Devices, systems, and methods for anomaly detection
Devices, systems, and methods obtain one or more training images; obtain a test image; select one or more associated pixels in the training images for a target pixel in the training images; calculate respective value relationships between a value of the target pixel and respective values of the associated pixels in the training images; select one or more associated pixels in the test image for a target pixel in the test image; and detect an anomaly in the target pixel in the test image based on the respective value relationships between the value of the target pixel and the respective values of the associated pixels in the training images and on respective value relationships between a value of the target pixel and respective values of the associated pixels in the test image.
1. A method comprising:
obtaining training images;
obtaining a test image;
selecting one or more respective associated pixels in the training images for each of a plurality of corresponding target pixels in the training images, wherein the plurality of corresponding target pixels correspond across the training images, and wherein the one or more respective associated pixels of each target pixel in the plurality of corresponding target pixels are included in a same training image of the one or more training images as the target pixel;
selecting one or more respective associated pixels in the test image for a target pixel in the test image, wherein each of the one or more respective associated pixels in the test image corresponds to at least one of the one or more respective associated pixels in the training images; and
detecting an anomaly in the target pixel in the test image based on respective values of the plurality of corresponding target pixels and respective values of the respective associated pixels of the plurality of target pixels in the training images and on a value of the target pixel and respective values of the respective associated pixels in the test image,
wherein detecting the anomaly in the target pixel in the test image includes:
generating a predicted target-pixel value based on the respective value of at least one of the plurality of corresponding target pixels and the respective values of the associated pixels of the at least one of the plurality of corresponding target pixels in the training images, and
calculating, for the test image, a target-prediction-to-target-difference between the predicted target-pixel value and the value of the target pixel in the test image.
2. The method of claim 1 , wherein selecting the one or more respective associated pixels in the training images for each of the plurality of corresponding target pixels in the training images includes:
selecting one or more respective candidate associated pixels in the training images for each of the plurality of corresponding target pixels in the training images,
calculating respective value differences between values that are based on the respective values of the plurality of corresponding target pixels and values that are based on respective values of the respective candidate associated pixels of the plurality of corresponding target pixels,
calculating respective variations of the respective value differences, and
selecting the one or more respective associated pixels in the training images for each of the plurality of corresponding target pixels in the training images based on the respective variations of the respective value differences.
3. The method of claim 1 , wherein detecting the anomaly in the target pixel in the test image further includes:
comparing the target-prediction-to-target-difference in the test image to a centrality and a variation of target-prediction-to-target differences across the training images.
4. The method of claim 3 , wherein the centrality of the target-prediction-to-target differences across the training images is a median or a mean.
5. The method of claim 3 , wherein the variation of the target-prediction-to-target differences across the training images is a median-absolute deviation or a standard deviation.
6. The method of claim 1 , wherein at least some of the one or more respective associated pixels of each target pixel in the plurality of corresponding target pixels have different distance ranges from the target pixel.
7. One or more non-transitory computer-readable media storing instructions that, when executed by one or more computing devices, cause the one or more computing devices to perform operations comprising:
obtaining training images;
selecting a plurality of corresponding target pixels in the training images;
selecting, for each corresponding target pixel in the plurality of corresponding target pixels, one or more respective corresponding associated pixels in the training images, wherein each target pixel in the plurality of corresponding target pixels and the one or more respective associated pixels of the target pixel are included in a same training image of the training images, and wherein the corresponding target pixels and the corresponding associated pixels correspond across the training images;
obtaining a test image;
selecting a target pixel in the test image corresponding to the plurality of corresponding target pixels in the training images;
selecting one or more respective associated pixels in the test image for the target pixel in the test image, wherein each of the respective associated pixels in the test image corresponds to a respective one or more of the respective corresponding associated pixels in the training images; and
performing anomaly detection on the target pixel in the test image based on a value of the target pixel and respective values of the respective associated pixels in the test image and on respective values of the plurality of corresponding target pixels and respective values of the respective corresponding associated pixels of the plurality of corresponding target pixels in the training images, wherein performing anomaly detection on the target pixel includes
generating a predicted target-pixel-value model based on the respective values of the plurality of corresponding target pixels and the respective values of the respective corresponding associated pixels of the plurality of corresponding target pixels in the training images,
generating a predicted target-pixel value based on the predicted target-pixel-value model and on the respective values of the respective associated pixels in the test image, and
calculating a respective value difference between the value of the target pixel in the test image and the predicted target-pixel value.
8. The one or more non-transitory computer-readable media of claim 7 , wherein selecting, for each corresponding target pixel in the plurality of corresponding target pixels, the one or more respective corresponding associated pixels in the training images includes:
selecting, for each corresponding target pixel in the plurality of corresponding target pixels, one or more respective candidate associated pixels in the training images;
calculating, for each corresponding target pixel in the plurality of corresponding target pixels, respective value differences between a value of the corresponding target pixel and values of the respective candidate associated pixels of the target pixel;
calculating, for each corresponding target pixel in the plurality of corresponding target pixels, respective variations of the respective value differences; and
selecting, for each corresponding target pixel in the plurality of corresponding target pixels, the one or more respective corresponding associated pixels in the training images based on the respective variations of the respective value differences.
9. The one or more non-transitory computer-readable storage media of claim 7 , wherein at least some of the one or more respective associated pixels of each target pixel in the plurality of corresponding target pixels have different distance ranges from the target pixel.
10. The one or more non-transitory computer-readable storage media of claim 7 , wherein the predicted target-pixel-value model is an affine model with respect to the respective values of the respective associated pixels in the test image.
11. The one or more non-transitory computer-readable media of claim 7 , wherein performing anomaly detection on the target pixel in the test image further comprises:
calculating a centrality and a variation of the respective value difference between the value of the target pixel and the predicted target-pixel value based on the respective values of the plurality of corresponding target pixels and on the respective values of the respective corresponding associated pixels in the training images; and
generating an anomaly score based at least in part on the respective value difference, the centrality, and the variation.
12. The one or more non-transitory computer-readable media of claim 11 , wherein the centrality of the respective value difference between the value of the target pixel and the predicted target-pixel value is a median or a mean.
13. The one or more non-transitory computer-readable media of claim 11 , wherein the variation of the respective value difference between the value of the target pixel and the predicted target-pixel value is a median-absolute deviation or a standard deviation.
14. A device comprising:
one or more computer-readable storage media; and
one or more processors in communication with the one or more computer-readable storage media to cause the device to perform operations comprising:
obtaining training images;
obtaining a test image;
selecting one or more respective associated pixels in the training images for each of a plurality of corresponding target pixels in the training images, wherein the plurality of corresponding target pixels correspond across the training images, and wherein the one or more respective associated pixels of each target pixel in the plurality of corresponding target pixels are included in a same training image of the one or more training images as the target pixel;
selecting one or more respective associated pixels in the test image for a target pixel in the test image, wherein each of the one or more respective associated pixels in the test image corresponds to at least one of the one or more respective associated pixels in the training images; and
detecting an anomaly in the target pixel in the test image based on respective values of the plurality of corresponding target pixels and respective values of the respective associated pixels of the plurality of target pixels in the training images and on a value of the target pixel and respective values of the respective associated pixels in the test image, wherein detecting the anomaly in the target pixel in the test image includes:
generating a predicted target-pixel value based on the respective value of at least one of the plurality of corresponding target pixels and the respective values of the associated pixels of the at least one of the plurality of corresponding target pixels in the training images, and
calculating, for the test image, a target-prediction-to-target-difference between the predicted target-pixel value and the value of the target pixel in the test image.
15. The device of claim 14 , wherein selecting the one or more respective associated pixels in the training images for each of the plurality of corresponding target pixels in the training images includes:
selecting one or more respective candidate associated pixels in the training images for each of the plurality of corresponding target pixels in the training images;
calculating respective value differences between values that are based on the respective values of the plurality of corresponding target pixels and values that are based on respective values of the respective candidate associated pixels of the plurality of corresponding target pixels;
calculating respective variations of the respective value differences; and
selecting the one or more respective associated pixels in the training images for each of the plurality of corresponding target pixels in the training images based on the respective variations of the respective value differences.
16. The device of claim 14 , wherein detecting the anomaly in the target pixel in the test image further includes:
comparing the target-prediction-to-target-difference in the test image to a centrality and a variation of target-prediction-to-target differences across the training images.
17. The device of claim 16 , wherein the centrality of the target-prediction-to-target differences across the training images is a median or a mean.
18. The device of claim 16 , wherein the variation of the target-prediction-to-target differences across the training images is a median-absolute deviation or a standard deviation.
19. The device of claim 14 , wherein at least some of the one or more respective associated pixels of each target pixel in the plurality of corresponding target pixels have different distance ranges from the target pixel.