JP2010159979A - Visual examination method and visual examination system - Google Patents

Abstract

An object of the present invention is to provide an appearance inspection method and a system capable of determining a lethality according to an appropriate standard in an appearance inspection during manufacturing and a subsequent fatality inspection.
A function which is a quality of a function determined by a final function inspection using a function inspection device for each appearance class which is a defect type learned by learning an appearance class classifier (404, 406). Learning the function class classifiers (407, 408) by teaching the correspondence of the class, and extracting the fatal factor that is a quantitative description regarding the appearance feature during the examination that determines the quality of the final function, It has a learning process of determining a lethality judgment condition based on the lethality factor that is the extracted quantitative description.
[Selection] Figure 4

Description

  The present invention relates to a defect classification method and apparatus for detecting a defect based on an image of an object and classifying the defect, for a thin display such as a PDP or TFT, a thin film device such as a semiconductor wafer or a hard disk substrate, and the like. The present invention relates to a device manufacturing method using the.

  Thin film devices such as plasma displays, liquid crystal displays, semiconductor wafers, and hard disk magnetic heads are manufactured through a number of processing steps. In the manufacture of such a thin film device, appearance inspection is performed for each of a series of processes for the purpose of improving and stabilizing the yield. In appearance inspection, the surface of the target device is irradiated with lamp light, laser light, electron beam, etc., and the pattern defect or foreign material is detected based on the image obtained by detecting reflected light, scattered light, secondary electrons, etc. Detect defects such as. The defect feature quantity such as brightness and size is calculated from the image of the defect portion obtained at this time or the review image obtained by imaging the detected defect position at a higher magnification, and the defect is classified based on them. .

  In defect classification of thin film displays such as plasma displays and liquid crystal displays, classification of defect types and determination of lethality are performed. The classification of the defect type is to classify scratches, foreign matter, pattern formation defects and the like according to the appearance, and is performed for the purpose of process monitoring. The lethality determination is performed for the purpose of determining whether or not to flow the product to the next process.

  Japanese Unexamined Patent Publication No. 2000-306964 (Patent Document 1) discloses an inspection data processing method for performing defect type classification and criticality determination. Note that the fatality determination described in Patent Document 1 is performed based on a determination rule defined for each defect existence area. Further, the case where the discrimination rule changes for each defect type is also taken into consideration. Further, it is described that the rule is corrected based on the probe inspection result.

  (Patent Document 2) describes a defect analysis in which a fatal defect and a non-fatal defect are classified as teaching data for teaching type defect classification by matching a defect image obtained in the middle of manufacturing and a final electrical test result. A method is disclosed. According to this method, it is possible to accurately classify the fatal defect.

JP 2000-306964 A JP 2001-230289 A G.H.John and P. Langley: Estimating Continuous Distributions in Bayesian Classifiers, Proceedings of the 11th Conference on Uncertainty in Artificial Intelligence, pp338-345 (1995)

  The lethality determination in the inspection during the manufacturing is performed for the purpose of preventing a defective product from flowing to a subsequent process. If the criteria for determining the criticality are too strict, there is a risk of disposing of a product that does not become defective. If it is too loose, there is a risk that the processed product will be wasted, so an appropriate standard is required. It is.

  In the method disclosed in Patent Document 1, the criticality determination is performed based on a determination rule determined for each defect existing region and defect type. However, since the discrimination rule needs to be set manually in advance, there is a problem that it is difficult to set an appropriate standard. Although there is a description that the rule is corrected based on the probe inspection result, the rule framework needs to be determined in advance. In addition, this method cannot be applied if there is no knowledge about lethality.

  In the method disclosed in Patent Document 2, the criticality determination is performed using a classifier learned by teaching defect data classified based on the final match of the electrical test results. According to this method, it is possible to accurately classify a fatal defect, but the classifier is a black box, and there is a problem that it is difficult to obtain knowledge about appearance features that influence the fatality. Was.

  On the other hand, in order to provide feedback to process state management and design, it is necessary to extract knowledge about appearance features that influence the lethality.

  In view of the above problems, an object of the present invention is to provide a visual inspection method and system capable of performing criticality determination based on appropriate standards in visual inspection during manufacturing and subsequent criticality inspection. It is in.

  Another object of the present invention is to provide an appearance inspection method and system capable of extracting knowledge about the appearance characteristics of defects that influence the fatality based on a match with the final function inspection. It is to provide.

  In order to achieve the above object, according to the present invention, during learning, a plurality of defects generated in a learning object are imaged using an appearance inspection apparatus to detect an image signal, and based on the detected image signal By teaching, as a classification condition, a first learning step for calculating the feature amount of each defect, and an appearance class that is a defect type classified based on the feature amount of each defect calculated in the first learning step A function inspection device is used for each of the second learning step for learning the appearance class classifier and the appearance class that is the defect type classified based on the feature amount of the defect type learned in the second learning step. Learning the function class classifier by teaching the correspondence of the function class that is the quality of the function determined by the final function test, and the appearance characteristics during the test that determine the quality of the final function. A third learning step for extracting a fatal factor which is a quantitative description, and a fatality determination condition is determined based on the fatal factor which is the quantitative description extracted in the third learning step. A learning process having a fourth learning step; and a plurality of defects generated in an inspection object being manufactured at the time of inspection are imaged using an appearance inspection apparatus to detect an image signal, and the detected image signal A first inspection step that calculates the feature amount of each defect based on the classification conditions learned in the second learning step based on the feature amount of each defect calculated in the first inspection step. A second inspection step for classifying the appearance class as the defect type using an appearance class classifier, and the quantification in the fourth learning step for each appearance class classified in the second inspection step. In the description A visual inspection method characterized by having an inspection process and a lethality third inspection prediction determines fatality based on the criticality judgment condition determined based on factors step that.

  Further, in the inspection process, the present invention further discards or modifies the product based on the result of the fatality prediction determination for each appearance class obtained in the third inspection step, and proceeds to the next process. Or a fourth inspection step for determining whether to flow to the next process as it is.

  In the learning process, the present invention further includes a fourth learning for selecting a defect type and a feature to be monitored based on a fatality factor that is the quantitative description extracted in the third learning step. And selecting each appearance class classified in the second inspection step based on the fatal factor that is the quantitative description in the fourth learning step. The frequency distribution of the detected defect types and features is calculated and recorded at regular intervals, and the occurrence of a fatal failure is detected by detecting the shift to the more fatal side by monitoring the frequency distribution. And a third inspection step.

  Further, according to the present invention, in the learning process, a defect type and a feature to be monitored are selected based on a fatality factor that is the quantitative description extracted in the third learning step, and each feature is selected. A fourth learning step for setting a threshold value for each of the appearance classes classified in the second inspection step in the inspection process, and the quantitative learning in the fourth learning step. The number of defects on the side of high fatality that is greater than or less than the threshold value of the defect type and feature of the monitoring object selected based on the fatality factor that is a description is recorded at regular intervals, and It has the 3rd inspection step which presumes a cause based on change of the number of defects, and counteracts.

  In the learning process, the present invention further includes, for each appearance class that is the defect type classified based on the feature type of the defect type learned in the second learning step and for each type of the learning product. In addition, the function class classifier is learned by teaching the correspondence of the function class that is the quality of the function determined by the final function test using the function test apparatus, and the test function that determines the quality of the final function A third learning step of extracting a fatal factor that is a quantitative description of the appearance feature of the image, and in the inspection process, for each appearance class classified in the second inspection step, The fatality factor obtained based on the comparison between the product types of the fatality factor that is the quantitative description that is learned in the third learning step and output from the function classifier And having a third checking step of feeding back the design information of tolerance of each cultivar against.

  According to the present invention, the feature quantity and the quality of the final function test are input in association with each other, and the quantitative description regarding the external feature that determines the fatality is output based on the information. Gain reliable knowledge of appearance features. Further, since the criticality determination condition is set based on the criticality factor, it is possible to set an appropriate criticality determination criterion during the manufacturing.

  In addition, according to the present invention, a defect is detected by performing an appearance inspection of a product in the middle of manufacture, and the fatality of the defect is predicted and determined with high probability based on the detected defect image. Correct any defects that are predicted to be high or discard the device. As a result, the risk of discarding non-defective products and the risk of processing into defective products can be reduced, and the cost reduction effect can be obtained.

  Further, according to the present invention, a fatal factor is extracted based on an image of a defect in a product being manufactured, and process management is performed using this information. In other words, by selecting defect types and features that have a great influence on the lethality and monitoring the histogram, it becomes possible to detect abnormal signs and take measures based on the defect types and features that have changed.

  Further, according to the present invention, a fatal factor is extracted for each product type based on a defect image of a product being manufactured, and feedback to the design is performed using this information. In other words, by comparing the lethality factors between varieties, the resistance to various defects (which is unlikely to be defective) can be understood, so the design information of the varieties with low tolerance can be modified with reference to the design information of the varieties with high tolerance. This makes it possible to manufacture a device that is resistant to defects.

  Embodiments of a defect classification method, a defect classification apparatus, and a device manufacturing method according to the present invention will be described with reference to the drawings.

[First Embodiment]
A first embodiment of the present invention will be described.

  A simple configuration of a PDP (Plasma Display Panel) that is an object of the first embodiment will be described with reference to FIG. Striped rib barriers 102 are formed on the back glass 101, and the phosphor layers 103 that emit light of three colors of RGB are filled therein. A front plate glass 105 is arranged on the upper part of the rib, and gas is sealed in a gap with the back plate. A plasma discharge 107 is caused between the transparent electrode 106 on the front plate and the address electrode 104 in the back plate orthogonal to the front plate to generate ultraviolet rays. The phosphor in each pixel is excited to emit light by the ultraviolet rays, and the light emitting pixel 108 is configured to produce an image.

  Next, the manufacturing process of the PDP according to the first embodiment will be described with reference to FIGS. First, in the front plate process shown in FIG. 2A, after the glass substrate is washed (S100), a transparent ITO (Indium Tin Oxide) electrode 106 is formed by sputtering (S101). Next, a bus electrode is formed by photolithography (photomask molding => etching) or the like (S102). And after apply | coating and baking a dielectric film (S103-S105), the MgO film | membrane which is a protective film is vapor-deposited and formed into a film (S106).

  Similarly, the back plate process shown in FIG. 2B begins with glass substrate cleaning (S200), and after forming address electrodes 104 by photolithography (S201 to S206), a dielectric film is formed (S207). Thereafter, unlike the front plate process, a rib material is printed and dried to form a rib layer (S208), and then a sandblast mask is formed (S209). By rib blasting (S210), ribs are formed and then fired to complete the rib barrier 102 (S211). The phosphor paste 103 is filled into the rib barrier by printing or the like, and baked to fix the phosphor in the rib barrier (S212).

  In the manufacturing process of the front plate and the back plate, although not shown in FIG. 2, an appearance inspection is performed by using an appearance inspection device 41 and an inspection device (including an appearance class classifier and a function class classifier) 42 during the production. Do. A manufacturing process management device (not shown) corrects a defect that is predicted to be highly fatal based on a fatality factor by using the appearance inspection result obtained from the appearance inspection device 41 and the inspection device 42 during the manufacturing process. If it cannot be corrected, the panel is discarded and the manufacturing process is controlled so that defective products are not created. In addition, the manufacturing process management device (not shown) or the inspection device 42 monitors the status of the manufacturing apparatus and process by monitoring the state of defect occurrence. Appearance inspection device 41 performs appearance inspection after electrode pattern formation (ITO, bus, address), dielectric film formation, rib formation / sintering, phosphor coating, MgO film deposition, and foreign matter, scratches, voids The position coordinates of defects such as pattern formation defects and the defect images thereof are detected. After that, in the inspection device 42, the appearance class which is a defect type classified based on the feature amount of each defect by reviewing the defect based on the position coordinates of the defect detected in the appearance inspection, the defect image, and the like. The appearance class classifiers 404 and 406 are learned by teaching as classification conditions, and the appearance class that is the defect type is classified using the appearance class classifier in the learned classification conditions 4061. According to the appearance class, the fatality of the defect type is appropriately set based on the fatality judgment condition 4081 determined by the function classifiers 407 and 408 based on the fatality factor 4071 which is the quantitative description, and Predict and judge with high probability, confirm whether there is no problem, correct or discard based on the fatality of the defect type judged and judged, and the result is It is fed back to the device (not shown).

  As described above, when the front plate and the back plate are completed, in the assembly process shown in FIG. 2 (c), after both are assembled and sealed (S300), vacuuming and discharge gas introduction are performed and sealing is performed. (S301). Then, the drive circuit is attached to the panel (S302), assembled as a TV set (S303), and completed.

  After completion or after mounting the drive circuit panel, the function inspection device 43 performs a lighting inspection as a function inspection. In the lighting inspection, a test pattern is displayed, and a malfunction (lighting failure) such as a black spot that does not light up, a bright spot that always lights up, or display unevenness is detected.

  Here, an appearance inspection method performed in the middle of manufacturing so as not to make a defective product, which is a feature of the present invention, will be briefly described with reference to FIG. FIG. 3 is a schematic configuration diagram showing an example of an appearance inspection apparatus after phosphor application, for example.

  For example, in the appearance inspection apparatus 41 after applying the phosphor, the phosphor is excited by the ultraviolet light irradiation optical system 302 that irradiates the back plate 301 on which the phosphor has been formed with the ultraviolet light 303 and the irradiated ultraviolet light. A detection optical system 305 having a detection camera for detecting emitted light 304 that emits light, and detecting a detection signal detected by the detection camera and detecting a defect state by image processing (defect coordinate data and an image including a defect). Acquisition) and a signal processing system 306. The appearance inspection apparatus 41 performs an appearance inspection on the entire back plate by continuously scanning, for example, the back plate 301 or the detection optical system 305 including the ultraviolet light irradiation optical system 302 and the detection camera. Note that a detection camera may be a detection method in a direction perpendicular to the substrate 301 or a detection method from a low angle of 45 ° or less.

  Next, a first embodiment of an appearance inspection system according to the present invention will be described with reference to FIGS.

  FIG. 4 and FIG. 12 are diagrams showing a first embodiment of a system configuration for realizing the appearance inspection method according to the present invention. Appearance inspection device 41 has an appearance after each electrode plate is formed (ITO, bus, address), dielectric film is formed, ribs are formed / sintered, phosphor is applied, and MgO film is deposited on each of the front plate and the back plate. Inspection is performed to detect defects such as foreign matter, scratches, voids, pattern formation defects, and the like, and coordinate data 411 of the detected defects is transmitted to the inspection apparatus. The inspection device (fatality determination device: including an appearance class classifier and a function class classifier) 42 is configured by a computer that is operated according to a program, and functionally a review device 401, a defect area extraction unit 402, a feature calculation unit. 403, a defect classification unit 404, a criticality determination unit 405, a classification condition setting unit 406, a criticality factor extraction unit 407, and a criticality determination condition setting unit 408. Therefore, the review apparatus 401 may use a stage, an irradiation optical system, and a detection optical system used in the appearance inspection apparatus 41 in combination.

  At the time of learning (at the time of setting conditions), the inspection device 42 and the like include an appearance inspection device 41 or a review device 401 that detects an image signal by imaging a plurality of defects generated in the learning object, and the detected image signal The defect region extraction unit 402 and the feature calculation unit 403 that calculate the feature amount of each defect based on 411, and the appearance class that is a defect type classified based on the feature amount of each defect are taught as the classification condition 4061. The classification condition setting unit 406 that learns the appearance class classifier (defect classification unit) 404 by using the function classifying device 43 for each appearance class that is the defect type classified based on the learned feature amount of the defect type. The function class classifier is learned by teaching the function class correspondence 431 that is the quality of the function determined by the final function test using the function to determine the quality of the final function. A lethality factor extraction unit 407 that extracts a fatality factor 4071 that is a quantitative description of the appearance characteristic during the examination to the right, and a lethality factor based on the extracted fatality factor 4071 that is the quantitative description. A criticality determination condition setting unit 408 that determines a determination condition (criticality determination criterion) 4081. That is, an image is acquired by the review device 401 or the like, a defect region extraction unit 402 extracts a defect region in the acquired image, and a feature calculation unit 403 calculates a feature in the extracted defect region. Then, visual defect classification is performed using the GUI or the like of the review apparatus 401, and defect type information is added to the extracted image in each defect area. The defect classification condition setting unit 406 uses the information on the defect type added by the review device 401 and the feature amount data corresponding to the defect type (appearance class) calculated by the feature calculation unit 403 as learning data. By performing learning, the defect type classification condition 4061 set by the defect classification unit (appearance class classifier) 404 is set.

  At the time of inspection during manufacturing, the inspection device 42 and the like include an appearance inspection device 41 or a review device 401 that detects an image signal by imaging a plurality of defects generated in an inspection object or product during manufacturing, and the detected image. A defect region extraction unit 402 and a feature calculation unit 403 that calculate the feature amount of each defect based on the signal, and an appearance that is the defect type in the learned classification condition 4061 based on the calculated feature amount of each defect A defect classification unit (appearance class classifier) 404 that performs class classification, and the fatality determination condition (based on the fatality factor 4071 that is the quantitative description) for each appearance class in which the classification is performed ( And a fatality determination unit 405 that predicts and determines the fatality based on (Fatality determination criteria) 4081. That is, at the time of inspection during manufacturing, the review device 401 receives coordinate data 411 of various defects such as foreign matter, scratches, voids, pattern formation defects and the like from the appearance inspection device 41, and a stage (not shown) at the position of the coordinates. To obtain various images including various defects. The defect area extraction unit 402 extracts defect areas in which various defects exist from various images acquired by the review apparatus 401, and the feature calculation unit 403 performs various processes from images of various defect areas extracted by the defect area extraction unit 402. A feature amount (details will be described later) of a defective region is calculated. The defect classification unit 404 classifies defect types based on the classification condition 4061 set in advance by the classification condition setting unit 406 based on the feature amount of the defect area calculated and input by the feature calculation unit 403. For each defect type classified by the defect classification unit (appearance class classifier) 404, the criticality determination unit 405 is manufactured based on the feature amount according to the criticality determination condition 4081 preset by the criticality determination condition setting unit 408. The criticality prediction determination is performed in the middle, and the fatality prediction determination result for each defect type is quickly fed back to the manufacturing process to prevent defective products from being manufactured.

  The criticality factor extraction unit 407, which is a feature of the present invention, includes defect types added by the review device 401 and the like, final function inspection quality information 431 obtained from the function inspection device 43, and corresponding classification condition setting units. The feature amount data obtained from 406 is input, and for each defect type, a quantitative description of appearance features that influence the fatality by calculation for each feature (for example, non-defective products and feature values shown in FIGS. 10 and 11). The probability density distribution of defective products or the fatality shown in FIG. 12) is output as the fatality factor 4071. Note that the lethal factor 4071 is extracted for each defect type. The criticality determination condition setting unit 408 sets a criticality determination condition (criticality determination criterion) 4081 for the passability determination unit 405 to make a pass / fail prediction determination for each defect type based on them. The function inspection device 43 performs lighting inspection of the completed PDP panel. In the fatality factor extraction unit 407, for example, the position coordinates 431 of the defective lighting pixels obtained by the function inspection are matched with the position coordinates 411 of the defects obtained by the appearance inspection apparatus 41, so that the quality of the final function inspection is determined. Information 431 and feature amount data obtained from the classification condition setting unit 406 are associated with each other.

  Some appearance inspection apparatuses 41 cut out and store an image of a defective portion at the same time as detecting a defect. When the image acquired by the appearance inspection apparatus 41 is used, the review apparatus 401 inputs and displays the image acquired by the appearance inspection apparatus 41 without performing image acquisition. In addition, some appearance inspection apparatuses 41 cut out images of defective portions simultaneously with defect detection, perform feature calculation, and store them. When using the feature amount acquired by the appearance inspection apparatus 41, the inspection apparatus 42 does not perform defect area extraction by the defect area extraction unit 402 and feature calculation by the feature calculation unit 403, and the feature amount data is directly subjected to appearance inspection. Input from the device 41.

  Next, the operation of each part of the inspection apparatus (fatality determination apparatus) 42 according to the present invention will be described. That is, the review device 401 captures an image of a learning object (learning product) and an inspection object (product) or inputs them from the appearance inspection device 41 and acquires an image of a defective portion.

  Next, a defect area extraction unit 402 that inputs an image from the review apparatus 401 and performs a defect area extraction process on the input image will be described. The defective area extraction unit 402 converts the input image into a grayscale image when the input image is a color image.

  That is, an example of defect area extraction processing in the defect area extraction unit 402 will be described with reference to the flowchart shown in FIG. The defect area extraction unit 402 first performs shading correction on the input processing target image (detected image) (S401), and then creates a reference image (S402). A good product image may be stored in advance and used as a reference image. If the non-defective image cannot be saved, a reference image is created from the original image. Since the PDP repeats the same pattern for each display pixel of the display, the pitch on the image is calculated based on the design information and the imaging magnification, and an image obtained by shifting the original image by one pitch is used as a reference image. If the imaging magnification is not constant, the pitch is calculated using the regularity of the pattern. Next, the original image and the reference image are aligned, a difference is calculated, and a difference image is created (S403). On the other hand, a pattern area image is created from the original image (S404). The most likely pattern area is determined using the brightness information, and the pattern area is sequentially determined using the pitch information. Next, a threshold value is created using the variance of the original image and the difference image (S405). This acts to lower the threshold value in a portion where the brightness is uniform, and to increase the threshold value in a portion where the brightness variation is large, or in the vicinity of the pattern edge. Next, the difference image is binarized and fine noise is removed by expansion / contraction processing (S406). Next, labeling is performed (S407), and a defective area is extracted by selecting a label that seems to be defective (S408). In order to deal with the fact that the defect area is divided into a plurality of areas, a process such as merging those at a certain distance may be added. The degree of defect is determined in consideration of the size and brightness. Further, when the reference image is created from the original image, a label is also extracted by labeling on a defective area on the reference image. In order not to select this, if there is a label at a position shifted by one pitch as opposed to when the reference image is created, it is determined that the defect is on the reference image and is not selected. Through the above processing, the defect area extraction unit 402 extracts various defect areas in the detected image.

  As another example of the defect region extraction processing method in the defect region extraction unit 402, defect detection based on saliency is conceivable. In this method, a multi-dimensional pixel feature amount is calculated, and the distance from the pixel having the closest distance in the feature space is defined as the saliency, and if the saliency exceeds a predetermined threshold, it is determined as a defective pixel. Pixel features include various filter responses such as Gaussian and differential, SIFT (Scale Invariant Feature Transform) features that are invariant to rotation, scale changes, and illumination changes, and HOG (Histograms of Oriented Gradient) features that histogram the luminance gradient of local regions. One value is calculated for each pixel.

  Next, the feature calculation unit 403 that calculates the feature amounts of various defects using the information on the various extracted defect areas will be described. An embodiment for calculating the feature quantities of the various defects will be described with reference to FIGS.

  FIG. 6 is a diagram for explaining characteristics relating to the size and shape of defects. A square corresponds to a pixel, and a hatched pixel represents an extracted defect area. The size-related features include the number of pixels of the defective area, that is, the area, the size in the X and Y directions, the length, the width, the perimeter, and the like. As the length, the longest part is measured by the distance between the contour points. The width is measured at the portion where the distance is maximum with two straight lines (indicated by dotted lines) parallel to the line segment passing through the contour point and measuring the length. The perimeter is the total length of line segments connecting adjacent contour points.

  FIG. 7 represents the original gray image corresponding to the defect area. A feature relating to the brightness of the defect is calculated from the grayscale image. In addition to the maximum luminance and the minimum luminance shown in the figure, the average luminance of the pixels corresponding to the defect area is calculated. Further, the maximum value of the difference from the reference image in the defect area, the luminance variation, and the maximum luminance inclination are calculated. Although not shown, the maximum brightness, the minimum brightness, the average brightness, and the brightness variation of the reference image corresponding to the defect area are calculated as the features related to the brightness of the background.

  FIG. 8 shows a filter used to calculate features related to defects and background texture. Using these filters, the edge strength for each direction and the edge direction frequency maximum value are calculated. The edge strength for each direction is calculated by calculating the response of each pixel using one of the filters shown in FIG. 8 and summing up the pixels corresponding to the defective area. The filter response is a numerical value calculated by summing the center of the filter on the target pixel and the product of the pixel values at the position where the filter numerical value overlaps. Since the original image and the reference image are calculated using four-direction filters, eight features are calculated. For the edge direction frequency maximum value, for each pixel corresponding to the defective area, the maximum response direction among the four filters is set as the edge direction, a histogram of the edge direction is calculated, and the maximum frequency value is calculated.

  FIG. 9 is a diagram for explaining the feature representing the positional relationship between the defect and the pattern. In FIG. 9, reference numeral 901 denotes a pattern, and reference numerals 902a to 902g denote defects. As shown in FIG. 9, the inner distance and the outer distance are defined from the edge of the pattern. When considering the right edge of the pattern as a reference as shown in FIG. 9, the distance from the edge to the left edge of the defect is the inner distance, the distance from the defect right edge is the outer distance, the inner distance is positive to the left, the outer distance is The right direction is positive. The inner distance and the outer distance are shown on the right side of each defect using the symbols a to k and the pattern width w in the figure. The positional relationship between the defect and the pattern is any one of the defects 902a to 902g, those obtained by horizontally inverting these, or those rotated 90 degrees. Similarly, the definition of the inner distance and the outer distance may be considered by turning left or right or rotating 90 degrees.

  In addition, when the input image is a color image, the feature calculation unit 403 decomposes the input image into color components, calculates a color histogram of pixels corresponding to the extracted defect area, and represents the color of the defect. To do.

  Further, as another embodiment of the processing method in the feature calculation unit 403, a method called bug of future can be considered. In this method, as described above, a large number of pixel feature amounts are calculated, a large number of pixels are collected from the learning data, clusters are defined by clustering, and the clusters for each cluster are examined by examining which cluster each defective image pixel is. Calculate and characterize.

  Next, the defect classification unit 404 that classifies defect types that can be classified by appearance such as scratches, foreign matter, pattern defects, and protrusions will be described. Since the generation process of each defect is different, by managing the number of generations for each defect type, it is possible to determine the cause of the most problematic defect and the countermeasure. The defect classifying unit (appearance class classifying unit) 404 uses a teaching type classifier, and feature amounts for various defect areas calculated by the feature calculating unit 403 obtained by learning in advance by the classification condition setting unit 406. The defect type is classified according to the classification standard (classification condition) 4061 corresponding to.

  Next, the classification condition setting unit 406 that performs learning of the teaching type classifier using the defect type information and the feature amount data corresponding thereto as teaching data will be described. The defect type information is added by visually checking the defect image with the review device 401. The association with the feature amount data calculated from the defect image is possible by attaching a defect ID to various information. The number of teaching data is preferably 10 or more per class.

  There are various teaching type classifiers such as a k-nearest neighbor method, a minimum distance method, a subspace method, a local subspace method, a neural network, and a support vector machine. Moreover, you may combine them and may perform ensemble learning, such as bagging and boosting.

Next, the criticality factor extraction unit 407 and the criticality determination condition setting unit 408, which are features of the present invention, will be described. The fatality factor extraction unit 407 is based on past data, the defect type added by the review device 401 or the like, the final function inspection pass / fail information 431 obtained from the function inspection device 43, and the corresponding classification condition setting unit Using the feature data obtained from 406 as an input, let us know the criticality factor (determining quantitative description of external features in the middle of production that influence the quality of the final function: for example, the feature values shown in FIGS. 10 and 11 The probability density distribution of the non-defective product and the defective product or the fatality rate with respect to the feature value shown in FIG. 12) 4071 is extracted. Specifically, the learning data is divided into defect types, and for each defect type, a non-defective product probability density distribution and a non-defective product probability density distribution are calculated for each appearance feature that is a fatal factor 4071, and a GUI (Graphic User Interface) is calculated. ) Part (not shown), for example, as a graph. The feature value is normalized at the maximum and minimum. For simple calculation, it is only necessary to determine an equally divided section, calculate a histogram of good products and defective products, and divide them by the number of good products and the number of defective products, respectively. FIG. 10 shows an example of graph display. The horizontal axis represents feature values, and the vertical axis represents probability density. X _{1 to} X ₁₀ represent sections. The black of the bar graph represents the probability density distribution of good products, and the shaded area represents the probability density distribution of defective products.

  The probability density distribution of non-defective products and the probability density distribution of defective products are GHJohn and P. Langley: Estimating Continuous Distributions in Bayesian Classifiers, Proceedings of the 11th Conference on Uncertainty in Artificial Intelligence, pp338-345 (1995) According to the literature 1), it may be calculated using the following equation (1).

つまり、特徴値の確率密度分布とは、あるデータがクラスｋ（良品または不良品）に属する場合に特徴値がｘとなる条件付確率であるととらえる。ここで、ω_ｋはクラスｋ（良品または不良品）に属するデータの集合であり、ｎ_ｋはクラスｋ（良品または不良品）のデータの個数である。ｘ_ｊはクラスｋ（良品または不良品）に属するｊ番目の学習データを表す。図１１には、ＧＵＩ部（図示せず）に表示したグラフ表示の実施例を示す。横軸は特徴値、縦軸は確率密度分布を表す。良品を黒のラインで表し、不良品を網掛けのラインで表す。確率密度分布は、上記（１）式より、学習データ１個１個を中心とするガウス分布の和として計算されるため、ヒストグラムを利用するよりも、計算時間は多いが滑らかな分布が算出されるという特徴がある。

That is, the probability density distribution of feature values is regarded as a conditional probability that the feature value is x when certain data belongs to class k (good product or defective product). Here, ω _k is a set of data belonging to class k (good product or defective product), and _nk is the number of data of class k (good product or defective product). x _j represents j-th learning data belonging to class k (good or defective). FIG. 11 shows an example of a graph display displayed on a GUI unit (not shown). The horizontal axis represents the feature value, and the vertical axis represents the probability density distribution. Non-defective products are represented by black lines, and defective products are represented by shaded lines. Since the probability density distribution is calculated as the sum of Gaussian distributions centered on each piece of learning data from the above equation (1), a smooth distribution is calculated although the calculation time is longer than using a histogram. There is a feature that.

  In order to compare features, the error rate is calculated based on the probability density distribution. The error rate is the ratio of the area of the overlapping portion of the distribution of non-defective products and defective products shown by hatching in FIG. 11 to the total area. In other words, it represents the probability of misclassification under the criterion of classifying into a class having a higher probability density when given a feature value of unknown data. It can be said that the lower the error rate, the more external features that influence the fatality.

The lethality determination condition setting unit 408 is based on the lethality factor calculated by the lethality factor extracting unit 407 (quantitative description on the appearance characteristics during the production that influence the quality of the final function: for example, the formula (1) Based on the probability density distribution 4071 obtained in this manner, the criticality determination condition (criticality determination criterion) 4081 is determined. As a method for determining the lethality determination condition 4081 based on the above information, the flexible naive Bayes method described in Non-Patent Document 1 can be considered. In the flexible naive Bayes method, when a defect feature x (vector) having an unknown lethality is given, the defect belongs to a class k (good product or defective product) having the maximum probability P (ω _k | x). The class is determined to be a class k (non-defective product or defective product) having the maximum P (x | ω _k ) estimated from the learning data based on Bayes' theorem expressed by the following equation (2). It is determined that the class belongs to.

なお、特徴ベクトルｘのクラスｋ（良品または不良品）の確率密度分布Ｐ（ｘ｜ω_ｋ）は、特徴毎のクラスｋの確率密度の積として、次の（３）式により求められる。

Note that the probability density distribution P (x | ω _k ) of the class k (non-defective product or defective product) of the feature vector x is obtained by the following equation (3) as the product of the probability density of the class k for each feature.

ここで、ｘ_ｉは致命性が未知の欠陥の特徴ｘのｉ番目の要素を表す。また、ｘ_ｉｊはクラスｋに属するｊ番目の学習データのｉ番目の要素を表す。なお、全ての特徴を用いてもよいが、エラー率の低い特徴（図１１にハッチングで示す良品と不良品の分布の重なった部分の面積の全面積に対する割合が小さい特徴）を上位何個か選んでおけば、致命性に無関係の特徴を用いないことにより、判定精度が向上する可能性がある。

Here, x _i represents the i-th element of the feature x of the defect whose fatality is unknown. X _ij represents the i-th element of the j-th learning data belonging to the class k. Although all the features may be used, some of the top features that have a low error rate (features in which the ratio of the area where the distribution of the non-defective product and the defective product shown in FIG. If selected, the determination accuracy may be improved by not using a feature unrelated to fatality.

  A description will be given of another example of a quantitative description of appearance characteristics that influence the lethality in the lethality factor extracting unit 407, that is, a method for calculating the lethality factor. In this method, a threshold value vs. fatality rate is calculated for each appearance feature and displayed on a graph, for example. First, find the total number of good and defective products. Let them be Nok and Nng, respectively. The feature value is normalized with the maximum and minimum. Sort data in descending order of normalized feature values. Sequentially check one by one and count the number of non-defective products and defective products up to that point to Nok + and Nng +. The plus sign is added to mean that it is above the threshold. By calculating Nng + / (Nok +++ Nng +), the fatality rate is equal to or higher than the threshold value. Let Nok-Nok + be Nok- and Nng-Nng + be Nng-. The minus sign is given to mean less than the threshold. By calculating Nng − / (Nok− + Nng−), the fatality rate is less than the threshold value. As shown in FIG. 12, the fatality rate calculated on the horizontal axis for the feature value at that time is plotted on the vertical axis. Then, as indicated by reference numeral 1201 in FIG. 12, it can be said that it is an external feature that influences the fatality as the rise or fall of the fatality rate becomes sharper.

  As a method for determining the fatality determination condition (fatality determination criterion) 4081 based on the fatality factor 4071 in the fatality determination condition setting unit 408, threshold processing for each feature can be considered. Find the range in which the fatality rate is 100% by checking from the larger or smaller of each feature, and determine whether the threshold for this feature should be judged as fatal or above or below the threshold decide. If there is no place where the lethality rate becomes 100%, the feature is not used for the lethality judgment. If there is at least one feature that is determined to be fatal using all the remaining features, the defect is determined to be fatal. If there is no range in which the fatality rate is 100% for all features, an appropriate numerical value with a fatality rate of less than 100% is obtained, for example, a range in which 90% or more is obtained, and the criticality determination for this feature is performed as described above. A condition (fatal judgment criterion) 4081 is determined. In this case, if a determination is made using one feature, an error that determines a non-fatal defect as fatal occurs. That is, 10% of the cases determined to be fatal are errors. In order to reduce the error rate, for example, when it is determined to be fatal using two or more features, it may be determined to be fatal.

  As described above, the fatality determination condition setting unit 408 has described the method of setting the fatality determination condition 4081 based only on the teaching data. However, the determination threshold value can be adjusted using the knowledge of a person, It is good to be able to add or delete features used for.

  As described above, since the criticality determination condition (criticality determination criterion) 4081 is set by the criticality determination condition setting unit 408 based on the criticality factor 4071, the criticality determination unit 405 has an appearance during the manufacturing process. Appearance inspection is performed by the inspection apparatus 41, and the criticality of each defect type is appropriately predicted and determined according to the criticality determination condition 4081 based on the feature amount of each defect type obtained from the defect classification unit (appearance class classifier) 404. (Pass / Fail judgment) is performed, based on the fatality of each predicted defect type, whether there is a problem, correction, or disposal is confirmed, and the result is fed back to a manufacturing process management device (not shown) By doing so, it becomes possible to remove a defect that finally becomes a malfunction in an inspection during manufacturing.

  In addition, it is possible to realize an appearance inspection method and an appearance inspection apparatus capable of determining the fatality of a defect with an appropriate standard during the manufacturing process. As a result, it is possible to realize a device manufacturing line capable of suppressing waste that erroneously discards non-defective products and waste that causes defective products to flow to a subsequent process, and an effect of cost reduction can be obtained.

  Further, it is possible to realize an appearance inspection method and an appearance inspection apparatus that can obtain knowledge about the fatality factor according to the defect type. An embodiment of a device manufacturing line using the fatality factor 4071 extracted by the fatality factor extraction unit 407 will be described.

  First, an embodiment in which the fatality factor 4071 extracted by the fatality factor extraction unit 407 is used for quantitative process management in the device manufacturing line will be described. In the learning process, the lethality determination condition setting unit 408 selects an external feature having a large influence on the lethality as the lethality determination condition 4081 based on the fatality factor information 4071 extracted by the lethality factor extraction unit 407. . The lethality determination condition setting unit 408 further calculates a threshold value with less misjudgment as the lethality determination condition 4081 when determining the lethality for each selected feature using the feature alone. Keep it. Further, the fatality determination condition setting unit 408 further calculates a second threshold value as a fatality determination condition 4081 by shifting the threshold value by a predetermined ratio to the low fatality side.

  In the inspection process of process management execution, in the same process in the middle of manufacturing, the defect position is detected by the appearance inspection device 41 and the inspection device 42, the defect area extraction unit 402 acquires the defect image, and the feature calculation unit 403 detects the defect. Calculate features. The lethality determination unit 405 and the like calculate and record a histogram for each feature selected as the lethality determination condition 4081 at regular intervals, for example, once a day. The fatality determination unit 405 or the like monitors the transition of this information and detects a sign of the occurrence of a fatal defect by detecting a shift in the frequency distribution toward the higher fatality. Also, the fatality determination unit 405 and the like count and record the defect having higher fatality that is equal to or higher than the second threshold value or less than the second threshold value. It can be said that the number is the number of reserves of fatal defects. This is monitored, and when there is an increase in specific features and defect types, the cause is investigated and countermeasures are taken using that information as a clue. Note that the cause may be examined using the information as a clue and the countermeasure taken may be executed not by the inspection apparatus 42 but by a production line management apparatus (not shown).

  Next, an embodiment in which the criticality factor 4071 extracted by the criticality factor extraction unit 407 is used for feedback to the design will be described. The influence of the defect characteristics on the criticality varies depending on the defect type, but also varies depending on the size and shape of the product pattern. Therefore, the processing in the inspection apparatus 42 according to the present invention is preferably performed for each product type, and here, it is assumed that the processing is performed for each type. Therefore, at the time of inspection, the fatality determination unit 405 and the like are learned for each defect type (appearance class) classified by the defect classification unit (appearance class classifier) 404 and are learned by the fatality factor extraction unit (function class). The product class of the fatality factor 4071, which is the quantitative description output from the classifier 407, is compared between the product types, and the resistance of each product type to the fatality factor obtained based on the comparison (becomes defective). Feedback) to the design. That is, the fatality determination unit 405 and the like take out a certain defect type and appearance feature, and compare the threshold values calculated by the same method as described above, thereby comparing the tolerance of each type to the defect type and feature. be able to. A device resistant to defects can be manufactured by correcting the design of the low-tolerance product type with reference to the design information of the high-tolerance product type.

  As described above, the manufacture of the PDP has been described as an example, but the gist of the invention described here can be applied to other flat panel displays such as a liquid crystal panel and an organic EL.

[Second Embodiment]
The second embodiment according to the present invention is directed to semiconductor manufacturing.

  The semiconductor manufacturing process of the second embodiment will be described with reference to FIG. The semiconductor manufacturing process is roughly divided into a transistor forming process shown in FIG. 13A and a wiring forming process shown in FIG. 13B. The transistor formation process starts with isolation (S501), and is processed in the order of well formation by ion implantation (S502), gate insulating film and gate electrode formation (S503), and source and drain formation (S504). In the wiring formation step, a wiring pattern is formed by photolithography (S601), an insulating film is formed thereon (S602), and planarized by CMP (Chemical Mechanical Polishing) (S603). An interlayer connection hole called a contact hole is formed by photolithography (S604), metal is buried (S605), and planarization is performed by CMP (S606). This process is repeated to form a multilayer wiring. Various defects occur in the middle of these processes. The main defects are foreign matter, scratches, short and open pattern defects, and interlayer connection failures. In the semiconductor manufacturing process, in order to maintain and improve the yield, these defects are detected at an early stage by in-line wafer inspection, and measures are taken.

  In-line wafer inspection consists of defect determination by appearance inspection and defect classification by a review device. A wafer in the middle of manufacture is extracted, and the appearance of the wafer is inspected using an optical or electron beam inspection device to detect the position of the defect. The detected defects are reviewed with a high-resolution scanning electron microscope, and the types of defects occurring are identified by performing defect classification.

  A second embodiment of the appearance inspection system according to the present invention will be described with reference to FIG. FIG. 14 shows a second embodiment of an optical appearance inspection system configuration for a semiconductor wafer. Reference numeral 61 denotes a semiconductor wafer as an object to be inspected, 62 denotes a stage on which the semiconductor wafer 61 is mounted and moved, 63 denotes a detection unit, and a light source 601 for irradiating the semiconductor wafer 61 and light emitted from the light source 601 are collected. The illumination optical system 602, the objective lens 603 that forms an optical image obtained by illuminating the semiconductor wafer 61 with the illumination light condensed by the illumination optical system 602 and reflecting it, and the formed optical image according to the brightness The image sensor 604 converts the image signal. Here, the light source 601 is, for example, a lamp light source or a laser light source, and the image sensor 604 is, for example, a CCD linear sensor, a TDI sensor, or a photomultiplier. Reference numeral 64 denotes an image processing unit that detects defect candidates on a wafer as a sample from the image detected by the detection unit 63.

  The image processing unit 64 converts an input signal from the image sensor 604 of the detection unit 63 into a digital signal, an AD conversion unit 605, and performs image correction such as shading correction and dark level correction on the AD converted digital signal. A processing unit 606 includes a defect determination unit 607 that compares a detected image with a reference image detected from a corresponding position of an adjacent die, and outputs a portion whose difference value is larger than a separately set threshold value as a defect. Is done.

  Reference numeral 65 denotes an overall control unit, a storage device 608 that stores coordinates, feature amounts, images, and the like of detected defects, and a user interface unit that accepts inspection parameter changes from the user and displays detected defect information Reference numeral 609 denotes a CPU that performs various controls. A mechanical controller 610 drives the stage 62 based on a control command from the overall control unit 65. Note that the image processing unit 64 and the detection unit 63 are also driven by commands from the overall control unit 65.

  Reference numeral 66 denotes a review device (appearance inspection system) according to the present invention, which is not included in the inspection device, but can exchange data. A review device (appearance inspection system) 66 that implements the appearance inspection method according to the present invention will be described with reference to FIG. The review device (appearance inspection system) 66 includes a stage 71 on which a semiconductor wafer is mounted and moved, a detection system 72 that captures a high-resolution scanning electron beam image of a defect, an overall control unit 73, and a computer that is operated according to a program. Functionally, it includes a defect area extraction unit 402, a feature calculation unit 403, a defect classification unit 404, a fatality determination unit 405, and a classification condition setting unit 406.

  Hereinafter, processing in the functional inspection device 74, the defect information totaling unit 708, the fatality factor extraction unit 706, and the fatality determination condition setting unit 707, which are different from the first embodiment, will be described.

  In the function inspection device 74, an electrical function test of the semiconductor circuit is performed for each chip by probe inspection, and pass / fail is determined. The defect totaling unit 708 collects defect inspection information of a plurality of processes on the same wafer, and associates them with chip numbers based on position information of each defect. The chip pass / fail judgment information is input from the function inspection device 74, and pass / fail information is added to each defect based on the association between the defect and the chip number. Here, it should be noted that a plurality of defects are associated with one chip. Even if there are a plurality of defects associated with an acceptable chip, all of them are labeled as “good”. Further, if the defect associated with the rejected chip is only one defect, it is labeled “defective product”. In the case of multiple, the quality is unclear.

  The criticality factor extraction unit 706 performs the same processing as the criticality factor extraction unit 407 in the first embodiment by removing unknown defects, and calculates the probability density distribution of good and defective products of each feature. It is possible to calculate the threshold value versus the fatality rate. However, there is a possibility that there are very few defects that are not known, in particular, “non-defective products”. Therefore, a method of performing processing by attaching a label “defective product” to an unknown defect will be described.

  In the first method, the probability density distribution of the non-defective product and the defective product of each feature is calculated. Since there is a possibility that non-defective products are mixed in the distribution of defective products, the distribution of non-defective products is trusted, and if the probability density of non-defective products exceeds a threshold value close to 0, the probability density of defective products is forced to zero. To. It can be said that the larger the integral value of the part where the probability density of defective products is not 0, the more critical the influence of the fatality. The condition for determining the lethality based on this is defined by the above equation (3) using the probability density distribution of each feature corrected as described above.

  In the second method, a threshold value versus fatality rate is calculated for each appearance feature, and a threshold value for each feature is determined based on the threshold value. Find the range where the non-fatal rate is 100% by checking from the larger or smaller of each feature, and decide whether the threshold value for this feature should be judged as non-fatal To do. If there is no place where the non-lethal rate becomes 100%, the feature is not used for determination. If there is at least one feature determined to be non-fatal using all of the remaining features, the defect is determined to be non-fatal.

  The fatality factor 4071 extracted in this way can be used for process management and feedback to design, as in the first embodiment.

[Third Embodiment]
A third embodiment according to the present invention will be described with reference to FIG.

  FIG. 16 is a diagram showing a system configuration for realizing the third embodiment of the appearance inspection method according to the present invention. The third embodiment has a configuration in which the defect classification unit 404 and the classification condition setting unit 406 are excluded from the first embodiment shown in FIG. Further, an image acquisition device 409 is included instead of the review device 401. As long as an image can be acquired, a review device may be used as the image acquisition device 409. The appearance inspection device 41 detects a defect and sends defect coordinate data to an inspection device (fatality determination device) 42. At the time of learning and inspection, the image acquisition device 409 in the inspection device 42 acquires an image of a defective portion based on the defect coordinate data. A defect area is extracted from the image by the defect area extraction unit 402, and a feature amount is calculated from the image by the feature calculation unit 403.

  At the time of inspection, the criticality determination unit 405 performs criticality determination according to the criticality determination condition 4081 preset by the criticality determination condition setting unit 408 by learning based on the feature amount calculated by the feature calculation unit 403. .

  At the time of learning (condition setting), the fatality factor extraction unit 407 constituting the function class classifier inputs the defect type, the final function inspection pass / fail information, and the corresponding feature amount data, and for each feature. By the calculation, a quantitative description regarding the external feature that affects the fatality is output as the fatality factor 4071. Further, the fatality determination condition setting unit 408 constituting the function class classifier sets the fatality determination condition 4081 based on them. The association between the final function inspection pass / fail information and the feature amount data is obtained by matching the position coordinates of the defective portion obtained by the function inspection apparatus 43 with the position coordinates of the defect obtained by the appearance inspection apparatus 41.

  As described above, according to the third embodiment, since the criticality determination is performed not for each defect type but for the entire defect type, there is a possibility that the accuracy of the criticality determination may be reduced. This eliminates the need for confirmation of the criticality factor, so that criticality factor extraction and criticality determination condition setting can be performed in a short time.

  The present invention can be used for manufacturing thin displays such as PDP and TFT, thin film devices such as semiconductor wafers and hard disk substrates, photomasks, and films.

It is explanatory drawing of the structure of PDP which concerns on this invention. It is a flowchart for demonstrating the manufacturing process of PDP which concerns on this invention. It is a schematic block diagram which shows one Example of the external appearance inspection apparatus of PDP which concerns on this invention. It is a lineblock diagram showing a 1st embodiment of an appearance inspection system concerning the present invention. It is a flowchart explaining one Example of the processing method in the defect area | region extraction part of the inspection apparatus which concerns on this invention. It is explanatory drawing of the characteristic regarding the magnitude | size of the defect which concerns on this invention. It is explanatory drawing of the characteristic regarding the brightness of the defect which concerns on this invention. It is a figure which shows the filter for extracting the characteristic regarding the defect and background texture which concern on this invention. It is explanatory drawing of the characteristic showing the positional relationship of the defect which concerns on this invention, and a pattern. It is a graph display which shows the Example of the probability density distribution of the quality goods and inferior goods with respect to the feature value which is the quantitative description regarding the external appearance characteristic which influences fatality about each defect kind which concerns on this invention. It is a graph display which shows the other Example of the probability density distribution of the quality goods and inferior goods with respect to the feature value which is the quantitative description regarding the external appearance characteristic which influences fatality about each defect kind which concerns on this invention. It is a figure which shows the fatality rate with respect to the feature value which is the quantitative description regarding the external feature which influences fatality about 1st Embodiment of the external appearance inspection system which concerns on this invention, and each defect kind. It is a flowchart for demonstrating the manufacturing process of the semiconductor wafer which concerns on this invention. It is a schematic block diagram which shows one Example of the external appearance inspection apparatus of the semiconductor wafer which concerns on this invention. It is a block diagram which shows 2nd Embodiment of the external appearance inspection system (review apparatus) which concerns on this invention. It is a block diagram which shows 3rd Embodiment of the external appearance inspection system which concerns on this invention.

DESCRIPTION OF SYMBOLS 101 ... Back glass, 102 ... Rib barrier, 103 ... Phosphor layer, 104 ... Address electrode, 105 ... Front plate glass, 106 ... Transparent electrode, 107 ... Plasma discharge, 108 ... Light emitting pixel, 301 ... Back plate, 302 ... Ultraviolet light source , 303 ... Irradiation ultraviolet light, 304 ... Phosphor excitation light, 305 ... Camera, 306 ... Signal processing system, 41 ... Appearance inspection device, 42 ... Inspection device (lethality determination device), 43 ... Function inspection device, 401 ... Review device 402 ... Defect area extraction unit 403 ... Feature calculation unit 404 ... Defect classification unit (appearance class classifier) 405 ... Fatality determination unit 406 ... Classification condition setting unit (appearance class classifier) 4061 ... Classification condition, 407 ... Fatal factor extraction unit (functional class classifier), 4071 ... Fatal factor, 408 ... Fatality determination condition setting unit (functional class classifier), 4081 Criticality determination condition (criticality criterion), 409 ... image acquisition device, 901 ... pattern, 902a-g ... defect, 501 ... silicon substrate, 502 ... isolation, 503 ... well, 504 ... gate insulating film, 505 ... gate Electrode, 506... Source / drain electrode, 507... Wiring pattern, 508 .. insulating film, 509 .. contact hole, 61 .. inspection object (semiconductor wafer), 62. Control unit, 66 ... Review device, 601 ... Light source, 602 ... Illumination optical system, 603 ... Objective lens, 604 ... Image sensor, 605 ... AD conversion unit, 606 ... Pre-processing unit, 607 ... Defect determination unit, 608 ... Memory Device: 609 ... User interface unit, 610 ... Mechanical controller, 71 ... Stage, 72 ... Detection unit, 73 ... Control unit, 74 ... function testing device, 706 ... fatality factor extraction unit, 707 ... fatality judgment condition setting unit, 708 ... defect information compiling unit.

Claims (18)

In learning, a first learning is performed in which a plurality of defects generated in an object to be learned are picked up using an appearance inspection apparatus to detect an image signal, and a feature amount of each defect is calculated based on the detected image signal. And a second learning step of learning an appearance class classifier by teaching an appearance class that is a defect type classified based on the feature amount of each defect calculated in the first learning step as a classification condition And, for each appearance class that is the defect type classified based on the feature amount of the defect type learned in the second learning step, the function determined by the final function inspection using the function inspection device By learning the function class classifier by teaching the correspondence of the function class that is good or bad, let us extract the fatal factor that is a quantitative description of the external features during the inspection that influence the quality of the final function. A third learning step of a learning process and a fourth learning step of determining a critical judgment condition based on the criticality factor is a third learning steps the quantitative extracted with description,
At the time of inspection, a plurality of defects generated in an inspection object during manufacture are imaged using an appearance inspection apparatus to detect an image signal, and a feature amount of each defect is calculated based on the detected image signal. The defect type using the appearance class classifier with the classification condition learned in the second learning step based on the feature amount of each defect calculated in the first inspection step and the first inspection step. The second inspection step for classifying the appearance class and the appearance class classified in the second inspection step are determined based on the fatal factor that is the quantitative description in the fourth learning step. And a third inspection step for predicting and determining the criticality based on the determined criticality determination condition.   The appearance inspection method according to claim 1, wherein the learning object and the inspection object are thin displays, and the function inspection is a lighting inspection.   The appearance inspection method according to claim 1, wherein the learning object and the inspection object are semiconductor wafers, and the functional inspection is a probe inspection.   The feature amount of each defect calculated in the first learning step and the first inspection step is calculated based on the extraction of each defect region, and the feature amount representing the size and shape of each defect, each defect, and the background Including at least one of a feature value representing the brightness of the image, a feature value representing the texture of each defect and the background, a feature value representing the positional relationship between each defect and the pattern, and a feature value representing the color of each defect. The visual inspection method according to claim 1.   The extraction of each defect area is performed based on a comparison between each defect image obtained by imaging each defect and a reference image obtained by imaging from the same pattern area not including each defect. The appearance inspection method according to claim 4, wherein   The extraction of each defect area is performed based on the similarity between pixels or blocks of feature vectors calculated in pixel units or block units of each defect image obtained by imaging each defect. The appearance inspection method according to claim 4, wherein   In the third learning step, the fatality factors are extracted by inputting defect types and final function inspection pass / fail information and feature data corresponding thereto, dividing the feature data into defect types, The appearance according to claim 1, wherein the defect type is calculated by calculating a probability density distribution of a non-defective product and a defective product for each feature, and the criticality determination condition is determined based on the calculation of the probability density distribution. Inspection method.   In the third learning step, the fatality factors are extracted by inputting defect types and final function inspection pass / fail information and feature data corresponding thereto, dividing the feature data into defect types, The defect type is calculated by calculating a fatality rate curve with respect to a threshold value for each feature, and the criticality determination condition sets a threshold value for each feature based on the fatality rate curve, and sets the threshold value for each set feature. The appearance inspection method according to claim 1, wherein the appearance inspection method is determined by a combination of threshold processing.   The appearance inspection method according to claim 7 or 8, wherein, in the third learning step, the fatality determination condition to be determined is adjustable. A first learning step of imaging a plurality of defects generated in a learning product using an appearance inspection device to detect an image signal and calculating a feature amount of each defect based on the detected image signal during learning And a second learning step of learning an appearance class classifier by teaching an appearance class that is a defect type classified based on the feature amount of each defect calculated in the first learning step as a classification condition; The quality of the function determined by the final function inspection using the function inspection device for each appearance class that is the defect type classified based on the feature amount of the defect type learned in the second learning step Learn the function class classifier by teaching the correspondence of the function class, and extract the fatal factor that is a quantitative description of the appearance features during the examination that influence the quality of the final function Third learning step and, a learning process and a fourth learning step of determining a critical judgment condition based on the criticality factor is the quantitative description extracted in the learning step the third that,
At the time of inspection, a plurality of defects generated in a product being manufactured are imaged using an appearance inspection apparatus to detect an image signal, and a feature amount of each defect is calculated based on the detected image signal An appearance that is the defect type using an appearance class classifier with a classification condition learned in the first learning process based on the inspection step and the feature amount of each defect calculated in the first inspection step A second inspection step for classifying the class and an appearance class classified in the second inspection step are determined based on the fatal factor that is the quantitative description in the fourth learning step. A third inspection step for predicting and determining the lethality based on the fatality determination condition, and discarding the product based on the fatality prediction determination result for each appearance class obtained in the third inspection step, Correct and next Or flow to step appearance inspection method characterized by it and a testing process and a fourth checking step that determines flow to the next step. A first learning step of imaging a plurality of defects generated in a learning product using an appearance inspection device to detect an image signal and calculating a feature amount of each defect based on the detected image signal during learning And a second learning step of learning an appearance class classifier by teaching an appearance class that is a defect type classified based on the feature amount of each defect calculated in the first learning step as a classification condition; For each appearance class that is the defect type classified based on the feature amount of the defect type learned in the second learning step, the function determined by the final function inspection using the function inspection device Learning the function class classifier by teaching the correspondence of the function class that is good or bad, and extracting the fatal factor that is a quantitative description of the external features during the examination that determines the quality of the final function And a fourth learning step of selecting a defect type and a feature to be monitored based on a fatality factor that is the quantitative description extracted in the third learning step. Process,
At the time of inspection, a plurality of defects generated in a product being manufactured are imaged using an appearance inspection apparatus to detect an image signal, and a feature amount of each defect is calculated based on the detected image signal An appearance class that is the defect type using an appearance class classifier with the classification condition learned in the second learning step based on the inspection step and the feature amount of each defect calculated in the first inspection step The second inspection step for classifying and the appearance classes classified in the second inspection step are selected based on the fatal factor that is the quantitative description in the fourth learning step. In addition, the frequency distribution of the defect types and features to be monitored is calculated and recorded at regular intervals, and a sign of the occurrence of a fatal failure is detected by detecting a shift to a more fatal side by monitoring the frequency distribution. Third test Appearance inspection method characterized by comprising an inspection step of a step. A first learning step of imaging a plurality of defects generated in a learning product using an appearance inspection device to detect an image signal and calculating a feature amount of each defect based on the detected image signal during learning And a second learning step of learning an appearance class classifier by teaching an appearance class that is a defect type classified based on the feature amount of each defect calculated in the first learning step as a classification condition; For each appearance class that is the defect type classified based on the feature amount of the defect type learned in the second learning step, the function determined by the final function inspection using the function inspection device Learning the function class classifier by teaching the correspondence of the function class that is good or bad, and extracting the fatal factor that is a quantitative description of the external features during the examination that determines the quality of the final function And selecting a defect type and a feature to be monitored on the basis of the fatality factor that is the quantitative description extracted in the third learning step, and setting a threshold value for each feature. A learning process comprising: a fourth learning step to:
At the time of inspection, a plurality of defects generated in a product being manufactured are imaged using an appearance inspection apparatus to detect an image signal, and a feature amount of each defect is calculated based on the detected image signal Appearance that is the defect type using an appearance class classifier in the inspection step and the classification condition learned in the second learning step based on the feature amount of each defect calculated in the first inspection step A second inspection step for classifying the class and an appearance class classified in the second inspection step are selected based on the fatal factor that is the quantitative description in the fourth learning step. Record the number of defects on the critical side that are greater than or less than the threshold value and less than the threshold value of the monitored defect type and characteristics, and determine the cause based on the change in the number of defects. Third test to estimate and take measures Appearance inspection method characterized by comprising an inspection step of a step. A first learning step of imaging a plurality of defects generated in a learning product using an appearance inspection device to detect an image signal and calculating a feature amount of each defect based on the detected image signal during learning And a second learning step of learning an appearance class classifier by teaching an appearance class that is a defect type classified based on the feature amount of each defect calculated in the first learning step as a classification condition; For each appearance class that is the defect type classified based on the feature type of the defect type learned in the second learning step and for each type of the learning product, a final function is obtained using a function inspection device. Learn the function class classifier by teaching the correspondence of the function class that is the quality of the function determined by the inspection, and quantitative about the appearance features during the inspection that determine the quality of the final function A learning process and a third learning step of extracting lethality factor is mentioned,
At the time of inspection, a plurality of defects generated in a product being manufactured are imaged using an appearance inspection apparatus to detect an image signal, and a feature amount of each defect is calculated based on the detected image signal Appearance that is the defect type using an appearance class classifier in the inspection step and the classification condition learned in the second learning step based on the feature amount of each defect calculated in the first inspection step A second inspection step for classifying the class, and the quantification that has been learned in the third learning step and output from the functional class classifier for each appearance class that has been classified in the second inspection step. And a third inspection step for feeding back to the design information on the resistance of each product to the lethality factor obtained based on a comparison between the product types of the lethality factor which is a typical description; Appearance inspection method characterized in that it comprises. At each time of learning and inspection, an image signal is detected by imaging a plurality of defects generated in each of the learning object and the inspection object being manufactured, and the feature of each defect is based on the detected image signal An inspection device for calculating the quantity;
At the time of learning, an appearance class classifier that learns by teaching an appearance class that is a defect type classified based on the feature amount of each defect calculated by the inspection apparatus as a classification condition;
A function determined by a final function inspection using a function inspection device for each appearance class that is the defect type classified based on the feature amount of the defect type learned by the appearance class classifier during learning By learning the correspondence of the function class that is the quality of the product, the fatal factor that is a quantitative description regarding the external feature during the inspection that influences the quality of the final function is extracted, and the extracted A functional class classifier that determines the criticality criteria based on criticality factors that are quantitative descriptions;
The inspection apparatus further includes, at the time of inspection, for each defect generated in the inspection object being manufactured, calculated by the inspection apparatus, based on the learned classification condition based on the feature amount of each defect. The appearance class that is the defect type is classified using an appearance class classifier, and the fatal description that is the quantitative description that is learned and output from the functional class classifier for each appearance class that has been classified. An appearance inspection system that predicts and determines fatality based on the fatality determination condition determined based on a sex factor.   The appearance class classifier includes, during learning, a classification condition setting unit that teaches, as a classification condition, an appearance class that is a defect type classified based on a feature amount of each defect calculated by the inspection apparatus; A defect classifying unit configured to classify the appearance class as the defect type using an appearance class classifier under the learned classification condition based on the feature amount of each defect. The visual inspection system according to claim 14.   The appearance class classifier is based on a review device that displays defect images and teaches defect types during learning, a feature amount of each defect calculated by the inspection device, and teaching of the defect types by the review device A classification condition setting unit that teaches an appearance class that is a defect type to be classified as a classification condition, and at the time of inspection, using the appearance class classifier with the learned classification condition based on the feature amount of each defect The appearance inspection system according to claim 14, further comprising a defect classification unit that classifies an appearance class that is a defect type.   The function class classifier extracts a fatal factor that is a quantitative description of an appearance characteristic during inspection that determines the quality of the final function for each appearance class as the defect type during learning. And a fatality determination condition setting unit that determines a fatality determination condition based on the fatality factor that is the quantitative description extracted by the fatality factor extraction section. The appearance inspection system according to claim 14.   The inspection device, at the time of inspection, determines the lethality based on the fatality determination condition determined based on the fatality factor that is the quantitative description output from the function class classifier for each appearance class. The visual inspection system according to claim 14, further comprising a fatality determination unit that performs prediction determination.

Images