Defect inspection device

Abstract

To provide a defect inspection device that is robust against noise and manufacturing errors without using design data during inspection.SOLUTION: A defect inspection device includes: a feature amount calculation unit that calculates a feature amount based on a captured image of a sample; an image information reduction unit that generates a latent variable by reducing an amount of information of the feature amount; a statistics estimation unit that estimates image statistics that a normal image can take based on the latent variable; and a defect detection unit that detects defects in an inspection image based on the image statistics and the inspection image of the sample.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a defect inspection apparatus, for example, an apparatus for inspecting defects or foreign matter contamination using an image.

Background Art

[0002] In semiconductor inspection, SEM images taken with a scanning electron microscope (SEM) are used. As a conventional inspection method, there is a Die-to-Database inspection that detects defects by comparing design data of a semiconductor circuit with an SEM image. However, with the miniaturization of semiconductor circuit patterns, it is becoming difficult to form a circuit pattern as designed on a wafer, and manufacturing errors of the circuit occur. For this reason, false detections occur due to noise and manufacturing errors in the Die-to-Database inspection by pixel comparison, and there is a limit to improving inspection performance.

[0003] On the other hand, in Patent Document 1, a probability distribution of luminance values that a normal circuit pattern can take is predicted from a reference image such as design data, and a region of an inspection image having a luminance value deviating from this probability distribution of luminance values is determined as a defect. A technique is disclosed.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] By distribution comparison, inspection robust against minute manufacturing errors and noise becomes possible, but in Patent Document 1, design data is required for estimation of this distribution. For this reason, when design data cannot be used at the time of inspection, the method of Patent Document 1 cannot be applied.

[0006] One inspection method that does not use design data is die-to-die inspection, which detects defects by comparing the inspection image with a reference image, which is an image of the same shape but at a different location. However, like die-to-database inspection, die-to-die inspection is susceptible to noise due to pixel comparison, and there are limits to improving inspection performance. Semiconductor devices containing undetected defects may become defective products in other inspections performed during final testing, potentially reducing yield.

[0007] The present invention aims to provide a defect inspection device that is robust against noise and manufacturing errors, without using design data during inspection. [Means for solving the problem]

[0008] An example of a defect inspection device according to the present invention is: A feature calculation unit calculates feature quantities based on the captured image of the sample, An image information reduction unit that generates latent variables by reducing the amount of information in the aforementioned feature quantities, A statistics estimation unit estimates the image statistics that a normal image can take based on the aforementioned latent variables, A defect detection unit detects defects in the inspection image based on the aforementioned image statistics and the inspection image of the sample. It is equipped with. [Effects of the Invention]

[0009] The defect inspection device according to the present invention can perform robust inspections against noise and manufacturing errors without using design data during inspection. [Brief explanation of the drawing]

[0010] [Figure 1] A diagram showing the configuration of a defect inspection apparatus according to Embodiment 1 of the present invention. [Figure 2] Configuration diagram of the image information reduction unit according to Example 1. [Figure 3] Configuration diagram for model parameter learning according to Example 1. [Figure 4]A diagram showing the processing flow of model parameter learning according to Example 1. [Figure 5] A schematic diagram showing the images generated at each step of model parameter learning according to Example 1. [Figure 6] A block diagram of the defect inspection according to Example 2 of the present invention. [Figure 7] A schematic diagram showing the estimated luminance value distribution according to Example 2. [Figure 8] A block diagram of the model parameter learning according to Example 3 of the present invention. [Figure 9] A block diagram of the image information reduction unit for pseudo defect creation according to Example 3. [Figure 10] A diagram showing the processing flow of model parameter learning according to Example 3. [Figure 11] A schematic diagram showing the images generated at each step of model parameter learning according to Example 3. [Figure 12] A block diagram of the model parameter learning according to Example 4 of the present invention. [Figure 13] A block diagram of the defect inspection apparatus according to Example 5 of the present invention.

Mode for Carrying Out the Invention

[0011] Hereinafter, embodiments of the present invention will be described based on the accompanying drawings. [Example 1] The defect inspection apparatus according to Example 1 of the present invention will be described with reference to FIG. 1. FIG. 1 is a block diagram of the defect inspection apparatus according to Example 1. The defect inspection apparatus can be configured using, for example, a known computer and includes arithmetic means and storage means. The arithmetic means includes, for example, a processor, and the storage means includes storage media such as a semiconductor memory device and a magnetic disk device. A part or all of the storage media may be a non-transitory storage medium.

[0012] The defect inspection device may also include input / output means. The input / output means includes, for example, input devices such as a keyboard and a mouse, output devices such as a display and a printer, and communication devices such as a network interface.

[0013] The storage means may store a program. By executing this program, the computer may perform the functions described in this embodiment. That is, this program may cause the computer to function as a defect inspection device.

[0014] The storage means of the defect inspection device stores the reference image 1, the inspection image 2, and the detection result 15. The reference image 1 and the inspection image 2 are captured images of specific points of the sample, respectively. In this embodiment, it is preferable that the reference image 1 and the inspection image 2 are images captured at different points having the same shape or a similar shape.

[0015] By the cooperation of the calculation means and the storage means of the defect inspection device, the feature amount calculation unit 11, the image information reduction unit 12, the statistical amount estimation unit 13, and the defect detection unit 14 are configured. The feature amount calculation unit 11, the image information reduction unit 12, and the statistical amount estimation unit 13 are trained models that have learned parameters by machine learning. The parameters are stored as model parameters 10 in the storage means of the defect inspection device.

[0016] The feature amount calculation unit 11 calculates a feature amount based on the reference image 1. The feature amount calculation unit 11 can use, for example, an encoder in a Convolutional Neural Network (CNN). The feature amount is an amount including a plurality of components, and the specific form can be arbitrarily designed. In this embodiment, the feature amount is a third-order tensor having a size of a×b×c. Each component of the feature amount is, for example, a real number between 0 and 1.

[0017] The image information reduction unit 12 reduces the amount of information in the features calculated by the feature calculation unit 11. The configuration of the image information reduction unit 12 is shown in Figure 2. The image information reduction unit 12 includes a discrete representation calculation unit 21 and a latent representation calculation unit 22, and stores a plurality of feature vectors 23. Each feature vector 23 is assigned an index to identify it. The index is represented by a discrete representation (for example, an integer).

[0018] Feature vector 23 is a quantity containing multiple components (however, the number of components in the feature vector does not exceed the number of components in the feature quantity). In this example, the feature vector is an a-dimensional vector, i.e., it has a number of components. Each component of the feature vector is, for example, a real number between 0 and 1.

[0019] The discrete representation calculation unit 21 searches the feature vector 23 for similar values ​​for at least some of the input features and replaces the features with the index values ​​of those feature vectors.

[0020] In this embodiment, a feature quantity having size a × b × c is divided into b × c a-dimensional vector parts. For each vector part, one feature vector 23 (for example, the most similar feature vector) is selected from among several candidate feature vectors 23 based on the values ​​of the components of that vector part. Then, that vector part is replaced with the index of the selected feature vector 23.

[0021] In this specification, a discretized image is obtained by replacing the features with index values ​​in this manner. A discretized image can be represented as a quantity consisting of b × c indices.

[0022] The latent representation calculation unit 22 embeds feature vectors 23 corresponding to each index into the discretized image. This discretized image with the embedded feature vectors 23 will hereafter be referred to as a "latent variable" in this specification. The latent variable becomes a tensor with the same size as the original feature.

[0023] In this way, the image information reduction unit 12 generates latent variables by reducing the information content of the features. In particular, the image information reduction unit 12 reduces the information content of the features by replacing at least a portion of the features with an index (discrete representation) related to at least one of the multiple feature vectors 23. In this way, the process for reducing the amount of information becomes concretely achievable. Note that the specific process for reducing the amount of information is not limited to the example described above and can be designed as appropriate by those skilled in the art.

[0024] The statistics estimation unit 13 estimates the image statistics that a normal image can take based on latent variables. The statistics estimation unit 13 can be configured, for example, using a decoder in a CNN. In this embodiment, the image statistics are represented by the probability distribution of luminance values ​​for each pixel. The probability distribution of luminance values ​​can be represented, for example, by a Gaussian distribution, and one Gaussian distribution is defined for each pixel. The Gaussian distribution can be identified, for example, by the mean and standard deviation of that Gaussian distribution. Hereinafter, in this specification, this probability distribution of luminance values ​​may be referred to as the "luminance value distribution".

[0025] Here, the feature calculation unit 11, the image information reduction unit 12, and the statistical estimation unit 13 each perform the above processing using the model parameters 10. If the feature calculation unit 11 and the statistical estimation unit 13 are the encoder and decoder in a CNN, respectively, the model parameters 10 input to the feature calculation unit 11 and the statistical estimation unit 13 include the weights of the neural network. The model parameters 10 input to the image information reduction unit 12 include the indices and values ​​of the feature vectors 23.

[0026] The defect detection unit 14 detects defects in the inspection image 2 based on the image statistics estimated by the statistics estimation unit 13 and the inspection image 2. For example, for each pixel in the inspection image 2, it determines whether the brightness value of that pixel is within a predetermined normal range related to the brightness value distribution estimated by the statistics estimation unit 13. Then, it detects a region in the inspection image 2 consisting of pixels whose brightness values ​​are outside the normal range as a defect and outputs the detection result 15 to the user.

[0027] The calculation method for the model parameter 10 described above will be explained using the learning flow in Figure 4, along with the configuration diagram in Figure 3.

[0028] In Figure 3, the training image 3 is an image (for example, an SEM image) taken at any location on the wafer, and in this embodiment, the image may or may not contain defects in the sample. For this training image 3, the feature calculation unit 11 performs feature calculation (S101), the image information reduction unit 12 performs index replacement of the feature vector 23 (S102) and embedding of the feature vector (S103), and the statistic estimation unit 13 estimates the brightness value distribution (S104), as described above in relation to Figure 1.

[0029] The model evaluation unit 31 calculates the error between the estimated luminance value distribution and the training image 3 (S105). That is, the model evaluation unit 31 evaluates the error based on the image statistics estimated based on the training image 3 and the original training image 3. For example, for each pixel, a predetermined comparison value (for example, the mean value of a Gaussian distribution) can be calculated from the estimated image statistics, and the difference between this comparison value and the luminance value of the corresponding pixel in the training image 3 can be calculated as the error for each pixel. The sum of the absolute values ​​of the errors for each pixel, or the sum of the squares of the errors for each pixel, can then be used as the error between the estimated luminance value distribution and the training image 3.

[0030] The model parameter update amount calculation unit 32 calculates the amount of model parameter updates that reduce the obtained error (S106). That is, the model parameter update amount calculation unit 32 calculates the amount of model parameter updates for the feature calculation unit 11, the image information reduction unit 12, and the statistical estimation unit 13 based on the obtained error. For example, the stochastic gradient descent method can be used to calculate this update amount. These processes from S101 to S106 are repeated until a predetermined number of times is reached (S107).

[0031] Figure 5 shows a schematic representation of the images generated at each step of this learning process. In training image 301, the semiconductor circuit portion is shown in white. Training image 301 contains circuit distortion due to manufacturing errors and noise, which causes a decrease in defect detection performance.

[0032] From this training image 301, a discretized image 302 (conceptually illustrated) with reduced information content is generated by calculating features (S101) and replacing the indices of the feature vector 23 (S102). From this discretized image 302, the image statistics 303 are estimated by embedding feature vectors (S103) and estimating image statistics (S104). Note that the image statistics 303 in Figure 5 show the average value of the estimated brightness value distribution of each pixel.

[0033] For each training image 301, by training to minimize the difference between the image statistics 303 and the training image 301, it is possible to search for model parameters 10 that minimize the error, and to estimate a plausible brightness value distribution corresponding to the circuit pattern contained in the training image.

[0034] The effects of this embodiment will now be explained. This embodiment describes a method for estimating the luminance value distribution from latent variables by replacing feature quantities in an image with feature vectors. In this replacement, multiple candidate feature vectors are prepared, and a portion of the feature quantity is replaced by one feature vector selected from among these candidates (in this embodiment, this replacement process is repeated b × c times). By adjusting the number of candidate feature vectors prepared, the amount of information in the latent variables used to estimate the luminance value distribution can be controlled.

[0035] For example, by setting a small number of candidate feature vectors, the image information reduction unit 12 can remove information related to noise and manufacturing errors in the image from the feature quantities, as shown in Figure 5, and restore only the circuit pattern features common to the training images. This makes it possible to estimate, pixel by pixel, the probability distribution of brightness values ​​that a normal circuit image, with manufacturing errors and noise information removed from the reference image, can take. By determining areas of the inspection image with brightness values ​​that deviate from this probability distribution as defects, it becomes possible to perform defect inspection that is robust to noise and manufacturing errors. The number of feature vectors may be specified externally, or it may be automatically searched by optimization calculations to maximize inspection performance.

[0036] Thus, the defect inspection device according to this embodiment can perform robust inspections against noise and manufacturing errors without using design data during inspection.

[0037] [Example 2] The defect inspection apparatus according to Embodiment 2 of the present invention will be described below with reference to Figure 6. In the following, parts that are common to Embodiment 1 may be omitted from the explanation.

[0038] The difference between Example 2 and Example 1 is that while Example 1 used reference image 1, Example 2 does not use reference image 1. Instead, defects in the original inspection image 2 are detected by comparing it with the brightness value distribution estimated from the inspection image 2 itself via the feature calculation unit 11, image information reduction unit 12, and statistical estimation unit 13. In other words, in Example 2, the image captured for which feature quantities are calculated is the inspection image 2.

[0039] In Example 2, the statistical estimation unit 13 uses latent variables calculated by the image information reduction unit 12 to estimate a brightness value distribution in the inspection image 2 that restores the areas with defects in the sample to appear as normal areas. By comparing this generated brightness value distribution with the inspection image, defects are inspected using only the inspection image as input.

[0040] The following describes an example of how the statistical estimation unit 13 learns to estimate the brightness value distribution by restoring defective regions to resemble normal regions. When the feature calculation unit 11 and the statistical estimation unit 13 are encoders and decoders in a CNN, respectively, the receptive field can be widened by designing a large number of layers in the CNN, allowing the use of surrounding features in estimating the brightness value distribution. Such a configuration is effective for circuit patterns where the same pattern is repeated.

[0041] The effects of this embodiment will now be explained. In this embodiment, since the inspection is performed by self-referencing of inspection image 2, a reference image is not required, thus reducing the time required to acquire the reference image and improving the inspection throughput.

[0042] Furthermore, self-referencing of inspection image 2 allows for defect inspection without being affected by fluctuations in image quality information (such as distortion, brightness unevenness, and defocus in the image), which had been a cause of performance degradation in conventional die-to-die and die-to-database inspections. In other words, the image information reduction unit 12 does not remove information related to distortion, brightness unevenness, or defocus in the image from the feature quantities.

[0043] Figure 7 shows the inspection image and a schematic representation of the luminance value distribution, which is an image statistic. If there is a luminance unevenness in the upper right of inspection image 401 (an area shown by a dashed line), the luminance value distribution estimated from inspection image 401 via the feature calculation unit 11, image information reduction unit 12, and statistic estimation unit 13 (an example of the average value is shown as image statistic 402) can retain features related to image quality variations while excluding information on noise and manufacturing errors. By comparing this luminance value distribution with the inspection image, defects can be inspected without being affected by image quality variations in the inspection image.

[0044] [Example 3] A defect inspection apparatus according to Embodiment 3 of the present invention will be described with reference to Figure 8. In the following, parts common to Embodiment 1 may be omitted from the explanation.

[0045] The difference between Example 3 and Example 1 is that in Example 1 (Figure 3), the SEM images used for training (training image 3) are not limited to normal images, whereas in Example 3, they are limited to normal images 4 that do not contain defects in the sample. Also, in Example 3, the image information reduction unit 12 (Figure 2) in the model parameter learning process is replaced by an image information reduction unit 41 for creating pseudo-defects.

[0046] Figure 9 shows the configuration of the image information reduction unit 41 for creating pseudo-defects. In addition to the internal configuration of the image information reduction unit 12, the image information reduction unit 41 is newly equipped with an image information manipulation unit 42.

[0047] In conjunction with the configuration diagrams in Figures 8 and 9, we will explain how to calculate the model parameters using the learning flow in Figure 10.

[0048] Using the normal image 4 as a training image, the feature calculation unit 11 calculates features (S101). Then, based on the features related to the normal image 4, the discrete representation calculation unit 21 selects an index (first discrete representation) related to one feature vector 23 from among multiple indexes (discrete representations) of feature vectors 23, and generates a discretized image (S102).

[0049] The index values ​​of this discretized image are changed by the image information manipulation unit 42 (S201). Methods for this change include randomly assigning a certain index value to the discretized image, filling random regions with a specific index, or copying a certain region to another region.

[0050] In this way, the pseudo-defect creation image information reduction unit 41 reduces the amount of information in the features of the normal image 4 by replacing at least a portion of the features of the normal image 4 with an index (second discrete representation) different from the selected index (first discrete representation).

[0051] Feature vectors are embedded according to the index values ​​(S103), and the brightness value distribution is estimated using the discretized image with manipulated index values, similar to Figures 3 and 4 (S104). The error between the estimated brightness value distribution and the normal image 4 is calculated by the model evaluation unit 31 (S105). The model parameter update amount calculation unit 32 calculates the amount of model parameter update that reduces the obtained error (S106).

[0052] Figure 11 shows a schematic representation of the images generated at each step of this learning process. Features are calculated from the normal image 501 (S101), and by replacing the indices of the feature vector 23 (S102), a discretized image 502 (conceptually illustrated) with reduced information content is generated. This discretized image 502 is manipulated by the image information manipulation unit 42.

[0053] The processed discretized image 503 (conceptually illustrated) in Figure 11 is an example using a method that randomly assigns values ​​of certain indices to the discretized image. If these assigned indices are features of the semiconductor circuit pattern, the brightness value distribution estimated from this processed discretized image 503 becomes a pseudo-defect image in which the circuit pattern features are amplified, as shown in the image statistic 504. By learning to minimize the difference between this pseudo-defect image and the normal image 501, it is possible to search for model parameters 10 that minimize the error.

[0054] The effects of this embodiment will now be explained. By manipulating the discretized image with the image information manipulation unit 42, a pseudo-defect image that reproduces some kind of defect is generated. A model parameter 10 that minimizes the difference between this pseudo-defect image and the normal image is searched for, and by using this model parameter 10 in an inspection configuration as shown in Figure 1 or Figure 6, the statistical estimation unit 13 can estimate the brightness value distribution that a normal circuit image can take with higher accuracy from the inspection image containing the defect. By detecting defects by comparing this brightness value distribution with the inspection image 2, a reference image is not required, and the inspection throughput can be improved by reducing the time required to acquire the reference image.

[0055] [Example 4] A defect inspection apparatus according to Embodiment 4 of the present invention will be described with reference to Figure 12. In the following, parts common to Embodiment 1 may be omitted from the explanation.

[0056] The difference between Example 4 and Example 1 is that in Example 1, as shown in Figure 3, the same training image 3 was used for estimating statistics and evaluating the model, whereas in Example 4, as shown in Figure 12, statistics were estimated from defective images 5 that included defects in the sample, and these statistics were compared with normal images 4 that did not contain defects in the sample.

[0057] In other words, in both Example 1 and Example 4, the model evaluation unit 31 evaluates the error based on the image statistics estimated based on the first training image, which is an image of the sample, and the second training image, which is an image of the sample. However, in Example 1, the first training image and the second training image are the same image (i.e., training image 3), whereas in Example 4, the first training image is a defective image 5 containing defects in the sample, and the second training image is a normal image 4 of a normal sample.

[0058] By learning to minimize the difference between the luminance value distribution estimated from the defective image 5 via the feature calculation unit 11, image information reduction unit 12, and statistical estimation unit 13 and the normal image 4, a model parameter 10 that minimizes the error is searched for. By using this model parameter 10 in an inspection configuration as shown in Figure 1 or Figure 6, it becomes possible to estimate the luminance value distribution that the captured image of a normal circuit can take, based on the inspection image containing the defect, using the statistical estimation unit 13, and then perform the inspection.

[0059] In particular, if a pair of defective images and normal images taken at another location with the same pattern is available, the model parameter 10 search using the configuration of this embodiment (Figure 12) is preferable. However, if such a pair is difficult to obtain, the model parameter 10 search using the configuration of Embodiment 1 (Figure 3) may be preferable.

[0060] [Example 5] A defect inspection apparatus according to Embodiment 5 of the present invention will be described with reference to Figure 13. In the following, parts common to Embodiment 1 may be omitted from the explanation.

[0061] The difference between Example 5 and Example 1 is that in Example 5, the defect inspection device is newly equipped with a pattern position adjustment unit 51. The pattern position adjustment unit 51 adjusts the circuit position in the brightness value distribution estimated by the statistics estimation unit 13 to match the inspection image by aligning the image statistics related to the reference image 1 with the reference image 1. This adjustment method may be a general image processing pattern matching method, or deep learning may be used to pre-learn the amount of deviation according to the circuit pattern. Defects are inspected by comparing the brightness value distribution after position correction with the inspection image 2.

[0062] The effects of this embodiment will now be explained. When the image information reduction unit 12 reduces information by discretizing the feature quantities of the reference image 1, some information regarding the position of the semiconductor circuit may be lost. The pattern position adjustment unit 51 corrects the positional shift in the brightness value distribution caused by this loss of positional information, thereby improving inspection performance.

[0063] [Other examples] Although the above embodiments were described using semiconductor defect inspection as an example, the defect inspection apparatus according to each embodiment can also be applied to image-based defect inspection and foreign object detection in other industrial equipment. [Explanation of symbols]

[0064] 1…Reference image (imaging image) 2… Examination images (imaging images) 3…Training images (1st training image, 2nd training image) 4…Normal images (1st training image, 2nd training image) 5…Defective image (first training image) 10…Model parameters 11…Feature calculation unit 12…Image Information Reduction Unit 13…Statistics estimation part 14…Defect detection unit 15…Detection results 21...Discrete representation calculation section 22...Latent expression calculation section 23…Feature vectors 31…Model Evaluation Department 32...Model parameter update amount calculation unit 41...Image information reduction unit for creating pseudo-defects 42…Image Information Manipulation Unit 51...Pattern position matching section 301...Training image 302... Discretized image 303…Image statistics 401... Examination image 402…Image statistics 501...Normal image 502... Discretized image 503...Discretized image after operation 504…Image statistics

Claims

1. A feature calculation unit calculates feature quantities based on the captured image of the sample, An image information reduction unit that generates latent variables by reducing the amount of information in the aforementioned feature quantities, A statistics estimation unit estimates the image statistics that a normal image can take based on the aforementioned latent variables, A defect detection unit detects defects in the inspection image based on the aforementioned image statistics and the inspection image of the sample. Equipped with, The captured image is the examination image. A defect inspection device characterized by the following features.

2. A feature calculation unit calculates feature quantities based on the captured image of the sample, An image information reduction unit that generates latent variables by reducing the amount of information in the aforementioned feature quantities, A statistics estimation unit estimates the image statistics that a normal image can take based on the aforementioned latent variables, A defect detection unit detects defects in the inspection image based on the aforementioned image statistics and the inspection image of the sample. A defect inspection device comprising: The feature calculation unit, the image information reduction unit, and the statistical estimation unit are trained models whose parameters have been learned by machine learning. The aforementioned defect inspection device, A model evaluation unit that evaluates the error based on the image statistics estimated based on the first training image, which is an image of the sample, and the second training image, which is an image of the sample. A model parameter update amount calculation unit calculates the amount of update to the model parameters of the feature calculation unit, the image information reduction unit, and the statistical estimation unit based on the aforementioned error, Equipped with A defect inspection device characterized by the following features.

3. A feature calculation unit calculates feature quantities based on the captured image of the sample, An image information reduction unit that generates latent variables by reducing the amount of information in the aforementioned feature quantities, A statistics estimation unit estimates the image statistics that a normal image can take based on the aforementioned latent variables, A defect detection unit detects defects in the inspection image based on the aforementioned image statistics and the inspection image of the sample. Regarding the image statistics related to the captured image, a pattern position adjustment unit is provided to align them with the captured image, A defect inspection device characterized by comprising the following features.

4. A feature calculation unit calculates feature quantities based on the captured image of the sample, An image information reduction unit that generates latent variables by reducing the amount of information in the aforementioned feature quantities, A statistics estimation unit estimates the image statistics that a normal image can take based on the aforementioned latent variables, A defect detection unit detects defects in the inspection image based on the aforementioned image statistics and the inspection image of the sample. Equipped with, The image information reduction unit removes information related to noise and manufacturing errors in the image from the feature quantities. A defect inspection device characterized by the following features.

5. A feature calculation unit calculates feature quantities based on the captured image of the sample, An image information reduction unit that generates latent variables by reducing the amount of information in the aforementioned feature quantities, A statistics estimation unit estimates the image statistics that a normal image can take based on the aforementioned latent variables, A defect detection unit detects defects in the inspection image based on the aforementioned image statistics and the inspection image of the sample. Equipped with, The image information reduction unit does not remove information related to distortion, brightness unevenness, or defocus in the image from the feature quantities. A defect inspection device characterized by the following features.

6. The defect inspection apparatus according to any one of claims 1 to 5, characterized in that the image information reduction unit reduces the amount of information of the feature quantity by replacing at least a portion of the feature quantity with at least one discrete representation from a plurality of discrete representations.

7. The aforementioned image information reduction unit, Based on the features related to the first training image, a first discrete representation is selected from among multiple discrete representations. The defect inspection apparatus according to claim 2, characterized in that the amount of information of the features relating to the first learning image is reduced by replacing at least a portion of the features relating to the first learning image with a second discrete representation different from the first discrete representation.

8. The defect inspection apparatus according to claim 2, characterized in that the first learning image and the second learning image are identical images.

9. The defect inspection apparatus according to claim 2, characterized in that the first learning image is an image of a sample containing defects, and the second learning image is an image of a normal sample.

Images