetic processing device included in the pattern matching device 100 and perform the pattern matching processing shown by Figure 1 . Figure 1 the pattern matching processing shown.

[0097] The arithmetic processing unit 707 includes: a scenario creation unit 711 that sets the conditions of the template; a matching processing unit 712 that performs pattern matching processing based on the set template; and a pattern measurement unit 710 that performs measurement processing at the measurement positions determined by the matching processing unit 712.

[0098] Secondary electrons, reflected electrons, etc. obtained by scanning with an electron beam are captured by the detector 703, and an SEM image (corresponding to the SEM image 1104 of Figure 1 ) is generated based on the detection signal. The SEM image is sent to the arithmetic processing device 704 as the image to be searched by the matching processing unit 712 and as the measurement signal for the pattern measurement unit 710.

[0099] In Figure 7 , the control device 702 and the arithmetic processing device 704 are configured as different devices, but they can also be an integrated device.

[0100] Based on the signal of the electrons captured by the detector 703, it is converted into a digital signal by the A / D converter built in the control device 702. Based on this digital signal, image processing corresponding to the purpose is performed by the image processing hardware (CPU, GPU, ASIC, FPGA, etc.) built in the arithmetic processing device 704.

[0101] The scenario creation unit 711 includes a cropping unit 713. The cropping unit 713 reads the design data from the design data storage medium 705 and crops a part of the design data. The part cropped from the design data is determined based on pattern recognition data such as coordinate information set from the input device 130.

[0102] The scenario creation unit 711 creates pattern data for matching based on the cropped design data (layout data). The pattern data created here can be used as Figure 1 's design data 104.

[0103] The processing in the matching processing unit 712 is as described using Figure 1 . Design data, scenario information, image information, measurement results, etc. are stored in the memory 708.

[0104] Part or all of the processing in the arithmetic processing device 704 can also be implemented by an electronic computer equipped with a CPU, GPU, and a memory capable of storing images.

[0105] The input device 130 also functions as a shooting scenario creation device and creates a shooting scenario. The shooting scenario represents the measurement conditions, and includes, for example, the coordinates of the electronic components required for measurement and inspection, the types of patterns, shooting conditions (optical conditions, moving conditions of the workbench), etc.

[0106] The input device 130 may also have a function of comparing the input coordinate information, information related to the type of pattern with the layer information of the design data or the identification information of the pattern, and reading out the required information from the design data storage medium 705.

[0107] The design data stored in the design data storage medium 705 can be represented in any form. For example, it can be represented by the GDS format, the OASIS format, etc. Appropriate software for displaying the design data can display the design data in various formats based on the design data, or process it as graphic data. The graphic data can be line image information representing the ideal shape of the pattern formed based on the design data, or line image information that has been subjected to a deformation process similar to the actual pattern by performing exposure simulation on it.

[0108] It is also possible to Figure 1 pre-register the program for performing the processing described in in the storage medium, and execute the program by a control processor having an image memory and supplying the required signals to the scanning electron microscope.

[0109] <Embodiment 2>

[0110] Figure 8 FIG. shows a structural example of the pattern matching device 100 according to Embodiment 2 of the present disclosure. For the constituent elements that are the same as those in Embodiment 1 ( Figure 1 ), the same reference numerals are used in Figure 8 , and thus the repeated description is omitted below.

[0111] The pattern matching device 100 according to the present Embodiment 2 does not require manual input of the matching offset, shooting of the SEM image, etc. when generating the learning data, and can learn the learner offline without relying on the scanning microscope. Specifically, in addition to the structure described in Figure 1 , the pattern matching system 120 further includes a learning data generation unit 8206. The learning data generation unit 8206 generates a pseudo-template image (pseudo-SEM image 8216) required for learning and a generated matching offset 8215 (described later) from the searched image (design drawing 1102). The input of the learning data generation unit 8206 is not limited to the design drawing, and can also be set to other types such as SEM images.

[0112] Figure 9 FIG. shows a structural example of the learning data generation unit 8206. Image 901 is an example of the design drawing 1102.

[0113] The matching offset generation unit 904 generates a matching offset and uses it as the generated matching offset 903. As a method for generating the matching offset, for example, a method of randomly generating an offset within the possible range of the matching offset can be cited. The method for generating the matching offset is not limited to these methods and can also be a method for generating the matching offset.

[0114] The generated matching offset 903 and the image 901 are input to the image cropping unit 905, and the image cropping unit 905 crops an image 906 having the same size as the template image.

[0115] The pseudo-SEM image generation unit 907 generates a pseudo-SEM image 902 based on the image 906. The image quality of the generated pseudo-SEM image is adjusted according to the image style 908 (contrast, noise, pattern deformation, etc.). The image style 908 can be input by the user or randomly generated. The pseudo-SEM image generation unit 907 can be constituted by, for example, a model that transforms from a learned design drawing into an image, or a simulator that transforms from a design drawing into an SEM image. The pseudo-SEM image generation unit 907 is not limited to these, as long as it can transform a pseudo-image with adjustable image quality.

[0116] Using the pseudo-SEM image 8216 generated by the learning data generation unit 8206 and the design drawing 1102, the learning model of the related image estimation unit 1201 is learned as described in the first embodiment. The related image calculation unit 1202 generates a calculated related image 1213 based on the generated matching offset 8215 and the design drawing 1102.

[0117] Using the learned model possessed by the related image estimation unit 1201, the estimated related image 1211 between the design drawing 1102 to be matched and the SEM image 1104 is estimated. The matching offset calculation unit 1205 calculates the estimated matching offset 1301 from the estimated related image 1211.

[0118] As described above, by using the learning data generation unit 8206 in the learning stage, learning can be performed only with the design drawing 1102, so learning can be performed offline. Moreover, by learning using a pseudo-SEM image with adjustable style, a highly generalizable model can be learned.

[0119] Figure 10 This is a flowchart showing the operation of the pattern matching device 100 according to the second embodiment. The following describes Figure 10 each step.

[0120] ( Figure 10 : Step S1001)

[0121] The pattern matching system 120 obtains the data (design drawing 1102) stored in the storage medium, and the learning data generation unit 8206 generates learning data. The learning data consists of a pseudo-SEM image 8216 and a generated matching offset 8215.

[0122] ( Figure 10 : Step S1002)

[0123] The correlation image calculation unit 1202 receives the design drawing 1102 and the generated matching offset 8215, and calculates the correlation image 1213.

[0124] ( Figure 10 : Step S1003)

[0125] The correlation image estimation unit 1201 receives the design drawing 1102 and the pseudo-SEM image 8216, and generates an estimated correlation image 1211. The estimation error calculation unit 1203 uses an error function to calculate the difference between the estimated correlation image 1211 and the calculated correlation image 1213, that is, the estimation error 1214 of the correlation image estimation unit 1201. The estimation parameter update unit 1204 calculates the changes in the weights and biases of the neural network by backpropagating the estimation error 1214, and updates their values. By repeatedly performing such estimation and backpropagation more than once, learning is implemented.

[0126] ( Figure 10 : Steps S1004 - S1006)

[0127] After the learning is completed, the pattern matching system 120 obtains the design drawing 1102 and the SEM image 1104 (S1004). The obtained images are input to the correlation image estimation unit 1201. The correlation image estimation unit 1201 estimates the estimated correlation image 1211 between the input images (S1005). The matching offset calculation unit 1205 calculates the estimated matching offset 1301 from the estimated correlation image 1211 (S1006).

[0128] <Embodiment 2: Summary>

[0129] In the pattern matching device 100 of the present Embodiment 2, the learning data generation unit 8206 automatically generates learning data, which can save the user's effort in inputting the matching offset and obtaining images by shooting with a scanning microscope, and can learn the learner without relying on a scanning microscope.

[0130] <Embodiment 3>

[0131] In Embodiment 3 of the present disclosure, a pattern matching device 100 that can separately output the matching offsets of each layer in the matching of design data with multiple layers and SEM images will be described. The structure of the pattern matching device 100 is the same as that of Embodiments 1 to 2. Therefore, the matters related to the matching offsets of each layer and the like will be mainly described below.

[0132] In the process of forming a semiconductor pattern, due to the influence of processing accuracy, environment, etc., there may be a configuration deviation between the layers of the formed semiconductor pattern. As a result, the captured SEM image may have an interlayer deviation relative to the design drawing. In a design drawing and an SEM image with multiple layers, by separately estimating the matching offsets of each layer, there is an effect that, for example, overlay measurement can be performed in a subsequent measurement process, or the measurement position can be adjusted with high precision. The overlay measurement here is the measurement of the interlayer deviation. For example, the matching offsets of the upper layer pattern and the lower layer pattern are measured separately, and the difference is used as the interlayer deviation.

[0133] The learning process in Embodiment 3 will be described. The design drawing 1102 is a design drawing containing the design information of the upper and lower layers respectively. The SEM image 1104 is an SEM image corresponding to a part of the design drawing 1102. In the design drawing 1102, the lower layer pattern hidden under the upper layer pattern may be deleted in advance.

[0134] The design drawing 1102 and the SEM image 1104 are input to the related image estimation unit 1201, and the estimated related image 1211 is estimated. The estimated related image 1211 holds the related image of the upper layer and the related image of the lower layer in the form of a multi-channel image, for example.

[0135] The related image calculation unit 1202 calculates the calculated related image 1213 that becomes the supervision data. The calculated related image 1213 holds the calculated related image of the upper layer and the calculated related image of the lower layer in the same manner as the estimated related image 1211. The calculated related image of the upper layer is calculated using the upper layer design drawing and the upper layer matching offset as described in Embodiment 1. The upper layer matching offset is input from the input device 130 as described in Figure 4 as described. The calculated related image of the lower layer is calculated using the lower layer design drawing and the lower layer matching offset as described in Embodiment 1. The lower layer matching offset is input from the input device 130 as described in Figure 4 as described. When the lower layer pattern located under the upper layer pattern cannot be seen in the SEM image, the lower layer pattern hidden under the upper layer pattern may be deleted in advance on the design drawing.

[0136] As described in Embodiment 1, by repeating the estimation and backpropagation one or more times, a learning model that makes the estimated related image 1211 close to the calculated related image 1213 is learned.

[0137] Using the learned model included in the correlation image estimation unit 1201, estimate the estimated correlation image 1211 between the design drawing 1102 and the SEM image 1104 that are objects to be matched. The estimated correlation image 1211 includes an upper-layer estimated correlation image and a lower-layer estimated correlation image.

[0138] The matching offset calculation unit 1205 calculates the estimated matching offset 1301 from the estimated correlation image 1211. The estimated matching offset 1301 includes an upper-layer matching offset and a lower-layer matching offset. The upper-layer matching offset is calculated from the estimated upper-layer correlation image. The lower-layer matching offset is calculated from the estimated lower-layer correlation image.

[0139] As described above, by separately calculating the matching offsets of the upper and lower layers, in the subsequent measurement process, for example, there is an effect that overlay measurement can be performed or the measurement position can be adjusted with high precision.

[0140] For offline learning, the learning data generation unit 8206 described in Embodiment 2 can also be used to generate learning data. The learning data generation unit 8206 generates a pseudo-SEM image 8216 and a generated matching offset 8215 based on the design drawing 1102. When generating the pseudo-SEM image 8216, the learning data generation unit 8206 first slightly offsets the upper-layer pattern from the lower-layer pattern. Using this offset design drawing, the pseudo-SEM image 8216 is generated as described in Figure 9 For the generated matching offset, the deviation amount between the upper and lower layer patterns is added as the upper-layer matching offset and the lower-layer matching offset.

[0141] Using the pseudo-SEM image 8216 generated by the learning data generation unit 8206 and the design drawing 1102, the learning model of the correlation image estimation unit 1201 is learned as described in Figure 1 The correlation image calculation unit 1202 calculates the calculated correlation image 1213 based on the generated matching offset 8215 and the design drawing 1102.

[0142] By using the learning data generation unit 8206, learning can be performed only from the design drawing 1102, so learning can be performed offline. Moreover, by learning using the pseudo-SEM image 8216 with adjustable styles, a highly generalizable model can be learned.

[0143] Figure 11 It is a flowchart showing the operation of the pattern matching device 100 according to Embodiment 3. The following describes Figure 11 each step of

[0144] ( Figure 11 : Step S1101: Its 1)

[0145] The pattern matching system 120 acquires the learning data (design drawing 1102, SEM image 1104) stored in the storage medium. The user inputs the upper layer matching offset and the lower layer matching offset of the two manually by matching the design drawing and the SEM image in the Figure 4 illustrated GUI 401. The design drawing, the SEM image, and the matching offset are prepared as learning data.

[0146] ( Figure 11 : Step S1101: 2)

[0147] When the learning data generation unit 8206 is used, the learning data (design drawing 1102) stored in the storage medium is acquired. The learning data generation unit 8206 generates a pseudo SEM image and generates a matching offset. The design drawing, the pseudo SEM image, and the generated matching offset are prepared as learning data.

[0148] ( Figure 11 : Steps S1102 to S1103)

[0149] The correlation image calculation unit 1202 receives the upper layer part of the design drawing 1102 and the upper layer matching offset, and calculates the upper layer calculated correlation image (S1102). The correlation image calculation unit 1202 receives the lower layer part of the design drawing 1102 and the lower layer matching offset, and calculates the lower layer calculated correlation image (S1103).

[0150] ( Figure 11 : Step S1104)

[0151] The correlation image estimation unit 1201 receives the design drawing 1102 and the SEM image 1104 (pseudo SEM image in the case of using the image generation unit), and generates an estimated correlation image 1211 including the upper layer estimated correlation image and the lower layer correlation image. The estimation error calculation unit 1203 uses an error function to calculate the difference between the estimated correlation image 1211 and the calculated correlation image 1213, that is, the estimation error 1214 of the correlation image estimation unit 1201. The estimation parameter update unit 1204 calculates the changes in the weights and biases of the neural network by backpropagating the estimation error 1214, and updates their values. Learning is implemented by repeating such estimation and backpropagation one or more times.

[0152] ( Figure 11 : Steps S1105 to S1106)

[0153] After the learning is completed, the pattern matching system 120 acquires the design drawing 1102 and the SEM image 1104 (S1105). The acquired design drawing and SEM image are input into the learned model (S1106).

[0154] ( Figure 11 : Steps S1107 to S1108)

[0155] The related image estimation unit 1201 estimates the upper-layer estimation related image using the learned model (S1107). The matching offset calculation unit 1205 calculates the upper-layer matching offset from the upper-layer estimation related image (S1108).

[0156] ( Figure 11 : Steps S1109 - S1110)

[0157] The related image estimation unit 1201 estimates the lower-layer estimation related image using the learned model in the same manner as the upper-layer pattern (S1109). The matching offset calculation unit 1205 calculates the lower-layer matching offset from the lower-layer estimation related image (S1110).

[0158] ( Figure 11 : Step S1111)

[0159] The pattern matching system 120 can calculate the deviation between the previously described layers by calculating the difference (vector difference) between the calculated upper-layer matching offset and the lower-layer matching offset. Alternatively, instead of calculating the inter-layer deviation, the upper-layer matching offset and the lower-layer matching offset can be used as the measurement positions in the subsequent measurement process.

[0160] <Embodiment 3: Summary>

[0161] The pattern matching device 100 of this Embodiment 3, by separately calculating the matching offsets of the upper and lower layers, has the effect of being able to perform overlay measurement or accurately adjust the measurement position in the subsequent measurement process, for example.

[0162] <Regarding Modifications of the Present Disclosure>

[0163] The present disclosure is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments are embodiments described in detail for easy understanding of the present disclosure and are not limited to having all the structures described. Additionally, a part of the structure of one embodiment can be replaced with the structure of another embodiment, and also, the structure of another embodiment can be added to the structure of one embodiment. Further, for a part of the structure of each embodiment, other structures can be added, deleted, or replaced.

[0164] In Embodiment 3, the case of calculating the matching offset for each of the upper and lower layers has been described. However, in a specimen having three or more layers, similarly, the matching offset is calculated for each layer, and the vector difference between the respective matching offsets is obtained, thereby enabling the calculation of the positional deviation between the layers.

[0165] In the above-described embodiments, it has been described that the correlation image calculation unit 1202 calculates the correlation image (calculated correlation image 1213) between the image generated by cropping the design drawing 1102 and the design drawing 1102. This is most preferable if the correctness of the supervised data is required. On the other hand, there are also cases where supervised data with various precisions is required. In such cases, instead of or in combination with cropping the design drawing 1102, the following correlation images can be calculated and used as learning data. In the case of combination, the following correlation images can be used as a part of the learning data.

[0166] (First Variation Example of Calculated Correlation Image 1213)

[0167] At the same position as the SEM image 1104, calculate the correlation image between the design drawing of the same type as the design drawing 1102 and the design drawing 1102 as the calculated correlation image 1213. In this case, the effort of the user inputting the matching offset can be saved.

[0168] (Second Variation Example of Calculated Correlation Image 1213)

[0169] It can also be considered that 1102 is an SEM image and 1104 is an SEM image. The SEM image 1104 is an SEM image obtained by a means different from the SEM image 1102 (for example, an SEM image taken under conditions different from the shooting conditions of the SEM image 1102). In this case, calculate the correlation image between the design drawing at the same position as the SEM image 1102 and the design drawing at the same position as the SEM image 1104 (of the same type as the above design drawing) as the calculated correlation image 1213. In this case, even if there are multiple types of SEM images, it has the advantage of being able to unify the reference for calculating the calculated correlation image 1213.

[0170] (Third Variation Example of Calculated Correlation Image 1213)

[0171] In the case where 1102 is an SEM image and 1104 is an SEM image (obtained by a means different from the SEM image 1102), calculate the correlation image between the SEM image of the same type as the SEM image 1104 (such as a pseudo-SEM image generated by an image generation tool, etc.) at the same position as the SEM image 1102 and the SEM image 1104 as the calculated correlation image 1213. In this case, the effort of preparing the design drawing can be saved.

[0172] (Fourth Variation Example of Calculated Correlation Image 1213)

[0173] In the case where 1102 is an SEM image and 1104 is an SEM image (obtained by a means different from that of the SEM image 1102), a correlation image between a pseudo-SEM image (such as a pseudo-SEM image generated by an image generation tool, etc.) of the same type as the SEM image 1102 and the SEM image 1102 at the same position as the SEM image 1104 is calculated as the calculated correlation image 1213. In this case, the effort of preparing a design drawing can be saved.

[0174] (Variant example 5 of calculating the correlation image 1213)

[0175] In the case where 1102 is an SEM image and 1104 is a design drawing, a correlation image between a design drawing of the same type as the design drawing 1104 and the design drawing 1104 at the same position as the SEM image 1102 is calculated as the calculated correlation image 1213. In this case, there is an advantage that the criteria for calculating the calculated correlation image 1213 can be unified even if there are multiple types of SEM images.

[0176] (Variant example 6 of calculating the correlation image 1213)

[0177] In the case where 1102 is an SEM image and 1104 is a design drawing, a correlation image between an SEM image of the same type as the SEM image 1102 and the SEM image 1102 at the same position as the design drawing 1104 is calculated as the calculated correlation image 1213. In this case, the effort of preparing a design drawing can be saved.

[0178] Explanation of reference numerals

[0179] 100: Pattern matching device

[0180] 110: Storage medium

[0181] 1102: Design drawing

[0182] 1104: SEM image

[0183] 120: Pattern matching system

[0184] 1201: Correlation image estimation unit

[0185] 1202: Correlation image calculation unit

[0186] 1204: Estimation parameter update unit

[0187] 1205: Matching offset calculation unit

[0188] 1206: Estimation error calculation unit

[0189] 130: Input device

[0190] 8206: Learning data generation unit.

Claims

1. A pattern matching device that performs pattern matching between images, characterized in that the pattern matching device includes a computer system that outputs the offset between the first image and the second image as a pattern matching result by performing pattern matching between the first image and the second image, the computer system includes a learner that receives the first image and the second image as inputs, estimates and outputs a first correlation image that has a numerical value representing the correlation between the first image and the second image as a pixel value, the computer system calculates a second correlation image that has a numerical value representing the correlation between the first image and a derivative image generated from the first image as a pixel value, the learner is configured to be able to perform learning to reduce the difference between the first correlation image and the second correlation image, the computer system calculates the offset between the first image and the second image based on the first correlation image.

2. The pattern matching device according to claim 1, characterized in that the computer system calculates a numerical value representing the correlation between the first image and the derivative image according to the coordinate offset between the first image and the derivative image, the computer system calculates the second correlation image as an image that has the numerical value calculated according to the coordinate offset as the pixel value of the pixel corresponding to the coordinate offset.

3. The pattern matching device according to claim 1, characterized in that the computer system calculates the offset by calculating the amount by which the pixel value having a correlation between the first image and the second image in the first correlation image that is equal to or greater than a reference value deviates from the first image or the second image.

4. The pattern matching device according to claim 2, characterized in that the computer system further includes an interface that receives an input specifying the coordinate offset, the computer system calculates the second correlation image according to the coordinate offset specified by the input received by the interface.

5. The pattern matching device according to claim 1, characterized in that the computer system inputs a plurality of the first images acquired under different conditions to the learner and outputs the first correlation image representing the correlation between the second image and each of the first images, alternatively, the computer system inputs a plurality of the second images acquired under different conditions to the learner and outputs the first correlation image representing the correlation between the first image and each of the second images, alternatively, the computer system inputs a plurality of the first images acquired under different conditions to the learner and inputs a plurality of the second images acquired under different conditions, and outputs the first correlation image representing the correlation between each of the first images and each of the second images.

6. The pattern matching device according to claim 1, characterized in that On the basis that the computer system corrects at least any one of the first image and the second image so that the brightness value of a specified partial area in the first correlation image is higher than the brightness values of other areas of the first correlation image, the computer system inputs the corrected first image and the second image to the learner.

7. The pattern matching device according to claim 1, wherein the learner is configured to estimate a first-direction correlation image having, as pixel values, first-direction correlation values indicating the degree of correlation between the first image and the second image in a first direction. the learner is configured to estimate a second-direction correlation image having, as pixel values, second-direction correlation values indicating the degree of correlation between the first image and the second image in a second direction orthogonal to the first direction. The computer system generates the first correlation image by synthesizing the first-direction correlation image and the second-direction correlation image.

8. The pattern matching device according to claim 1, wherein the computer system further includes a learning data generation unit that generates learning data for the learner to perform learning. The learning data generation unit generates the derived image by shifting the coordinates of the first image, and provides the generated derived image and the first image to the learner as the learning data. The computer system calculates the second correlation image using the coordinate offset amount used by the learning data generation unit to generate the derived image, or the derived image generated by the learning data generation unit.

9. The pattern matching device according to claim 8, wherein the learning data generation unit provides, as the learning data, a deformed image generated by changing at least any one of the image quality of the derived image and the shape pattern included in the derived image to the learner.

10. The pattern matching device according to claim 1, wherein the computer system is configured to perform pattern matching on an image of a specimen having an upper layer and a lower layer. The first image includes a first upper-layer image of the specimen and a first lower-layer image of the specimen. The second image includes a second upper-layer image of the specimen and a second lower-layer image of the specimen. The learner receives the first upper-layer image and the second upper-layer image as inputs, and estimates and outputs a first upper-layer correlation image having, as pixel values, numerical values indicating the correlation between the first upper-layer image and the second upper-layer image. The learner receives the first lower-layer image and the second lower-layer image as inputs, and estimates and outputs a first lower-layer correlation image having, as pixel values, numerical values indicating the correlation between the first lower-layer image and the second lower-layer image. The computer system calculates an upper-layer offset amount between the first upper-layer image and the second upper-layer image based on the first upper-layer correlation image. The computer system calculates a lower-layer offset between the first lower-layer image and the second lower-layer image based on the first lower-layer related image.

11. The pattern matching device according to claim 10, wherein the computer system calculates and outputs a deviation amount between the upper layer and the lower layer of the specimen based on the upper-layer offset and the lower-layer offset.

12. The pattern matching device according to claim 1, wherein the computer system calculates a third correlation image having a numerical value representing the correlation between the first image and a third image of the same type as the first image at a position corresponding to the second image as pixel values. The learner is configured to be able to perform learning to reduce the difference between the first correlation image and the third correlation image.

13. The pattern matching device according to claim 1, wherein the computer system calculates a fourth correlation image having a numerical value representing the correlation between a fourth image of the same type as the second image at a position corresponding to the first image and the second image as pixel values. The learner is configured to be able to perform learning to reduce the difference between the first correlation image and the fourth correlation image.

14. The pattern matching device according to claim 1, wherein the computer system calculates a fifth correlation image having a numerical value representing the correlation between a fifth image corresponding to the position of the first image and a sixth image of the same type as the fifth image at a position corresponding to the second image as pixel values. The learner is configured to be able to perform learning to reduce the difference between the first correlation image and the fifth correlation image.

15. A pattern measurement system, characterized in that, comprising: the pattern matching device according to claim 1; and a scanning electron microscope that obtains the first image and the second image by photographing a specimen and provides them to the pattern matching device.

16. A pattern matching program that causes a computer system to perform pattern matching between images, wherein the pattern matching program causes the computer system to perform a step of outputting the offset between the first image and the second image as a pattern matching result by performing pattern matching between the first image and the second image, in the step of performing the pattern matching, the computer system performs a step of receiving the first image and the second image as inputs, providing the first image and the second image to a learner that estimates and outputs a first correlation image having a numerical value representing the correlation between the first image and the second image as pixel values, and receiving the first correlation image, in the step of performing the pattern matching, the computer system performs a step of calculating a second correlation image having a numerical value representing the correlation between the first image and a derivative image generated from the first image as pixel values, the learner is configured to be able to perform learning to reduce the difference between the first correlation image and the second correlation image. In the step of performing the pattern matching, the computer system performs the step of calculating an offset between the first image and the second image based on the first correlation image.

Images