System for generating an image and non-transitory computer readable medium

Abstract

To provide an image generating system capable of efficiently learning specific parts.SOLUTION: An image generating system is a system configured to generate converted images based on the input images generated by a charged particle beam device. The image generating system includes: one or more computer subsystems; and one or more components that are executed by one or more computer subsystems. One or more components include: first images that are input as teacher data during learning; and a learning model with parameters adjusted so that the input image of the system is converted by the input of the second image to suppress an error. The first image is a semiconductor pattern image that has been subjected to image correction using different parameters in at least two or more regions of the image, and the at least two or more regions are divided by the image segmentation.SELECTED DRAWING: Figure 11

Description

[Technical field]

[0001] The present disclosure relates to a method, system, and non-transitory computer readable medium for generating an image, and in particular to a method, system, and non-transitory computer readable medium for generating an image based on irradiation of a beam on a sample. [Background technology]

[0002] Patent Document 1 discloses a sample observation device that generates an image based on a detection signal obtained based on irradiation of a sample with a charged particle beam. Patent Document 1 discloses a sample observation device that generates a high-quality image from a low-quality image using a neural network such as deep learning, and the sample observation device includes an image conversion model for generating a high-quality image from a low-quality image. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2018-137275 A Summary of the Invention [Problem to be solved by the invention]

[0004] On the other hand, neural networks use two images, the image converted from a low-quality image and the high-quality image, as training data and proceed with learning to match the two images, so there are cases where the quality of specific parts, such as the edges of objects to be measured or inspected, cannot be sufficiently improved. Also, if learning is performed to match the entire image, it may take a lot of time to generate a model. [Means for solving the problem]

[0005] A method, a system, and a non-transitory computer-readable medium for efficiently learning a specific portion are described below. As one aspect for achieving the above object, a system configured to generate a converted image based on an input of an input image generated by a charged particle beam device is proposed, the system comprising one or more computer subsystems and one or more components executed by the one or more computer subsystems, the one or more components including a learning model with parameters adjusted to suppress an error between a first image input as training data during learning and a second image converted by inputting the input image to the system, the first image being a semiconductor pattern image, the first image being subjected to image correction with different parameters in at least two or more regions within the image, and the at least two or more regions being regions divided by image segmentation. Effect of the Invention

[0006] According to the present invention, different learning can be performed on each divided region (for example, wiring and circular structures), and therefore it is possible to generate an image according to the operator's preferences from a low-quality image. [Brief description of the drawings]

[0007] [Figure 1] FIG. 1 is a diagram showing an example of a scanning electron microscope (SEM) which is a type of image generation tool that generates an image to be input to the image generation system of the first embodiment. [Diagram 2] FIG. 2 is a diagram showing an example of an image generation system that generates a converted image from image data obtained by an image generation tool such as that shown in FIG. [Diagram 3] 1 shows an example of a neural network configuration. [Figure 4]1A and 1B are diagrams showing an example of a pattern to be scanned by a charged particle beam, an image obtained by scanning the pattern with the beam, and a signal waveform formed from luminance distribution information of the image. [Diagram 5] FIG. 13 is a diagram showing an example of a GUI screen that enables setting of image division conditions and weights of loss functions (error adjustment conditions) for divided regions. [Figure 6] FIG. 13 is a diagram showing an example of an image subjected to region division using luminance distribution information. [Figure 7] 4 is a flowchart showing a process of setting learning conditions for a learning module using the system of the first embodiment (FIG. 2), learning according to the set learning conditions, and generating an image using the trained learning module. [Figure 8] This is a conceptual diagram illustrating an example of performing learning according to the importance of each region by setting a different weighting coefficient for each position having a different relative positional relationship with the alignment pattern, and an example of setting a weighting coefficient λ3 for ROI 805 within the field of view 804 and a weighting coefficient λ4 for other regions 806 in the recipe in advance. [Figure 9] 1 shows a scanning electron microscope according to a modified example of the first embodiment. [Figure 10] 13 is a flowchart illustrating the operation of a modified example of the first embodiment. [Figure 11] FIG. 11 is a block diagram showing an example of an image generating system according to a second embodiment. [Figure 12] This figure shows an example of a high-quality image input to the computer system 202 illustrated in Figure 11, a region division image obtained by dividing the high-quality image by the image division processing unit 2060, and a brightness-adjusted image obtained by performing brightness adjustment on each divided region of the region division image by the region-by-region brightness correction unit 2070. [Figure 13] 13 shows an example of a GUI screen for setting a luminance correction condition in the second embodiment. [Figure 14] 10 is a flowchart illustrating an operation of the second embodiment. [Figure 15] FIG. 13 is a conceptual diagram illustrating the operation of a system according to a third embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] Hereinafter, the present embodiment will be described with reference to the accompanying drawings. In the accompanying drawings, functionally identical elements may be indicated by the same or corresponding numbers. Note that the accompanying drawings show embodiments and implementation examples according to the principles of the present disclosure, but these are for understanding the present disclosure and are not to be used to interpret the present disclosure in a limiting manner. The description in this specification is merely a typical example and does not limit the scope or application of the present disclosure in any sense.

[0009] In the present embodiment, the present disclosure is described in sufficient detail for a person skilled in the art to implement the present disclosure, but it should be understood that other implementations and forms are possible, and that changes in configuration and structure and substitutions of various elements are possible without departing from the scope and spirit of the technical ideas of the present disclosure. Therefore, the following description should not be interpreted as being limited thereto.

[0010] The embodiments described below relate to a method, system, and computer-readable medium for converting a first image into a second image having image quality different from that of the first image, and describe a method, system, and computer-readable medium for performing different conversions on multiple regions contained in an image.

[0011] [First embodiment] An image generating system according to a first embodiment will be described with reference to FIGS. 1 is a diagram showing an example of a scanning electron microscope (SEM), which is a type of image generation tool that generates an image input to the image generation system according to the first embodiment. In the following description, an SEM, which is a type of charged particle beam device, is described as an image generation tool, but the image generation tool is not limited to this, and for example, a focused ion beam device that generates an image based on the scanning of an ion beam can also be used as the image generation tool. Furthermore, it is also possible to use other image generation tools that can simplify processing by generating a low-quality image rather than a high-quality image.

[0012] 1 includes an imaging unit 101, a computer system 102, a signal processing unit 103, an input / output unit 104, and a storage unit 105. The storage unit 105 also functions as a non-transitory recording medium that stores a computer program that controls the operation of the system. The computer system 102 controls the following optical systems included in the imaging unit 101:

[0013] The imaging unit 101 includes an electron gun 106 that irradiates an electron beam 107, a focusing lens 108 that focuses the electron beam 107, and a focusing lens 109 that further focuses the electron beam 107 that has passed through the focusing lens 108. The imaging unit 101 further includes a deflector 110 that deflects the electron beam 107, and an objective lens 111 that controls the height at which the electron beam 107 is focused.

[0014] The electron beam 107 that has passed through the optical system of the imaging unit 101 is irradiated onto a sample 112 placed on a sample stage 113. Emitted electrons 114, such as secondary electrons (SE) and backscattered electrons (BSE), emitted from the sample 112 by irradiation with the electron beam 107 are detected by a lower detector 115 and an upper detector 116 that are installed on the trajectory of the sample. An aperture provided in the upper detector 116 allows the electron beam 107 to pass through. By making this aperture sufficiently small, it is possible to detect secondary electrons that are emitted from the bottom of a deep hole or deep trench formed on the sample 112, pass near the center of the pattern, and escape onto the sample surface. The emitted electrons 114 can be energy-discriminated by energy filtering using an energy filter 117a located immediately before the upper detector 116 or an energy filter 117b located immediately before the lower detector 115.

[0015] The imaging section 101 further includes a blanking deflector 118 that limits the electron beam 107 from reaching the sample 112 by deflecting the electron beam 107 off the optical axis, and a blanking electrode 119 that receives the electron beam 107 deflected by the blanking deflector 118.

[0016] The signal processing unit 103 generates an SEM image based on the outputs of the lower detector 115 and the upper detector 116. The signal processing unit 103 generates image data by storing the detection signal in a frame memory or the like in synchronization with the scanning of a scanning deflector (not shown). When storing the detection signal in the frame memory, the detection signal is stored at a position corresponding to the scanning position of the frame memory, thereby generating a signal profile (one-dimensional information) and an SEM image (two-dimensional information). In addition, by deflecting the secondary electrons with the deflector 120 as necessary, the secondary electrons passing near the optical axis after escaping from a deep hole or the like can be guided outside the opening of the lower detector 115 (the detection surface of the lower detector 115).

[0017] Fig. 2 is a diagram showing an example of an image generation system that generates a converted image from image data obtained by an image generation tool as exemplified in Fig. 1. The computer system 102 exemplified in Fig. 2 is composed of one or more computer subsystems including one or more CPUs and GPUs. The computer system 102 exemplified in Fig. 2 also includes one or more components executed by the one or more computer subsystems. The one or more computer subsystems may realize the processes described below by software using one or more processors, or may realize part or all of the processes by hardware such as electronic circuits.

[0018] 2 includes, for example, an image conversion unit 2030, a conversion error calculation unit 2040, a conversion parameter update unit 2050, and an image segmentation processing unit 2060. The computer system 202 is also configured to receive various types of information from an input device 203.

[0019] The image conversion unit 2030 takes an output image of an image generation tool as exemplified in FIG. 1 or an image (low-quality image) stored in the storage medium 201 as an input image, and generates a converted image based on the input image. The image conversion unit 2030 includes a learning model in which conversion parameters are adjusted, and executes image conversion using this learning model. The conversion parameters are appropriately updated by a conversion parameter update unit 2050 and supplied to the image conversion unit 2030. The learning model is composed of, for example, a neural network, and includes one or more input layers, one or more intermediate layers (hidden layers), and one or more output layers as exemplified in FIG. 3.

[0020] Neural networks can achieve proper output by learning to adjust parameters (weights, biases, etc.) so that the output layer produces the desired results (such as high-quality images or correct measurement values). Learning is performed by sequentially updating variables (weights and biases) using, for example, backpropagation, and the output error of the data is partially differentiated with the weights (including the activation function) to gradually adjust the output to the optimal value.

[0021] The conversion error calculation unit 2040 calculates the error between a converted image (a second image which is the output of the output layer) generated from a low-quality image by the image conversion unit 2030 and an image (a first image) input as correct answer data (a high-quality image (teacher data)). More specifically, the conversion error calculation unit 2040 calculates a mean absolute error, a mean square error, or the like, as a conversion error, from the pixel values ​​of the converted image derived by forward propagation and the pixel values ​​of the corresponding correct answer image. The conversion parameter update unit 2050 adjusts the conversion parameters (variables) of the neural network based on the conversion error so as to suppress the conversion error of each pixel, and supplies the result to the image conversion unit 2030.

[0022] The image division processing unit 2060 divides a high-quality image as training data into a plurality of images, and supplies region information relating to the division to the conversion error calculation unit 2040. The region information and weight information can be input from the input device 203.

[0023] By repeating the above-mentioned forward propagation and back propagation, the accuracy of the output can be improved, but until the weights for the neuron input are optimized, learning using many images (teaching data) is required. On the other hand, measurement and inspection of semiconductor device patterns are performed by measuring the dimensions between the edges of patterns contained in images and evaluating the shapes of patterns, so high accuracy is not required for parts other than edges, for example.

[0024] In this embodiment, a learning model capable of selectively converting a specific pattern contained in an image or a partial image such as the edge of a structure with high accuracy, and a system for performing learning on the learning model will be described. If the degree of learning can be changed according to the importance of each part of the image, rather than the entire image, it is possible to generate a high-quality image by performing advanced learning on important parts, while it is possible to improve the efficiency of learning by reducing the processing required for learning on unimportant parts.

[0025] In the system illustrated in FIG. 2, a high-quality image input as training data is divided by an image division processing unit 2060, and the image conversion unit 203 performs learning on a learning device according to the degree of learning for each divided area input by an input device 203. The degree of learning for each divided area is set, for example, by assigning a different weight to each divided area as a loss function, which is a function for calculating the error between the correct answer data (label) and the output of the learning device. For example, the loss function L is calculated based on an arithmetic expression such as [Equation 1]. D can be calculated.

[0026] [Number 1] L D =(λ1 L bg +λ2·L edge )

[0027] Here, L bg is the loss function set for the first region in the image (e.g., the background part other than the pattern edge), and L edge is the loss function set for the second region in the image (e.g., the pattern edge). λ1 and λ2 are the weighting coefficients of the respective loss functions. For example, by setting the coefficient λ2 for the edge portion to be larger than λ1, the error value for the region other than the edge portion is estimated relatively large, and as a result, it becomes possible to update the conversion parameters so as to focus on the edge portion and suppress the conversion error. A similar effect can be achieved by L bg is the loss function set for the entire image including the edge parts, L edge It can also be obtained by setting the loss function set only to the edge portion and setting λ1 and λ2 so that the error value of the edge portion is reflected more than the ratio of the total number of pixels of the actual edge to the total number of pixels of the image.

[0028] On the other hand, for parts other than edges that are not required for measurement or inspection, no learning is performed or learning is performed to a relatively low degree, so that the processing required for learning can be reduced by setting λ1<λ2 (λ1 includes zero). That is, in this first embodiment, in the learning model, different learning is performed for each different area in the image. By relatively increasing the weight of the loss function for the area required for measurement or inspection in the image and decreasing the weight of the loss function for the other areas, it is possible to improve the reproducibility of the required parts.

[0029] Next, an outline of the region division process in the image division processing unit 2060 illustrated in Fig. 2 will be described. Fig. 4 is a diagram showing an example of a pattern to be scanned with a charged particle beam, an image obtained by scanning the pattern with the beam, and a signal waveform formed from luminance distribution information of the image.

[0030] A charged particle beam is two-dimensionally scanned on a line pattern 401 having a cross-sectional shape as shown in Fig. 4(a) and the charged particles obtained by the scanning are detected, thereby forming a two-dimensional image 402 as shown in Fig. 4(b). By further performing a projection process (adding and averaging the signals of each pixel row), a signal waveform (profile) 404 as shown in Fig. 4(c) can be generated.

[0031] When a beam is scanned over a pattern having a cross section as shown in FIG. 4(a), the amount of charged particles emitted from a position 403 corresponding to the edge of the pattern is greater than the amount of charged particles emitted from other regions due to the edge effect. As described above, in measurements and inspections using images, the brightness information of the edge portion is important. For this reason, as shown in FIG. 4(c), a threshold Th (405) is set, and the coefficient for the conversion error of the region obtained from the region where the brightness exceeds a certain threshold or the region where the brightness exceeds the threshold and has a margin of several pixels is set to λ2, and the coefficient for the conversion error obtained from the other regions is set to λ1. This makes it possible to generate a learning model with excellent reproducibility of the edge portion.

[0032] In addition, when performing measurements based on identifying edge portions of patterns, it is desirable to extract not only the area extracted as an area having a brightness above a certain threshold, but also the surrounding area as the edge area. In particular, when measuring the dimension between edges, the reference position for measurement is determined using the brightness distribution information of the white band (high brightness area corresponding to the edge), so the brightness value around the white band also affects the measurement value. For this reason, it is desirable to use image data extracted from the automatically extracted white band area and the surrounding area of ​​a certain number of pixels (peripheral area) as training data. A specific method for extracting image data including the peripheral area will be described later with reference to Figure 6.

[0033] Fig. 5 is a diagram showing an example of a GUI screen that enables the setting of image division conditions and weights of loss functions (error adjustment conditions) for the divided regions. Note that Fig. 5 illustrates an example of a GUI that sets a weighting coefficient λ for each divided region, but the present invention is not limited to this, and other parameters that can adjust the processing load required for learning may be input.

[0034] The GUI screen illustrated in Fig. 5 can be displayed, for example, on the display device of the input device 203 illustrated in Fig. 1. Such a GUI screen enables the backpropagation conditions (learning conditions) to be set by the computer system 202 (computer subsystem). The GUI screen 501 illustrated in Fig. 5 includes an SEM image display area 502 and a loss function condition setting area 503.

[0035] Image segmentation conditions can be selected in the loss function condition setting area 503. Specifically, the loss function setting area 503 is provided with selection buttons for ROI (Region Of Interest) setting, which is selected when setting a desired area on the GUI, Auto Segmentation, which is selected when performing area division automatically, and Area division by brightness, which is selected when performing area division automatically according to brightness.

[0036] For example, when ROI setting is selected, an ROI setting frame 505 is displayed in the SEM image display area 502, which allows the user to set an ROI at any position and with any size using a pointer 504. After using the pointer 504 to set the position and size of the ROI, the user can set the degree of learning load for the ROI by setting the desired weighting coefficients (λ1, λ2) in the weighting coefficient setting frame 506. For areas not set as ROI, the weighting coefficients for Background (λ g ), you can select the weighting factor for the region other than the selected region. As mentioned above, the weighting factor (λ g ), by setting the weighting coefficients (λ1, λ2) of the region selected as the ROI relatively high, it is possible to improve the reproducibility and quality of the image within the ROI.

[0037] Alternatively, selecting Auto Segmentation causes the computer system 202 to automatically perform segmentation, for example by semantic segmentation.

[0038] Furthermore, when "Area division by brightness" is selected, the computer system 202 executes area division according to the brightness information in the image. Specifically, the area division is performed in the image by n-value processing (n≧2) according to the brightness. By providing a weighting coefficient setting field 507 for setting a weighting coefficient of the loss function for each area divided by selecting Auto Segmentation or "Area division by brightness", it is possible to set an appropriate weight for each area.

[0039] In this embodiment, an example will be described in which an image is divided and a loss function condition is defined according to the importance of the divided region, but the present invention is not limited to this. For example, it is also possible to selectively learn only the ROI and not learn other regions. That is, it is also possible to set the input device 203 in FIG. 1 not to input information related to division (region information, etc.) and not to execute division processing in the image division processing unit 203. When it is desired that only the edge portion used for measurement or inspection be a high-quality image, the portion other than the ROI may be a low-quality image, and such selective learning may be performed.

[0040] On the other hand, if the degree of processing required for learning can be set for each region as in this embodiment, it is possible to set appropriate learning conditions according to the purpose, taking into account the reproducibility of the image and the processing time required for learning. Furthermore, if an attempt is made to optimize an area other than the ROI, this may be a factor that hinders the improvement of the quality of the reproduced image within the ROI, but by enabling learning that improves the quality of the ROI, it is possible to improve the quality of the reproduced image at an early stage. Note that, instead of semantic segmentation, other division processing methods such as the K-means method can also be applied.

[0041] FIG. 2 illustrates a system that outputs a high-quality image by inputting a low-quality image, where the low-quality image is, for example, a small-frame image (one frame is, for example, one two-dimensional scan). In a scanning electron microscope or the like, in order to improve image quality, the same field of view (FOV) is scanned multiple times, and the obtained detection signals are integrated (averaged), thereby improving the S / N ratio of the image. However, multiple scanning of the beam may cause charge accumulation and pattern shrinkage. In this embodiment, a low-quality image is treated as a one-frame image, and a high-quality image serving as teacher data is treated as a multi-frame image such as 64 frames, thereby reproducing a high-quality image from a low-quality image.

[0042] Note that performing beam scanning over multiple frames may cause the pattern to shrink. For this reason, in this embodiment, for example, beam scanning over 64 frames is performed to obtain an image signal required to generate a high-quality image, and then a small number of frame images (for example, a one-frame image) are obtained. Then, this small number of frame images are converted into a converted image by the image conversion unit 2030, and a high-quality image such as the previously obtained 64-frame image is used as training data to calculate the conversion error between the two. This allows the learning device to learn while suppressing the effects of shrinkage and the like.

[0043] Incidentally, since the amount of shrinkage of the pattern is large at the beginning of beam scanning, it is advisable to start acquiring signals for generating multi-frame images once the shrinkage has settled down to a certain extent (for example, when scanning of n frames has been completed since the start of scanning), and to generate a high-quality image based on the detection signals thus obtained.

[0044] Fig. 6 is a diagram showing an example of an image in which region division is performed using luminance distribution information. When luminance distribution information is extracted from an image as shown in Fig. 6 by projection or the like, a signal waveform as shown in Fig. 4(c) can be generated. Here, a predetermined region based on the center in the x direction of the region exceeding a predetermined threshold, or a region extended by n pixels (n is any natural number) from the end of the region exceeding a predetermined threshold in the x direction is set as ROI 601, and it is possible to set a loss function condition (weighting coefficient, etc.) specific to that part, thereby improving the quality of not only high luminance parts but also parts that can be expressed as peak waveforms.

[0045] FIG. 7 is a flowchart showing the process of setting learning conditions for a learning module using the system of the first embodiment (FIG. 2), learning according to the set learning conditions, and generating an image using the trained learning module.

[0046] First, the computer system 202 acquires an output image of an image generation tool as shown in Fig. 1 or an image stored in the storage medium 201 (step S701). Meanwhile, the image segmentation processing unit 2060 executes region segmentation processing on the image according to region information specified by the input device 203 or the like (step S702). The image segmentation processing unit 2060 executes image segmentation based on image segmentation conditions set on a GUI screen as shown in Fig. 5, for example.

[0047] Next, the weighting coefficient of the loss function, which is one of the parameters of the load required for learning, is set for each divided region on the GUI screen shown in FIG. 5 or the like, thereby setting the learning conditions for each region (step S703).

[0048] Next, the image conversion unit 2030 generates a converted image (forward propagation) by inputting the low-quality image to the computer system 202 including the learning device. Furthermore, the conversion error calculation unit 2040 obtains the pixel-by-pixel difference between the converted image and a high-quality image input to the computer system 202 separately from the low-quality image, and calculates the error between the image generated by the forward propagation and the high-quality image which is the correct image. Here, the conversion parameter update unit 2050 executes backpropagation using the weighting coefficients assigned to each image region and the loss function of each region, calculates the changes in the weights and biases of the neural network, and updates the values ​​(step S704).

[0049] Learning is performed by repeating the above forward propagation and back propagation one or more times. It is also possible to apply an evolutionary algorithm as a feedback method.

[0050] As described above, the trained model has been subjected to selective advanced learning for specific regions, making it possible to perform estimation with good reproducibility (image conversion, step S705) while minimizing the effort and time required for training.

[0051] Next, an example will be described in which learning is performed according to the importance of each region by setting a different weighting factor for each position having a different relative positional relationship with the alignment pattern.

[0052] 8(a) is a diagram showing the positional relationship between an alignment pattern 801 formed on a sample such as a semiconductor wafer, a first field of view 802, and a second field of view 803. When positioning the field of view at a desired position using a scanning electron microscope such as that shown in FIG. 1, for example, a stage provided in the scanning electron microscope is moved so that the field of view position is positioned at the coordinates of the alignment pattern 801 or the desired position, and a low-magnification image is generated by scanning a beam over a wider range than the field of view at the desired position. After acquiring the low-magnification image, the position of the alignment pattern 801 is identified by performing pattern matching using a template image including the alignment pattern 801 registered in advance, and the field of view is moved to the first field of view 802 and the second field of view 803 that are in a known positional relationship with the alignment pattern 801.

[0053] The scanning electron microscope illustrated in Fig. 1 is controlled by an operating program called a recipe, and is controlled by one or more computer systems to move the field of view based on information on the distance ((-x1, -y1), (x2, y2)) between a preregistered alignment pattern 801 and a desired field of view position. As illustrated in Fig. 8(a), by defining the weighting coefficients (λ1, λ2) of the loss function according to the importance of each field of view position, it is possible to realize efficient learning.

[0054] Furthermore, as shown in FIG. 8B, a weighting factor λ3 for an ROI 805 within a field of view 804 and a weighting factor λ4 for a region 806 other than the ROI 805 may be set in advance in a recipe.

[0055] 9 and 10 show a modification of the first embodiment. FIG. 9 is a diagram showing an example of a scanning electron microscope equipped with BSE detectors 901 and 902 and an SE detector 908 according to a modified example of the first embodiment. The configuration other than that shown in FIG. 9 can be the same as that of the first embodiment (FIG. 1). In the example of FIG. 9, a negative voltage is applied to the sample 112 in order to achieve both high speed of the electron beam passing through the objective lens 111 and low acceleration of the electron beam reaching the sample 112. By applying the negative voltage to the sample 112, a deceleration electric field that decelerates the beam reaching the sample 112 is formed between the objective lens 111 and the sample 112. In such an optical system, secondary electrons 907 emitted from the sample 112 are accelerated almost along the beam optical axis 904 and deflected toward the SE detector 908 by, for example, a Wien filter 906.

[0056] On the other hand, backscattered electrons 903, which are accelerated more quickly than secondary electrons 907, are relatively unaffected by the deceleration electric field and are detected by BSE detectors 901, 902 arranged closer to the sample 112 than the SE detector 908. When the backscattered electrons 903 collide with, for example, a sidewall of a pattern, they are reflected in a direction inclined to the surface of the sample 112 and are detected by the BSE detectors 901, 902 after being focused by the objective lens 111. By arranging the BSE detectors 901, 902 at positions as shown in Fig. 9, it is possible to selectively detect backscattered electrons reflected by edges of the pattern, etc., and thus it is possible to clarify the contrast between the edges and other parts.

[0057] In the modification of the first embodiment, such a phenomenon is utilized to perform image region segmentation by using luminance distribution information of the BSE image, and the learning process of the learning device is executed. Fig. 10 is a flowchart showing the learning process of the learning device that performs learning by using both the BSE image and the SE image.

[0058] First, a BSE image and an SE image are obtained based on detection signals from the SE detector 908 and the BSE detectors 901 and 902 by a scanning electron microscope having an optical system as exemplified in FIG. 9 (steps S1001 and S1002).

[0059] Next, the image segmentation processing unit 2060 illustrated in FIG. 2 executes image segmentation processing using the luminance distribution information of the BSE image (step S1003). As described above, backscattered electrons reflected by the edges of the pattern, etc. are detected by the BSE detectors 901 and 902, while backscattered electrons emitted in other directions are not detected. Therefore, the BSE image has a clear contrast between the edge portion and the other portions. Therefore, the image segmentation processing unit 2060 extracts, for example, a high luminance portion above a predetermined threshold in the BSE image, and sets the portion, or a portion obtained by expanding the high luminance portion by several pixels, as the ROI. In addition, the other portions are defined as the background.

[0060] The computer system 202 sets information on the loss function (weighting coefficients in the case of this embodiment) input from a GUI screen such as that shown in FIG. 5 as loss function information corresponding to each divided region (step S1004), and stores the information in, for example, the storage unit 105. When an image is divided into two regions, an ROI and a background, as described above, the weighting coefficients λ1 and λ g Set.

[0061] The computer system 202 performs learning of the learning device by performing different learning (one or more repetitions of forward propagation and back propagation) for each region using the weight coefficients set for each region as described above (step S1005). According to the method illustrated in FIG. 10, efficient learning can be performed.

[0062] [Second embodiment] Next, an image generation system according to a second embodiment will be described with reference to Fig. 11 to Fig. 14. The block diagram of Fig. 11 shows an example of an image generation system according to the second embodiment. The same components as those in the first embodiment (Fig. 2) are given the same reference numerals in Fig. 11, and therefore duplicated descriptions will be omitted below.

[0063] The system illustrated in Fig. 11 is a system that makes it possible to generate a better-looking image rather than simply producing a high-quality image. Specifically, the system in Fig. 11 includes a per-region luminance correction unit 2070 in addition to the components of the system in Fig. 2. The per-region luminance correction unit 2070 is configured to perform luminance correction processing for each of the images divided by the image division processing unit 2060. In other words, the per-region luminance correction unit 2070 is configured to correct the luminance value so that each divided region has a predetermined luminance.

[0064] Fig. 12 is a diagram showing an example of a high-quality image input to the computer system 202 shown in Fig. 11, an area division image obtained by dividing the high-quality image by the image division processing unit 2060, and a luminance-adjusted image obtained by performing luminance adjustment on each of the divided areas of the area division image by the area-by-area luminance correction unit 2070. The high-quality image 1201 shown in Fig. 12 is an image of a semiconductor device as an example, and includes upper layer wiring 1202, lower layer wiring 1203, a circular structure 1204, and a circular structure 1205 overlapping with the wiring, etc.

[0065] The image division processing unit 2060 generates a region-divided image by performing, for example, semantic segmentation or region division according to brightness. When performing region division according to brightness, it is possible to perform processing such as dividing the image for each predetermined brightness range. For example, the upper layer pattern 1202 included in the high-quality image 1201 is often brighter than the lower layer pattern 1203. This is because secondary electrons are more likely to be emitted closer to the sample surface, and the amount of secondary electrons detected is relatively large. Furthermore, the shape of the pattern and its positional relationship with other structures may be converted into feature quantities, and classification may be performed according to the feature quantities. In the example of FIG. 12, an example is shown in which the image is divided into five regions A to E. Specifically, in FIG. 12, one image is divided into five regions: (1) region A corresponding to the upper layer wiring 1202, (2) region B corresponding to the lower layer wiring 1203, (3) region C corresponding to the circular structure 1204, (4) region D corresponding to the circular structure 1205 overlapping with the upper layer wiring 1202, and (5) region E which is the background.

[0066] For each region divided as described above, the region-by-region luminance correction unit 2070 performs luminance correction based on luminance correction information input from the input device 203 or the like. The conditions for luminance correction may be set on a GUI screen, for example, as shown in Fig. 13. The GUI screen 1301 shown in Fig. 13 is configured to enable luminance adjustment for each region divided by the computer system 202.

[0067] The GUI screen 1301 is provided with a divided image display area 1302 and a brightness condition setting field 1303. The divided image generated by the image division processing unit 2060 is displayed in the divided image display area 1302. The brightness condition setting field 1303 is provided with a slide bar 1304 for adjusting the brightness of each divided area, a brightness setting field 1305 for setting the brightness numerically, and the like.

[0068] The initial brightness of the divided image is displayed in the brightness setting field 1305, and the operator can adjust each area to a desired brightness using the brightness setting field 1305 or the slide bar 1304. In addition, by adjusting the display so that the brightness of each area of ​​the divided image display area 1302 changes in conjunction with the adjustment of the brightness using the slide bar 1304 or the like, the operator can set conditions while checking the appearance of the converted image.

[0069] In the second embodiment, an example in which the brightness of each area is adjusted manually will be described, but it is also possible to provide a mode in which light and dark are inverted so that, for example, dark areas are brightened and bright areas are darkened, and to adjust the brightness automatically. Also, a mode in which circular structures are made highly bright and other areas are made low may be prepared and used depending on the purpose of measurement or inspection.

[0070] Fig. 14 is a flowchart showing a process of setting the learning conditions of the learning device and a learning process in the system shown in Fig. 11. According to this flowchart, an image can be generated in which the luminance of any region included in the image is selectively changed.

[0071] 11 first acquires a high-quality image as training data, and generates a segmented image using the image segmentation processing unit 2060 (steps S1401 and S1402). The segmented image is displayed on a GUI screen or the like displayed on a display device or the like of the input device 203.

[0072] The operator sets appropriate luminance correction conditions while referring to the divided image displayed in the divided image display area 1302 and the brightness of each divided area on the display device (step S1403). The area luminance correction unit 2070 performs luminance correction of the image using the set luminance correction value conditions (step S1404), and sends the luminance-corrected image data to the conversion error calculation unit 2040 as corrective data (label).

[0073] On the other hand, the image conversion unit 2030 acquires a low-quality image for learning (step S1405) to perform forward propagation using the low-quality image as input, and generates a converted image (step S1406). The conversion error calculation unit 2040 calculates the error between the converted image generated through forward propagation and the brightness corrected image that is the output of the region-by-region brightness correction unit 2070. The conversion parameter update unit 2050 receives the conversion error in pixel units as teacher data, and sets conversion parameters that minimize loss (steps S1407, S1408). Learning is performed by repeating the processes of steps S1401 to S1408 one or more times.

[0074] According to a computer system 202 as illustrated in FIG. 11, different learning can be performed on each divided area (e.g., wiring and circular structures), making it possible to generate an image according to the operator's preferences from a low-quality image.

[0075] In the above embodiment, an example has been described in which the luminance of each divided region is changed. However, it is also possible to color a specific pattern with a color other than gray levels by image processing, and use this as training data.

[0076] [Third embodiment] Next, an image generation system according to a third embodiment will be described with reference to Fig. 15. The system of this third embodiment may have a basic configuration similar to that of the previously described embodiment (Figs. 2 and 11). However, in addition to (or instead of) the configuration of the previously described embodiment, this third embodiment is configured to be able to execute different learning for each region by changing the scanning speed for each region.

[0077] FIG. 15 is a conceptual diagram for explaining the features of the third embodiment. FIG. 15(a) shows an example in which a low-speed scanning region and a high-speed scanning region are set within one field of view (FOV). Switching between high-speed scanning and low-speed scanning within the FOV can be realized, for example, by changing the scanning signal supplied to the scanning deflector of the electron microscope as shown in FIG. 15(b). The slope of the signal waveform indicates the scanning speed, and the example of FIG. 15(b) shows that the scanning speed of region (b) is 1 / 2 that of regions (a) and (c). The faster the scanning speed, the less the beam irradiation amount per unit area, so that the effects of charging and the like can be reduced, while the amount of detection signal decreases due to the smaller irradiation amount.

[0078] In the third embodiment, for example, a measurement standard or an area to be inspected can be set as an ROI, and the ROI can be selectively scanned at a low speed. This makes it possible to obtain necessary information while suppressing the effects of static electricity in the FOV.

[0079] As described above, in the first embodiment (FIG. 2), the ROI portion is selectively trained, and other regions are not trained, or the weight λ is lowered, thereby enabling efficient training.

[0080] In the third embodiment, the low quality image and the high quality image can be generated by changing the scanning speed as well as the frame number. Also, instead of changing the scanning speed, the low quality image and the high quality image can be generated by other means of changing the beam dose per unit area.

[0081] (others) Another example of a method for performing different learning on a region-by-region basis will be described. The secondary electrons generated by scanning an electron beam over a sample are emitted in greater quantities at the edge of a pattern, etc., compared to a flat sample surface, and are displayed with high brightness on the image. On the other hand, backscattered electrons emitted from the sample are suitable for generating a contrast image according to the material contained in the sample. Furthermore, an electron microscope equipped with an optical system as exemplified in FIG. 9 can selectively detect backscattered electrons heading in a specific direction, making it possible to generate an image with clear contrast for the edge of a structure on the sample.

[0082] The above-mentioned characteristics of detected electrons and other examples of generating learning models according to the purpose of measurement or inspection are described below. First, in the case of a learning model that generates an image for measuring the dimension between pattern edges, it is desirable to increase the weighting coefficient of the high brightness area (edge ​​part) of the SE image. Also, if it is necessary to clarify the material contrast, it is possible to increase the weighting coefficient of the part in the BSE image where the brightness changes significantly (part where the material changes). By using an optical system that can simultaneously detect secondary electrons and backscattered electrons as shown in the example of Figure 9, it is possible to generate SE images and BSE images and provide a system that can perform multiple learning according to the purpose, and it becomes possible to efficiently generate multiple models according to the application.

[0083] In the above embodiment, a system including a learning device that generates (estimates) a high-quality image by image conversion of a low-quality image has been described, but instead of a high-quality image, edge-to-edge dimension values, profile waveforms, contour lines, roughness evaluation values, etc. obtained from a high-quality image may be used as correct answer data to learn and estimate these values ​​or data. For example, in the case of edge-to-edge dimensions, the correct answer data is a dimension value obtained by a measurement algorithm from a high-quality image including two or more edges to be measured for dimensions, and learning is performed by updating the conversion parameters so as to minimize the error between the dimension value obtained by forward propagation of a low-quality image of the same region as the high-quality image and the dimension value of the correct answer data.

[0084] The same is true for other estimation targets; it is possible to train a learning machine by extracting correct data from high-quality images and performing learning using backpropagation.

[0085] The present invention is not limited to the above-described embodiments, and includes various modified examples. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration. [Explanation of symbols]

[0086] 100...scanning electron microscope system, 101...imaging unit (scanning electron microscope), 102...computer system, 103...signal processing unit, 104...input / output unit, 105...memory unit, 106...electron gun, 107...electron beam, 108...focusing lens, 109...focusing lens, 110...deflector, 111...objective lens, 112...sample, 113...stage, 114...emitted electrons, 115...deflector, 116...detection aperture, 117a, 117b...energy filter, 118...secondary electrons, 119...detector, 120...secondary electrons, 121...detector, 123..., 201...storage medium, 202...computer system, 203...input device, 2030...image conversion unit, 2040...conversion error calculation unit, 2050...conversion parameter update unit, 2060...image segmentation processing unit, 2070...region-specific brightness correction unit.

Claims

1. 1. A system configured to generate a transformed image based on an input of an input image generated by a charged particle beam device, comprising: one or more computer subsystems and one or more components executed by the one or more computer subsystems; The one or more components include a learning model in which parameters are adjusted to suppress an error between a first image input as training data during learning and a second image converted by inputting the input image to the system; The first image is a semiconductor pattern image, The first image is subjected to image correction using different parameters in at least two or more regions within the image; An image generation system, wherein the at least two or more regions are regions divided by image segmentation.

2. In claim 1, The image generation system, wherein the image correction includes at least one of luminance correction, brightness correction, contrast correction, and color correction.

3. 1. A non-transitory computer readable medium storing program instructions executable on a computer system to perform a computer implemented method for generating a transformed image based on input of an input image generated by a charged particle beam device, the method comprising: The computer-implemented method includes a learning model having parameters adjusted to suppress an error between a first image input as training data during learning and a second image transformed by inputting the input image to the system; The input image is a semiconductor pattern image, The first image is subjected to image correction using different parameters in at least two or more regions within the image; The at least two regions are regions divided by image segmentation.

4. In claim 3, The image correction includes at least one of a luminance correction, a brightness correction, a contrast correction, and a color correction.

5. 1. A method of generating an image in a system configured to generate a transformed image based on input of an input image generated by a charged particle beam device, comprising: The system includes one or more computer subsystems and one or more components executed by the one or more computer subsystems; The one or more components include a learning model in which parameters are adjusted to suppress an error between a first image input as training data during learning and a second image converted by inputting the input image to the system; The input image is a semiconductor pattern image, The first image is subjected to image correction using different parameters in at least two or more regions within the image; An image generating method, wherein the at least two regions are regions divided by image segmentation.

6. In claim 5, The image generating method, wherein the image correction includes at least one of luminance correction, brightness correction, contrast correction, and coloring correction.

Images