Defect inspection apparatus
A defect inspection apparatus includes: a feature value calculation unit calculating a feature value based on a captured image of a sample; an image information reduction unit generating a latent variable by reducing an information quantity of the feature value; a statistic value estimation unit estimating an image statistic value that can be taken by a normal image based on the latent variable; and a defect detection unit detecting a defect in an inspection image based on the image statistic value and the inspection image of the sample.
1 . A defect inspection apparatus for improving robustness to process and imaging variations and increasing inspection throughput by generating per-pixel normal-image statistics from a vector-quantized latent representation, the apparatus comprising:
an imaging subsystem configured to capture, from a sample, an inspection image; and
one or more processors and a non-transitory memory storing instructions that, when executed by the one or more processors, cause the apparatus to:
compute, from a captured image of the sample, a feature representation;
generate a latent variable by reducing an information quantity of the feature representation,
estimate, based on the latent variable, an image statistic value that can be taken by a normal image, and
detect a defect in the inspection image based on the image statistic value and the inspection image of the sample.
2 . The defect inspection apparatus according to claim 1 , wherein the captured image is the inspection image.
3 . The defect inspection apparatus according to claim 1 , wherein generating the latent variable comprises reducing the information quantity of the feature representation by replacing at least part of the feature representation with at least one discrete representation selected from a plurality of discrete representations.
4 . The defect inspection apparatus according to claim 3 , wherein the plurality of discrete representations are entries of a learned codebook and the replacing comprises substituting, for respective spatial locations, indices identifying the entries.
5 . The defect inspection apparatus according to claim 1 , wherein the one or more processors are further configured to execute trained models whose parameters are learned by machine learning for the feature-representation computation, for the latent-variable generation, and for the image-statistic estimation, and wherein the one or more processors are further configured to:
evaluate an error based on an image statistic value estimated from a first learning image as a captured image of the sample and from a second learning image as a captured image of the sample, and
calculate parameter update quantities for the trained models based on the error.
6 . The defect inspection apparatus according to claim 5 , wherein
the one or more processors are configured to select, for the first learning image, a first discrete representation from among the plurality of discrete representations based on a feature representation related to the first learning image, and
to reduce an information quantity of the feature representation related to the first learning image by replacing at least part of that feature representation with a second discrete representation different from the first discrete representation.
7 . The defect inspection apparatus according to claim 5 , wherein the first learning image and the second learning image are the same image.
8 . The defect inspection apparatus according to claim 5 , wherein the first learning image is a sample image including a defect and the second learning image is an image of a normal sample.
9 . The defect inspection apparatus according to claim 5 , wherein generating training data further comprises creating a pseudo-defect learning image by operating indices of the latent variable and using the pseudo-defect learning image for model training together with a corresponding normal image.
10 . The defect inspection apparatus according to claim 1 , further comprising one or more processors configured to perform pattern position matching to align the image statistic value related to the captured image with respect to the captured image.
11 . The defect inspection apparatus according to claim 1 , wherein generating the latent variable includes deleting intra-image noise and manufacturing-error-related information from the feature representation.
12 . The defect inspection apparatus according to claim 1 , wherein generating the latent variable does not delete information related to intra-image distortion, luminance unevenness, or defocusing from the feature representation.
13 . The defect inspection apparatus according to claim 1 , wherein the image statistic value comprises, for each pixel, at least one of a central-tendency value and a dispersion value predicted for a normal image.
14 . The defect inspection apparatus according to claim 1 , wherein the apparatus operates without design data of the sample and without using a separate reference image.