Defect detection system
A computing system generates a training data set for training the prediction model to detect defects present in a target surface of a target specimen and training the prediction model to detect defects present in the target surface of the target specimen based on the training data set. The computing system generates the training data set by identifying a set of images for training the prediction model, the set of images comprising a first subset of images. A deep learning network generates a second subset of images for subsequent labelling based on the set of images comprising the first subset of images. The deep learning network generates a third subset of images for labelling based on the set of images comprising the first subset of images and the labeled second subset of images. The computing system continues the process until a threshold number of labeled images is generated.
1. A computing system comprising:
a processor; and
a memory having programming instructions stored thereon, which, when executed by the processor, performs operations comprising:
generating a training data set for training a prediction model to detect defects present in a target surface of a target specimen by:
identifying a set of images for training the prediction model, the set of images comprising a first subset of images, wherein each image of the first subset of images is labeled with labels identifying defects on a respective specimen;
generating, by a deep learning network, a second subset of images for subsequent labelling based on the set of images comprising the first subset of images, the generating comprising:
for each image in the set of images, generating a first probability map corresponding to a first probability of each pixel in the image belongs to an image foreground first category,
for each image in the set of images, generating a second probability map corresponding to a second probability of each pixel in the image belongs to an image background,
for each image in the set of images, generating an unseen metric based on the first probability map and the second probability map, wherein the unseen metric corresponds to a likelihood of the image including pixels unseen by the deep learning network, and
ranking each image in the set of images based on the unseen metric, wherein the second subset of images is generated using images that exceed a threshold unseen metric value;
prompting an operator to label each image in the second subset of images; and
aggregating the first subset of images and the second subset of images, wherein each image in the first subset of images and the second subset of images are labeled; and
training the prediction model to detect defects present in the target surface of the target specimen based on the training data set.
2. The computing system of claim 1 , wherein the unseen metric for the image is based on an aggregate of unseen metrics for each pixel in the image.
3. The computing system of claim 2 , wherein generating, by the deep learning network, the second subset of images for subsequent labelling based on the set of images comprising the first subset of images comprises:
generating an alpha metric for the image based a corresponding unseen metric.
4. The computing system of claim 2 , wherein generating, by the deep learning network, the second subset of images for subsequent labelling based on the set of images comprising the first subset of images comprises:
generating a threshold metric for the image based on a number of pixels in the image having an unseen score that exceeds a threshold value.
5. The computing system of claim 1 , further comprising:
receiving, from an imaging apparatus, a target image of the target surface of the target specimen;
detecting, by the prediction model, one or more defects present in the target surface of the target specimen; and
based on the detecting, generating a graphical output illustrating the one or more defects.
6. A method comprising:
generating, by a computing system, a training data set for training a prediction model to detect defects present in a target surface of a target specimen by:
identifying a set of images for training the prediction model, the set of images comprising a first subset of images, wherein each image of the first subset of images is labeled with labels identifying defects on a respective specimen;
generating, by a deep learning network, a second subset of images for subsequent labelling based on the set of images comprising the first subset of images, the generating comprising:
for each image in the set of images, generating a first probability map corresponding to a first probability of each pixel in the image belongs to an image foreground first category,
for each image in the set of images, generating a second probability map corresponding to a second probability of each pixel in the image belongs to an image background,
for each image in the set of images, generating an unseen metric based on the first probability map and the second probability map, wherein the unseen metric corresponds to a likelihood of the image including pixels unseen by the deep learning network, and
ranking each image in the set of images based on the unseen metric, wherein the second subset of images is generated using images that exceed a threshold unseen metric value;
prompting an operator to label each image in the second subset of images; and
aggregating the first subset of images and the second subset of images, wherein each image in the first subset of images and the second subset of images are labeled; and
training, by the computing system, the prediction model to detect defects present in the target surface of the target specimen based on the training data set.
7. The method of claim 6 , wherein the unseen metric for the image is based on an aggregate of unseen metrics for each pixel in the image.
8. The method of claim 7 , wherein generating, by the deep learning network, the second subset of images for subsequent labelling based on the set of images comprising the first subset of images comprises:
generating an alpha metric for the image based a corresponding unseen metric.
9. The method of claim 7 , wherein generating, by the deep learning network, the second subset of images for subsequent labelling based on the set of images comprising the first subset of images comprises:
generating a threshold metric for the image based on a number of pixels in the image having an unseen score that exceeds a threshold value.
10. The method of claim 6 , further comprising:
receiving, by the computing system from an imaging apparatus, a target image of the target surface of the target specimen;
detecting, by the prediction model, one or more defects present in the target surface of the target specimen; and
based on the detecting, generating, by the computing system, a graphical output illustrating the one or more defects.
11. A non-transitory computer readable medium having one or more sequences of instructions, which, when executed by one or more processors, causes a computing system to perform operations comprising:
generating, by the computing system, a training data set for training a prediction model to detect defects present in a target surface of a target specimen by:
identifying a set of images for training the prediction model, the set of images comprising a first subset of images, wherein each image of the first subset of images is labeled with labels identifying defects on a respective specimen;
generating, by a deep learning network, a second subset of images for subsequent labelling based on the set of images comprising the first subset of images, the generating comprising:
for each image in the set of images, generating a first probability map corresponding to a first probability of each pixel in the image belongs to an image foreground first category,
for each image in the set of images, generating a second probability map corresponding to a second probability of each pixel in the image belongs to an image background,
for each image in the set of images, generating an unseen metric based on the first probability map and the second probability map, wherein the unseen metric corresponds to a likelihood of the image including pixels unseen by the deep learning network, and
ranking each image in the set of images based on the unseen metric, wherein the second subset of images is generated using images that exceed a threshold unseen metric value;
prompting an operator to label each image in the second subset of images; and
aggregating the first subset of images and the second subset of images, wherein each image in the first subset of images and the second subset of images are labeled; and
training, by the computing system, the prediction model to detect defects present in the target surface of the target specimen based on the training data set.
12. The non-transitory computer readable medium of claim 11 , wherein the unseen metric for the image is based on an aggregate of unseen metrics for each pixel in the image.
13. The non-transitory computer readable medium of claim 12 , wherein generating, by the deep learning network, the second subset of images for subsequent labelling based on the set of images comprising the first subset of images comprises:
generating, by the computing system, a threshold metric for the image based on a number of pixels in the image having an unseen score that exceeds a threshold value.
14. The non-transitory computer readable medium of claim 11 , further comprising:
receiving, by the computing system from an imaging apparatus, a target image of the target surface of the target specimen;
detecting, by the prediction model, one or more defects present in the target surface of the target specimen; and
based on the detecting, generating, by the computing system, a graphical output illustrating the one or more defects.
15. The computing system of claim 1 , wherein the deep learning network is a modified U-net architecture that does not include the final softmax layer in order to generate pixel probabilities per category.
16. The computing system of claim 15 , wherein a convolutional layer padding of the deep learning network matches an input size and a number of feature maps used by convolutional layers of the deep learning network.