TW201510878A - Measurement device - Google Patents

Abstract

An objective of the present invention is to provide a measurement device whereby it is possible to easily reuse an identification screen which has been used in carrying out pattern matching in the past. A measurement device according to the present invention appends unique identification information to matching identification screens for determining matching success and/or algorithm identification screens for selecting algorithms, stores same in a storage unit, and calls the identification screens with the identification information as a key.

Description

Measuring device

Regarding a measuring device using pattern matching, the pattern matching uses an image of a sample.

In an apparatus for measuring/inspecting a pattern formed on a semiconductor wafer, the field of view of the inspection apparatus is aligned with a desired measurement position using a template matching technique. The template matching is a process of finding an area that is most consistent with a previously registered template image from the image of the search target.

As an example of a measurement/inspection apparatus using template matching, an apparatus for measuring a pattern on a semiconductor wafer using a scanning electron microscope is exemplified. In this device, although the field of view of the device is moved to the approximate position of the measurement position by the movement of the platform, only by the positioning accuracy of the platform, there is a large offset in the image taken at a high magnification of the electron microscope. The situation. Moreover, it is not limited to placing the wafer on the platform in the same direction every time, and there is a coordinate system of the wafer placed on the platform (for example, the arrangement direction of the wafers of the wafer, etc.) is not completely consistent with the driving direction of the platform. The situation. This also causes the shift in the image taken by the high magnification of the electron microscope. And, in order to obtain high magnification electrons at the desired viewing position The microscope image has a case where the electron beam is deviated only by a small amount (for example, several tens of μm or less) and is irradiated onto a target position on the observation sample (a case called electron beam shift). Even in this electron beam shift, the accuracy of the deflection control of the beam alone causes a shift in the irradiation position from the desired observation position.

Template matching is performed in order to correct such an offset and perform measurement/checking at the correct position. First, alignment is performed in multiple stages by using alignment of an optical camera of a lower magnification than that of an electron microscope image and alignment using an electron microscope image.

For example, when an optical camera is used to perform alignment of a coordinate system of a wafer placed on a platform, an image of a plurality of wafers (for example, wafers at the left and right ends of the wafer) located at a position separated on the wafer is used. alignment. First, a unique identical pattern (a pattern that is relatively located at the same position in each wafer) in each wafer or in the vicinity is registered as a template. As the pattern registered at this time, it is often used to form an alignment pattern for optics on a wafer. Next, in each of the wafers, the platform is moved so that the template is registered, and the image is acquired on each wafer. Perform template matching on the obtained image, and calculate the offset of the platform movement based on the obtained matching positions, and use the offset as the correction value of the platform movement to align the coordinates of the platform movement. System and wafer coordinate system.

When the alignment of the electron microscope is used, a unique pattern close to the measurement position is registered as a template in advance, and the relative coordinates of the measurement position observed from the template are memorized in advance. Image taken from an electron microscope When the measurement position is obtained, template matching is performed on the captured image, and the matching position is determined, and the relative coordinate portion that has been previously memorized from here is formed as the measurement position.

The field of view of the device is moved to the desired measurement position using template matching as described above.

In the case where the offset of the platform movement or the offset of the electron beam shift is large, the pattern for alignment cannot be printed in the image captured by the electron microscope. In this case, the situation in which the alignment pattern (peripheral search) is again searched for in the vicinity of the imaging position or the measurement is interrupted and the alignment failure (measurement interruption) is transmitted to the user is performed by an alarm. In order to carry out this processing, it is necessary to determine whether or not there is a pattern for alignment in the image.

In this determination, a matching score such as a template matching is used (for example, a correlation coefficient in a normalized correlation calculus). If the matching score is higher than the preset reference value (hereinafter, the reference value is referred to as "Score Acceptance"), it is determined that there is a pattern in the field of view, and if the matching score is lower than the acceptance score, it is determined that there is no pattern in the field of view. .

In template matching, when the appearance of the template and the image being searched are large, there is a possibility that the matching fails. The reason why the template is different from the appearance of each of the images to be searched is, for example, the following: (1) the imaging conditions of the inspection device when the template is registered and the inspection when the image to be searched is photographed. The difference between the imaging conditions of the device becomes large; (2) the difference between the performance of the semiconductor pattern captured when the template is registered and the result of the semiconductor pattern captured by the searched image becomes larger; and (3) registration The manufacturing process and the semiconductor pattern of the template are searched The manufacturing process of the semiconductor pattern in which the image is taken is different. Not limited to the above example, there is a case where the appearance of the image in the template and the image to be searched becomes large due to various factors.

As one of the methods for avoiding the failure of matching due to the deviation of the template from the appearance of the image being searched, it is proposed to use various kinds of image feature quantities extracted from the template or the searched image or from the template and the image to be searched. A method of judging the success or failure of the matching of various image feature quantities calculated by the mutual relationship (Patent Document 1). In this method, in the feature amount space formed by using a plurality of types of image feature quantities, a recognition surface for judging whether the matching is successful or not is obtained by mechanical learning (Support Vector Machine: SVM, etc.) (matching determination side) When the interface (super surface) is matched and the object to be inspected is matched, the identification surface is used to find the correct matching position. In the method, the matching is made robust by using a plurality of pieces of information (image feature quantities) obtained from the images.

Further, as another method for avoiding the failure of the matching due to the deviation between the template and the appearance of the image to be searched, a condition (a size of the pattern, an imaging magnification, etc.) for the degree of the deviation, the tendency or the cause of the deviation (the size of the pattern, the imaging magnification, etc.) is considered. Differences, to change the matching algorithm or to adjust the method of matching before processing (smoothing, edge enhancement, etc.). Further, there is an example in which the matching is succeeded by changing the imaging conditions (the imaging magnification, the acceleration voltage of the irradiated electrons in the electron microscope, the type of the detector, and the like) of the apparatus.

[Previous Technical Literature] [Patent Literature]

[Patent Document 1] Japanese Patent Application No. 2012-167363

The technique described in Patent Document 1 is a case where a new type of image (unused as an image at the time of learning) is input at the time of inspection, and the matching is unstable. In this case, it is necessary to use the new kind of image for learning again. However, when re-learning, there is a possibility that the identification surface used up to this point and the additional information associated therewith cannot be used.

For example, in a semiconductor production line, when a past manufacturing process is reproduced, even if a new type of image is input at the time of matching implementation, the new image is actually an image used in the past. In this case, the reliability of the matching performance of the previously learned recognition surface is higher than that of the learning surface that is learned again. In this case, it is preferable to consider returning the device setting to the previous recognition surface. However, since the recognition surface is a multi-dimensional plane in the multi-dimensional feature amount space, it is difficult to reproduce the recognition surface by manual operation. In the above Patent Document 1, a method for eliminating the problem is not described. Further, there is an example in which not only the recognition surface is restored to the previous setting, but also other device conditions (auxiliary information) when the same recognition surface is used in combination, and the previous setting is restored.

As a way to avoid matching failures due to the separation of the templates from the appearance of the image being searched, as described in the prior art, consider changing Match the algorithm, adjust the pre-processing (image processing) setting parameters, or adjust the imaging conditions. In this case, the user must select and determine settings suitable for the observation object from among the various options. In this operation, it is necessary to confirm each observation object, and in order to determine the setting of the matching success, it is necessary to make an attempt error, and the user has to stop the production line when attempting to change the setting because of a lot of complicated work. There is a flaw in reducing production efficiency. Moreover, there are also cases where the appropriate settings cannot be found.

Therefore, it is desirable to have a method of automatically selecting an appropriate setting in response to an object of observation. Since the method is implemented when matching, the user does not need to select (mode selection) settings, and is therefore referred to as non-pattern matching. In order to achieve modeless matching, consider the method of mechanical learning. That is, it is considered to use an identification surface obtained by mechanical learning to select an appropriate mode in response to the state of the observation object. However, even in the present method, when the recognition surface is learned again, it is difficult to restore the problem of setting the recognition surface up to this point.

The present invention has been made in view of the above-described problems, and it is an object of the present invention to provide a measuring device that can easily reuse the identification surface used in pattern matching in the past.

The measurement device of the present invention stores the unique identification information in at least one of the matching recognition surface for determining the success or failure of the matching and the algorithm recognition surface for selecting the algorithm, and stores the identification information in advance in the memory unit, and The identification information is set to be an index (KEY) and the identification faces are read.

According to the measuring device of the present invention, the identification surface used in the past image matching can be easily read and reused by the identification information unique to the identification surface.

100‧‧‧Learning Data Generation Department

101‧‧‧Image acquisition unit for learning

106‧‧‧Data Generation Department

110‧‧‧ Resumekeeping Department

111‧‧‧ Identification Face Learning Department

113‧‧‧ Identification Information Addition Department

117‧‧‧ Resumekeeping Department

118‧‧‧Memory Department

120‧‧‧Template Matching Department

121‧‧‧ Algorithm Selection Department

122‧‧‧ Matching success or failure determination department

140‧‧‧ Resume Information Selection Department

142‧‧‧ Resume Information Reading Department

1000‧‧‧Measurement device

FIG. 1 is a functional block diagram of the measuring device 1000.

Fig. 2 is a view showing a configuration of a scanning electron microscope which can be mainly used for measuring a pattern size of a semiconductor element formed on a semiconductor wafer.

FIG. 3 is a diagram for explaining an image set 102 and a learning material 107.

FIG. 4 is a flow chart for explaining processing of acquiring the image set 102 and the learning material 107 by the measuring device 1000.

FIG. 5 is a view showing an example of a character file 500 in which the content of the learning material 107 is described.

FIG. 6 is a view showing an example of the configuration of the history information 600.

FIG. 7 is a view showing an example of a character file 700 in which the content of the history information 600 is described.

[Fig. 8] A diagram showing a case where the matching face recognition matching algorithm is used to determine whether the matching is successful or not using the matching recognition face.

FIG. 9 is a functional block diagram illustrating an operation when the measuring device 1000 learns the recognition surface again.

FIG. 10 is a view for explaining an example of a hardware configuration of the measuring device 1000.

FIG. 11 is a view showing a GUI 1100 showing and editing the history information 600.

FIG. 12 is a diagram illustrating a GUI 1200 for searching the history information 600 stored in the storage unit 118.

FIG. 13 is a diagram illustrating a GUI 1300 for monitoring an operation state of an identification surface in the measurement device 1000.

Hereinafter, embodiments of the present invention will be described using the drawings. In the drawings, the same reference numerals are used, and the same members are denoted unless otherwise specified.

1 is a functional block diagram of a measuring device 1000 of the present invention. The measuring device 1000 defines the measurement position on the sample by template matching. Further, it is configured to mechanically learn the recognition surface used for the success or failure of the matching or the selection of the algorithm, and to give the unique identification information to the learned recognition surface and memorize it, and then read the information using the identification information. The identification surface used in the past.

The measurement device 1000 includes a learning material generation unit 100, a history storage unit 110, a template matching unit 120, and a history information selection unit 140. The learning data generating unit 100 generates learning materials for acquiring a plurality of sample images and learning the recognition surface. The history storage unit 110 uses the generated learning material to identify the matching recognition surface for determining whether the matching is successful or not, and the algorithm identification surface for selecting the matching algorithm. At least one (in the following description, both of them have mechanical learning), and the unique identification information is associated with and recognized by the recognition surface obtained by the learning. The template matching unit 120 performs template matching using the recognition surface. The history information selection unit 140 sets the identification information as an index and reads the recognition surface from the storage unit at the request of the user.

The learning material generation unit 100 includes a learning image acquisition unit 101 and a data generation unit 106. The learning image acquisition unit 101 acquires the image set 102 for learning. The image set 102 is an integration of the template image 103 and the matching position image. The matching position image includes at least one of the matching correct position image 104 and the matching incorrect position image 105, and each of the template images includes one or more images of one or more of the images. . The data generation unit 106 assigns a label necessary for the subsequent mechanical learning to the image data for the image set 102, and outputs it as the learning material 107.

Further, in the latter stage of mechanical learning, there are an example of learning a matching recognition surface for determining the success or failure of matching, and an example of learning an algorithm recognition surface for selecting a matching algorithm (matching mode selection). Learning the label when matching the recognition surface, specifically, the label indicating the image of the correct position in the image matching the correct position (herein referred to as the front side), indicating not in the image matching the incorrect position The label of the image in the correct position (referred to herein as negative). The label for learning the algorithm recognition surface is described in detail in FIG. 8 which will be described later, and refers to the matching algorithm name (matching pattern name) obtained after matching the correct position image. The image set 102 at this time does not match the correct position image, and may also be the searched image itself. Label and learning resources The material is not limited to those shown here, as long as the identification surface for determining the success or failure of the matching or the identification surface for selecting the algorithm can be obtained by mechanical learning.

The history storage unit 110 includes an identification surface learning unit 111, an identification information adding unit 113, a history storage unit 117, and a storage unit 118. The recognition surface learning unit 111 performs mechanical learning using the learning material 107, and further obtains a desired recognition surface. The identification surface can be obtained by a method already described in Patent Document 1 or the like. As a mechanical learning method, for example, an SVM or the like can be used. A specific example of the identification surface will be described later with reference to Fig. 8. The identification surface is an identification surface for identifying the success or failure of the matching in the feature amount space formed by the plurality of feature quantities extracted from the learning material (identification) Boundary, the same below) or select the matching face of the matching algorithm. The identification information adding unit 113 adds unique identification information to the identification surface data 112, which combines the obtained identification surface and additional information related to the identification surface (described in detail in FIG. 6). An example of the identification information will be described in detail with reference to FIG. 6 which will be described later, and includes a measurement processing program 115 describing the measurement procedure, auxiliary information 116 related to the measurement, and the like. The history storage unit 117 stores the identification surface data 112 to which the identification information is added, in the storage unit 118.

The template matching unit 120 includes at least one of the algorithm selection unit 121 and the matching success/failure determination unit 122. Hereinafter, for convenience of explanation, both are provided. The algorithm selection unit 121 performs template matching by using the measurement processing program 114 (identification information is given) and selecting a matching algorithm corresponding to the measurement target 130. About use A specific example of the matching algorithm to select the matching algorithm will be described later with reference to FIG. 8. The matching success/failure determination unit 122 obtains a matching result in which template matching is performed. For the matching process, the method described in, for example, Patent Document 2 can be used. Briefly, the measurement target, the template image, or the feature amount obtained from the correlations are drawn on the feature amount space obtained by the recognition surface learning unit 111, and based on the recognition surface held by the recognition surface data 112. The matching correct position of the measurement target 130 is obtained.

The algorithm selected by the algorithm selection unit 121 is different from the algorithm itself, and includes the difference in the setting parameters, the difference in the previous processing in the image processing, and the difference in the parameters of the pre-processing, even if the algorithm is the same.

When the measurement device 1000 does not use the algorithm selection unit 121, the matching algorithm specified in advance is used, and the recognition surface is used only in the determination of the success or failure of the matching. When the matching success determination unit 122 is not used, the matching plane selection matching algorithm is used, and in the template matching success determination, the general matching method is used without using the recognition surface (as described, for example, in Patent Document 2) Image reference matching of conventional methods, etc.).

When the matching result 136 obtained by the template matching using the identification surface is available, the matching result 136 is also associated with the above-mentioned identification information, and may be stored in the storage unit 118.

The history information selection unit 140 is provided with a history information reading unit 142. The history information reading unit 142 is based on the selection information 141 (search keyword (component name, creation date and time, producer, etc.), external bar) The component (the element pattern width is 20 nm or less, made after 2012, etc.), etc., and the identification surface that satisfies the condition is searched from the memory unit 118 and read, and the selection information 141 is designated to read the desired recognition surface and described later. The conditions of the resume information 600. The history information reading unit 142 transmits the read identification surface information 143 to the identification information adding unit 113. The identification information adding unit 113 adds the identification information 143 to the measurement processing program 115 in the same manner as described above. Thereby, the desired recognition surface stored in the storage unit 118 can be searched, and the measurement processing program 114 can be combined to perform pattern matching.

2 is a view showing a configuration of a scanning electron microscope (SEM) for measuring a pattern size of a semiconductor element formed on a semiconductor wafer as an example of the measuring device 1000.

The electron gun 201 is used to generate an electron beam. The deflector 204 and the objective lens 205 are controlled at any position on the semiconductor wafer 203 as a sample placed on the stage 202 so that the electron beam is irradiated to the focal point. The secondary electrons are released from the semiconductor wafer 203 irradiated with the electron beams, and are detected by the secondary electron detector 206. The detected secondary electrons are converted into digital signals by the A/D converter 207 and stored in the image memory 215 in the processing unit 214. The CPU 216 is for performing recognition processing using image processing and mechanical learning for the purpose. The template matching of the present invention is implemented by the processing unit 214. The display device 220 is for displaying the processing result. The optical camera 211 is used to perform light alignment lower than that of the aforementioned electron microscope. The signal obtained by taking the semiconductor wafer 203 by the optical camera 211 is also The A/D converter 212 is converted into a digital signal (when the output of the optical camera 211 is a digital signal, the A/D converter 212 is not required), and is stored in the image memory 215 in the processing unit 214, and the CPU 216 Image processing is performed for this purpose. In the case where the reflected electron detector 208 is provided, the reflected electrons emitted from the semiconductor wafer 203 are detected by the reflected electron detector 208. The detected reflected electrons are converted into digital signals by the A/D converter 209 or 210 and stored in the image memory 215 in the processing unit 214, and the CPU 216 performs image processing for this purpose.

FIG. 3 is a diagram for explaining the image set 102 and the learning material 107. The examples shown in Fig. 3 will be described below.

Fig. 3(a) shows an image obtained in order to learn the recognition surface. The image for learning the recognition surface refers to the template image 301 and the searched image 302 which are searched objects in the graphic matching. For the template image 301, an image obtained by an actual device (for example, an image obtained by an electron microscope or an image obtained by an optical camera 211) may be used, or may be as described in JP-A-2002- It is produced based on the design information of the semiconductor element as shown in the publication No. 328015. When the template image is created based on the design data, it is not necessary to specifically acquire the image by the measuring device 1000 in order to create the template. Within the searched image 302, the location 303 that conforms to the desired matching pattern will be formed to match the correct location. Matching positions other than the correct position 303 all match the incorrect position. In Fig. 3(a), the position 304 is used as an example of matching an incorrect position.

Fig. 3(b) shows the map included in the learning material 107. A picture like the example of the data. As illustrated in Fig. 1, the learning material 107 includes a combination of a template image 311 and a matching position image, and a label is attached to the matching image. Fig. 3(b) shows an example of an additional tag for learning the identification face of the success or failure of the match. A label of the front side is attached to the matching correct position image 312, and a negative label is attached to the matching incorrect position image 313. The recognition surface can be mechanically learned by using the material to which the label is attached.

In the learning material 107 for learning the identification face of the matching algorithm, the matching algorithm name is formed as a tag. The name of the matching algorithm in this case is not only the type of matching method, but also different algorithms can be given even if the setting parameters of the algorithm are different or the setting parameters of the previous processing in the image processing are different. name. Thereby, in the case where the setting parameters are different, it is also possible to process the different matching modes, and the difference of the setting parameters is also included in the range, and learning to select the matching face of the appropriate matching mode.

Fig. 3(c) is a diagram showing conditions for changing the learning material 107. When the learning material 107 is acquired, the more the variation of the various learning materials that are expected to be included in each of the examination objects is obtained, the more likely the matching becomes stronger. The conditions for causing the variation of the learning material 107 include a variation 321 of the pattern shape, a variation 322 of the device condition, and a variation 323 set by the selected matching algorithm.

As the pattern shape 321, there is a design variation 330 (designed shape). For example, consider the pattern of lines & spaces, the pattern of through holes, the pattern of connections, various other patterns, and the size and configuration of each pattern. Variations such as spacing. Since the pattern shape categories of the samples have a large influence on the learning result of the recognition surface, it is desirable to learn the recognition surface in each of the pattern shapes in advance and recognize the pattern shape category. The other pattern shapes 321 described below are also the same.

As the other pattern shape 321, there is a variation 331 due to a change in conditions of the semiconductor manufacturing process. For example, there are cases where the size of the pattern is different from the setting data or the roughness is generated on the line.

The other pattern shape 321 is a variation due to the structure or manufacturing process of the semiconductor element. For example, when the alignment of the upper and lower layers in the multilayer structure pattern is shifted, when the upper and lower layers are also observed in the observation image, the shape of the pattern observed differs depending on the degree of the offset. Moreover, due to the difference between the upper and lower layers, there is a case where the appearance of the edge of the image is different. Further, since the inclination angle of the side wall of the pattern is different, even if the side wall portion is captured in the observation image in the plan view, the design data and the appearance are different. Or depending on the material, there is a difference in the contrast of the images. Further, in the multiple exposure (SADP or the like) of the semiconductor process, there is a case where the appearance and the design data are different due to the alignment shift between the exposure layers.

As a variation of the device condition 322, there is a variation 335 set by the device. For example, there are differences in settings such as imaging magnification, acceleration voltage of irradiated electrons, probe current, focus setting value, frame addition amount of an image, and difference in detector. As a variation of other device conditions 322, there is a machine difference 336 between device settings. The variation of the other device condition 322 is a time-history change 337 of the execution value set by the device. Patterns of other interference conditions 338 are also considered to be added. For example, there is an electromagnetic field from the outside. The situation of noise or mechanical vibration.

The variation of the matching algorithm setting 323 is the type of the algorithm to be selected, the difference in the setting parameters of the algorithm, the difference in the processing before the image processing, or the difference in the setting parameters of the pre-processing.

The image set 102 can be obtained under various conditions by changing the manner in which the image set 102 is obtained by changing the condition of the learning material 107 listed above. It is considered that a more robust match can be formed by using the method of generating the learning material 107 using the imaginary variation to perform mechanical learning. However, it is not necessary to have many conditions to make the match strong. Further, the change condition is not limited to the conditions listed here, and the image set 102 may be acquired using other conditions as long as the condition of the image can be changed.

FIG. 4 is a flow chart for explaining the process of acquiring the image set 102 and the learning material 107 by the measuring device 1000. The learning image acquisition unit 101 performs steps S401 to S410, and the data generation unit 106 performs steps 421 to S422. The steps of FIG. 4 will be described below.

(Figure 4: Step S401)

The learning image acquisition unit 101 acquires the template image 311.

(Fig. 4: step S402)

The learning image acquisition unit 101 moves the visual field so that the observation target on the wafer enters the field of view of the apparatus (as explained in the prior art, since the movement of the visual field shifts, the observation target is not necessarily limited to the observation target. After entering the field of view, the measurement device 1000 acquires an image of the field of view and further becomes a searched image. In this step, the photographing condition information 430 describing the conditions when the image is taken is the input of step S422.

(Fig. 4: Step S403)

The learning image acquisition unit 101 performs template matching between the template image 311 and the image to be searched. The algorithm setting information 431 describing the setting of the matching algorithm used at this time is input to step S422.

(Fig. 4: step S404)

After the measurement performed after matching or matching is successful, the process proceeds to step S405, and if it fails, the process proceeds to step S408.

(Figure 4: Steps S405~S407)

The learning image acquisition unit 101 cuts out the matching position (S405), and registers the cut image as the matching correct position image 312 (S406). Further, an arbitrary position other than the image corresponding to the correct position may be cut out (S407), and the image cut out in step S409, which will be described later, is registered as the matching incorrect position image 313.

(Figure 4: Steps S408~S409)

The learning image acquisition unit 101 cuts out the matching position (S408), and The cut image is registered as the matching incorrect position image 313 (S409). The image cut out in step S407 can also be registered as the matching incorrect position image 313.

(Fig. 4: step S410)

The learning image acquisition unit 101 determines whether or not the matching processing is completed for all the learning material conditions. If it is not finished, the process returns to step S401 and a matching is performed for the remaining conditions. For all the conditions (for example, all the conditions described in FIG. 3(c)), the image set obtained by ending the matching process is formed as the image set 102. Here, in fact, although the template matching is performed to determine whether the matching is successful or not, the user can also determine whether the matching is successful or not by visual confirmation.

(Figure 4: Steps S421~S422)

The data generation unit 106 assigns a label to the image set 102 (S421). The data generating unit 106 associates the template image with the matching image and the corresponding tag for each template image, and stores the template image in the storage unit 118. A corresponding example will be described with reference to Fig. 7 which will be described later. Further, the imaging condition information 430 or the algorithm setting information 431 may be combined and stored in the storage unit 118. The contents of the imaging condition information 430 and the algorithm setting information 431 are, for example, part or all of the contents described in FIG. 3(c).

FIG. 5 is a view showing an example of a character file 500 in which the content of the learning material 107 is described. For template image 510, matching result graph Like 511 and 512, the image file name is described. For the matching position image, the image file name is described in the result of each additional label. For example, as shown in FIG. 5, it is described as a front image file name list 511 and a negative image file name list 512. Further, the imaging condition information 430 and the algorithm setting information 431 may be described. These descriptions refer to the information shown in Figure 3(c). The text file 500 is created in step S422 and stored in the storage unit 118.

FIG. 6 is a view showing an example of the configuration of information stored in the storage unit 118 by the history storage unit 117 (hereinafter referred to as history information 600). The history information 600 is a combination of the recognition surface used in the past pattern matching and the information corresponding thereto, and has a management number/character string 610, an algorithm recognition surface attribute 611, a matching recognition surface attribute 612, and a measurement processing program name 613. The auxiliary information 614, the matching performance information 615, and the learning information information 616.

The management number/character string 610 is inherent identification information that is used to define the identification surface stored by the memory unit 118 or to define the identification surface used by the measurement device 1000 to perform template matching. As long as the recognition surface can be uniquely identified, it is not limited to the description.

The algorithm identification surface attribute 611 is used to select attribute information of the recognition surface of the matching algorithm, and includes a learning method category 620, a learning method additional information 621, a number of matching algorithms, and a name 622 of each matching algorithm. The feature quantity category 623 used in the learning. The algorithm identification surface attribute 611 is not limited to these, and other information that the display algorithm recognizes the attributes of the surface may also be used.

The learning method category 620 specifies the type of learning method (for example, the non-linear soft boundary SVM) used when the recognition surface learning unit 111 performs mechanical learning. The algorithm selection unit 121 also uses a recognition method corresponding to the learning method category 620. In the identification surface construction in the SVM, the required support vector information and the category information attached to each of the foregoing support vectors (the matching success or the type of matching algorithm) may also be included.

The learning method additional information 621 is information describing the setting parameters in the mechanical learning method implemented by the recognition surface learning unit 111. For example, if it is a non-linear soft boundary SVM, it includes the type of core function (Gaussian function, polynomial kernel function, etc.), the coefficient of the core function (if Gaussian function, γ value, etc.), and the setting is imposed on the non-satisfied face. The cost coefficient of the degree of loss of the identified sample, and the like. In addition, information about the support vectors obtained by mechanical learning can also be included.

The feature amount category 623 specifies the type of the feature amount used in the machine learning performed by the recognition surface learning unit 111. Multiple categories can also be specified. The recognition surface learning unit 111 obtains the recognition surface in the feature amount space composed of the feature amount. Further, when the matching algorithm is selected by the algorithm selecting unit 121, the feature amount of the measurement target is also drawn in the feature amount space, and the recognition surface is recognized. An example of the type of the feature amount is described in Patent Document 2.

The matching identification surface attribute 612 is used to determine attribute information of the recognition surface of the matching success or not, and includes a learning method category 624, a learning method additional information 625, and a feature quantity category 626. These contents are used in addition to the viewpoints used to determine whether the match is successful or not. The algorithm identifies the polygon attribute 611 with the same information.

The measurement processing program name 613 is a file (processing program file) name in which the measurement procedure performed by the measurement device 1000 is described. This file describes various settings, processing execution steps, and the like of the apparatus for performing the desired measurement.

The auxiliary information 614 includes a general attribute 627 and a measurement object attribute 628. The general attribute 627 refers to the date and time when the history information 600 is created, the name of the producer, and the annotation information of the content arbitrarily recorded by the user. The measurement object attribute 628 refers to information such as the original information of the component. For example, the engineering name of the measurement object, the component structure name (line & space, via array, SAxP, FinFET, DSA, etc.), the assumed size of the component (line width, aperture, etc.). In addition to the above, other information may be used as long as it is available to define the identification information 600.

The matching performance information 615 is information about the matching performance when the template matching is performed by using the algorithm identification surface attribute 611 and the respective recognition surfaces corresponding to the matching recognition surface attribute 612. For example, the match between the correct rate, the matching score matching the correct position, and the matching score of the matching failure location (eg, matching scores for matching incorrect locations/matching scores matching the correct position), success of matching (delimited template) Success in matching with each combination of images being searched). The matching performance information 615 is not limited to the above content, and may be information that can indicate matching performance.

The learning information 616 is the learning material 107 and its attribute information. For example, the learning image 629 describes the file name of the template image, the file name matching the image of the correct position, and the matching incorrect position. The file name of the image. The imaging condition information 630 is information such as a measurement device ID, an imaging magnification, an imaging range, an acceleration voltage of an irradiation electron, a probe current, a frame addition amount, and a focus characteristic value. The measurement target information 631 describes the line width, the edge thickness, and the like. The update history information 616 is a history information that describes the creation of the learning materials. For example, when a new learning material is created based on a certain learning material (herein referred to as original learning material), the management number/character string 610 of the original learning material is recorded and the learning is performed again. Execution conditions (name of additional learning materials, etc.). When a plurality of different original learning materials are used to create learning materials, the plurality of original learning materials are recorded. When performing multiple re-learning, all of the update histories can be recorded. In addition, an update history of only a part of the re-learning may be described as needed. In this way, the update history information 616 is used to trace the production history of the learning materials. The learning material information 616 is not limited to the above content, and may be information that can represent the original information of the learning materials.

The history storage unit 117 sets the management number/character string 610 as an index, and associates the above-described respective information included in the history information 600 in the storage unit 118. For other information, the management number/character string 610 can be similarly indexed and corresponding, and stored. Thereby, the past history of the identification surface and the associated information can be stored.

FIG. 7 is a view showing an example of a character file 700 in which the content of the history information 600 is described. Further, the history storage unit 117 creates a text based on information input on the screen described in, for example, FIG. 11 which will be described later. The file 700 is stored in the memory unit 118. In the management number/character string 610, the history storage unit 117 automatically numbers the numbers different from the management number/character string 610 of the stored character file 700, or the user performs the screen on the screen described in FIG. 11 which will be described later. Input. By adding the management number/character string 610 to the text file 700, each piece of information can be associated and stored in the memory unit 118.

The description form of each piece of information is not limited to those shown in FIG. Further, each piece of information may be divided into a plurality of files and described. In this case, the correspondence is maintained by assigning the management number/character string 610 to each file and correspondingly, and correspondingly matching the files based on the file name.

FIG. 8 is a diagram showing a case where the algorithm is used to identify the face selection matching algorithm, and the matching recognition surface is used to determine whether the matching is successful or not. For the sake of simple description, the identification surface in the two-dimensional feature quantity space using two feature quantities is exemplified, but three or more types of feature quantities may be used.

Fig. 8(a) is a diagram showing a case where the matching recognition surface 800 is used to determine whether or not the matching is successful. The method described in Patent Document 2 is a method for determining whether or not the matching is successful using mechanical learning. For example, when the SVM is used, in the feature amount space formed by the plurality of feature quantities obtained from the learning material 107 at the time of learning, the matching recognition face 800 that divides the feature amount into two types of matching success and matching failure is obtained. . When the matching is performed, the feature quantity is obtained from the measurement target, and the same feature quantity is drawn in the same feature quantity space as that at the time of learning, and the matching is determined based on which type of the feature quantity belongs to the matched identification surface 800. Whether or not. If counted If the measured feature quantity belongs to the matching success category, it is determined that the matching is successful, and if it belongs to the matching failure category, it is determined that the matching fails.

Fig. 8(b) is a diagram showing a case where the algorithm is used to identify the face 850 and the matching algorithm is selected. Here, although an example in which any of the three matching algorithms is selected is shown, the number of algorithms as selection candidates may be arbitrary. Similarly to FIG. 8(a), in the feature amount space formed by the plurality of feature quantities obtained from the learning material 107 at the time of learning, the feature amount is divided into three types corresponding to the matching algorithms. The algorithm identifies face 850. When the matching is performed, the feature amount is obtained from the measurement target, and the same feature amount is drawn in the same feature amount space as that at the time of learning, and the type of the feature amount belongs to which is divided by the algorithm recognition surface 850. The matching algorithm used.

FIG. 9 is a functional block diagram showing an operation when the measuring device 1000 learns the recognition surface again. When the new learning material 107 different from the learning material 107 used for learning is added and the recognition surface is learned again, the newly added learning material 107, the learned learning material 107, and the stored learning unit 107 are stored. Learn to use the information 616, etc. to learn again. Since the history information 600 obtained by the previous learning is managed, it is possible to easily select the existing learning materials for re-learning and to perform the learning again efficiently.

FIG. 10 is a view for explaining a hardware configuration example of the measuring device 1000. In the measurement device 1000, it is not always necessary to mount all the components in one hard body, and the components may be dispersed and mounted on a plurality of devices. In the example shown in FIG. 10, a connection is made in the network 1100. Measuring devices 1001 to 1003, memory unit 118, and computer 1004.

The computer 1004 is configured to implement at least a part of the functions of the measurement device 1000. For example, the function of the history information selection unit 140 and the output unit 135 can be installed on the computer 1004, and an operation input from the user can be received via the GUI described later with reference to FIG. 11, or the user can be presented with the matching result 136. The memory unit 118 can share the measurement devices 1001 to 1003. The history information 600 may be a different management number/character string 610 attached to each measurement device, and may be shared as long as the management number/character string 610 is not repeated. The measurement devices 1001 to 1003 include functions other than the functions of the computer 1004 and the memory unit 118 among the functions of the measurement device 1000.

The number of measuring devices is arbitrary, and is not limited to three. The number of the computers 1004 is not limited to one, and the functions of the measuring device 1000 may be dispersed and installed on a plurality of computers.

FIG. 11 is a view showing a GUI 1100 showing and editing the history information 600. The GUI 1100 can be configured, for example, as an operation screen on a display. The GUI 1100 is provided with an item corresponding to at least a part of the history information 600. In the example shown in FIG. 11, the management number/character string 1110, the algorithm recognition plane attribute 1111, the matching recognition plane attribute 1112, the measurement processing program file name 1113, the auxiliary information 1114, the matching performance 1115, and the learning material can be displayed. Information 1116 and edited.

The user inputs the value to the management number/character column 1110, and presses the recognition face to read the button 1133. The history information reading unit 142 reads the identification information 600 from the storage unit 118 and displays it on On the GUI1100. When the user specifies the measurement processing program file name of the measurement device 1000 in the measurement processing program file name 1113 and presses the connection button 1132, the displayed history information 600 and the designated information can be connected by the same management number/character string 1100. Measurement processing program. When the history information storage button 1131 is pressed, the history storage unit 117 stores the displayed history information in the storage unit 118. The matching performance 1115 can also be read from the memory unit 118 and the past matching result can be obtained. In this case, the button 1120 is read by the matching result, and the data describing the matching result is read. The learning material information 1116 is a list of specified images. Also, an image for learning can be selected from the list. Further, an image (button 1121) may be deleted from the learning material 107, and a new image (button 1122) may be added. When the learning button 1130 or the learning button 1134 is pressed again, the recognition surface learning unit 111 learns or learns the recognition surface again based on the displayed content.

FIG. 12 is a diagram showing a GUI 1200 for searching the history information 600 stored in the storage unit 118. The user inputs the search key 1201 and presses the search button 1202. The history information reading unit 142 displays the matching list on the list display unit 1210 in accordance with the history information 600 and the search key 1201. The list display unit 1210 displays at least a part of the history information 600. When any of the history information 1211 displayed on the list display unit 1210 is selected and the detailed display button 1212 is pressed, the history information reading unit 142 reads the history information 1211 from the storage unit 118 in detail and displays it on the GUI 1100, for example.

When the selection button 1213 is pressed, the selected history information 1211 is selected. Attached to the measurement handler and used this material for matching after implementation.

FIG. 13 is a diagram illustrating a GUI 1300 for monitoring an operational state of an identification surface in the measurement device 1000. Since the recognition surface used for pattern matching by the measurement device 1000 is stably performed, there is a case where the conditions of the measurement target or the like are switched. By using the GUI 1300, it is possible to visually confirm whether or not to switch, or to visually confirm the change history of the recognition surface up to this point. The GUI 1300 displays the operational status of the recognition surface in each of the measuring devices 1000. Specifically, the following information is displayed on each measurement device 1000 by the chart 1310.

The horizontal axis of the graph 1310 indicates the date and time and the management number/character string 610 of the identification surface used by the measuring device 1000 in each date and time. The vertical axis of the graph 1310 indicates the matching error rate and the matching score when the graph matching is performed by the measuring device 1000.

The user understands that when the matching score decreases with time and time, there is a problem that the matching is unstable. Similarly, it can be easily understood that when the error rate increases, the matching becomes unstable. Further, when the matching score is equal to or lower than the predetermined threshold, the point of the matching instability may be displayed on the GUI 1300, and the change of the recognition surface or the learning again may be promoted.

When the detailed display button 1311 is pressed, the history information 600 of the identification surface displayed on the graph 1310, the matching performance up to the measurement device 1000, and the matching performance of the identification surface in the other measurement device 1000 can be displayed. Wait.

<Summary of the Invention>

As described above, the measurement device 1000 of the present invention can provide the identification information exemplified as the history information 600 to at least one of the identification surface for determining the success or failure of the matching and the identification surface of the selection matching algorithm. The identification information is set as an index to read out the identification faces used in the past. Thereby, even if the recognition surface is updated by, for example, re-learning, the previously generated recognition surface and the associated history information 600 can be read, and template matching can be performed using this.

Moreover, since the setting of the identification surface used in the past can be easily returned to the measuring device 1000 according to the present invention, the user can learn again and update the identification surface without worrying about the deterioration of the matching performance.

Further, according to the measuring apparatus 1000 of the present invention, matching can be performed with a small number of matching algorithms as compared with a method of selecting a matching algorithm by mechanical learning and using a combination of a plurality of matching algorithms (integrated learning). Thereby, the matching processing time can be shortened and the productivity of the measuring device 1000 can be improved.

The history information 600 is desirably configured so that the user can easily read each of the identification faces later. Therefore, the history information 600 is desirably configured to display the learning results of the respective recognition surfaces well, and to use information that allows the user to easily understand the characteristics. The learning result of the recognition surface is greatly affected by the shape pattern type of the measurement object, and the shape pattern category is also easy to understand for the user. Therefore, using this as the history information 600 is useful. Moreover, since the pattern shape is significantly different in appearance or learning result due to the photographing magnification, It is also useful to use pattern shapes and photographic magnification at the same time. Further, since it is desired to use the recognition surface of the pattern shape well, it is also useful to use the matching performance when the pattern matching is performed using the pattern shape as the history information 600.

The present invention is not limited to the above embodiments, and various modifications can be included. The above embodiments are described in detail in order to clearly understand the present invention, and are not limited to having all of the constituents described. For example, the present invention is also applicable to an inspection apparatus for inspecting a sample using pattern matching. In the above description, an example in which a scanning electron microscope is used as the measuring device is shown. However, the present invention is not limited thereto, and can be widely applied to a measuring device that performs template matching using a sample image.

100‧‧‧Learning Data Generation Department

101‧‧‧Image acquisition unit for learning

102‧‧‧Photo Collection

103‧‧‧Template image

104‧‧‧ Matching the correct position image

105‧‧‧Matched incorrect position image

106‧‧‧Data Generation Department

107‧‧‧Learning materials

110‧‧‧ Resumekeeping Department

111‧‧‧ Identification Face Learning Department

112‧‧‧identification information

113‧‧‧ Identification Information Addition Department

114‧‧‧Measurement Processing Program

115‧‧‧Measurement Processing Program

116‧‧‧Auxiliary information

117‧‧‧ Resumekeeping Department

118‧‧‧Memory Department

120‧‧‧Template Matching Department

121‧‧‧ Algorithm Selection Department

122‧‧‧ Matching success or failure determination department

130‧‧‧Measured objects

135‧‧‧Output Department

136‧‧‧ Match results

140‧‧‧ Resume Information Selection Department

141‧‧‧Select information

142‧‧‧ Resume Information Reading Department

143‧‧‧identification information

Claims (12)

A measuring device that performs pattern matching on an image of a sample, and is characterized in that the learning data generating unit includes a combination of a template image used for performing the pattern matching and a result image obtained by matching the pattern. a plurality of learning materials are produced; the mechanical learning unit uses the learning data to obtain at least one of a matching recognition surface and an algorithm recognition surface by mechanical learning, and the matching recognition surface is used to determine that the graphic matching is successful. Whether the recognition surface of the feature quantity space is used for selecting an identification surface on a feature quantity space of an algorithm used for performing the pattern matching; the history storage unit is configured to describe the matching identification surface At least one of the matching recognition surface data or the algorithm recognition surface data describing the recognition surface of the algorithm corresponds to the unique identification information and is stored in the storage unit; the selection unit sets the identification information as an index and reads Corresponding to the aforementioned matching identification surface data or the corresponding algorithm identification surface data; and the matching part, according to the former Result selection unit read out using the face identification information or the matching algorithm to identify the surface of the pattern matching data used in the embodiment of the past, determining the success or pattern matching algorithm to select the foregoing. The measuring device according to claim 1, wherein the history storage unit sets a type of the shape pattern of the sample and The identification information is associated with and stored in the storage unit, and the selection unit sets the type of the shape pattern as an index and receives a request to read the matching identification surface data or the algorithm identification surface data. And reading the matching identification surface data or the algorithm identification surface data corresponding to the type of the shape pattern from the memory unit. The measurement device according to claim 2, wherein the history storage unit associates the imaging magnification of the shape pattern with the identification information, and stores the image in the memory unit, wherein the selection unit The type of the pattern and the imaging magnification of the shape pattern are indexed, and when a request for reading the matching identification surface data or the algorithm identification surface data is requested, the type corresponding to the shape pattern and the foregoing are read from the memory unit. The aforementioned matching identification surface data of the photographic magnification of the shape pattern or the aforementioned algorithm identification surface data. In the measurement device of the third aspect of the invention, the history storage unit uses the type of the shape pattern and the imaging magnification of the shape pattern to match the matching accuracy rate of the pattern matching with the identification information. Correspondingly, the selection unit is configured to read the matching identification surface data or the calculation when the type of the shape pattern, the imaging magnification of the shape pattern, and the matching accuracy ratio are indexed. When the method of identifying the surface data is requested, the matching identification surface data or the algorithm identification surface data corresponding to the type of the shape pattern, the imaging magnification of the shape pattern, and the matching accuracy rate are read from the memory unit. In the measurement device of the fourth aspect of the invention, the history storage unit stores the matching recognition surface attribute of the attribute of the matching recognition surface, and the algorithm recognition surface attribute in which the attribute of the algorithm recognition surface is described. The processing program information of the measurement step of the measuring device, the auxiliary information associated with the pattern matching, the matching performance information describing the performance index of the pattern matching, and at least one of the learning materials and the identification information are described. Corresponding to, and stored in the aforementioned memory. The measurement device of claim 5, wherein the history storage unit uses at least one of the following as the matching identification surface attribute, and includes: a learning method name used when mechanically learning the matching recognition surface And used for mechanically learning the core function category when matching the recognition surface, the coefficient in the core function, and the feature quantity category used when mechanically learning the matching recognition surface. The measurement device of claim 5, wherein the history storage unit uses at least one of the following as the algorithm recognition surface attribute, and includes: learning for mechanically learning the recognition surface of the algorithm The name of the technique, the type of the core function used when the machine learns the recognition surface of the algorithm, the coefficient in the core function, the number of the algorithms that are candidates for selection, the name of the algorithm that is the candidate for selection, and the use of mechanical learning. The feature quantity category when the aforementioned algorithm recognizes a face. For example, the measuring device of claim 5, wherein The history storage unit uses at least one of the following as the matching performance information, and includes the separation of the matching score in the pattern matching and the success of the matching in the mechanical learning. The measurement device according to the first aspect of the invention, wherein the learning data generating unit includes a learning image acquiring unit that acquires a plurality of images for capturing the sample, and uses a plurality of types of matching algorithms. Performing pattern matching on the plurality of images, and cutting out the portion of the image in which the pattern matching has been successfully matched as the correct position image, and cutting out the other portions as the incorrect position image Further, the aforementioned correct position image and the aforementioned incorrect position image are used as the resultant image obtained by the aforementioned pattern matching. The measuring device according to claim 5, wherein the measuring device provides a GUI for displaying at least one of the following and changing the screen, and includes the identification information and the shape pattern of the sample. The type, the photographic magnification of the shape pattern, the matching correctness rate, the matching recognition surface attribute, the algorithm recognition surface attribute, the processing program information, the auxiliary information, the matching performance indicator, and the learning data. The measuring device of claim 5, wherein the measuring device provides a GUI for specifying at least one of the following and searching for the identification information that is stored in the memory unit, and includes: The identification information, the type of the shape pattern of the sample, the imaging magnification of the shape pattern, the matching accuracy rate, the matching recognition surface attribute, the algorithm recognition surface attribute, the processing program information, the auxiliary information, and the matching performance indicator And the aforementioned learning materials. The measuring device according to claim 1, wherein the measuring device uses the mechanical learning matching identification information or the mechanical learning algorithm to recognize face information, and the matching score and the graphic are performed when the graphic matching is performed. At least one of the matched error rates outputs an information describing its time series change.