Devices, systems, and methods for topological normalization for anomaly detection

Abstract

Some embodiments of devices, systems, and methods obtain at least one first image, wherein the at least one first image is defined in an image space; select at least one feature in the at least one first image; define a topology based on the at least one feature; generate a topology mapping between the topology and the image-space topology; obtain a plurality of anomaly scores, wherein each anomaly score of the plurality of anomaly scores was generated based on a respective detection area in a second image; map the plurality of anomaly scores to the topology based on the topology mapping; and normalize each anomaly score in the plurality of anomaly scores based on the respective neighboring anomaly scores in the topology, thereby generating normalized anomaly scores.

1 . A device comprising:
 one or more computer-readable storage media; and   one or more processors that are in communication with the one or more computer-readable storage media and that are configured to cause the device to perform operations including:
 obtaining a plurality of anomaly scores, wherein each anomaly score of the plurality of anomaly scores corresponds to a respective detection area in an image; 
 generating a respective adjustment factor for each of one or more detection areas, wherein the respective adjustment factor of each detection area of the one or more detection areas is based on the anomaly score of the detection area and on the anomaly scores of the detection areas that are in a respective neighborhood of the detection area; and 
 adjusting at least one anomaly score, which corresponds to a detection area of the one or more detection areas for a which a respective adjustment factor was generated, based on the respective adjustment factor of the corresponding detection area. 
   
     
     
         2 . The device of  claim 1 , wherein generating the respective adjustment factor for each of the one or more detection areas includes:
 identifying one or more outlier anomaly scores in the plurality of anomaly scores; and   replacing the one or more outlier anomaly scores with one or more inlier anomaly scores.   
     
     
         3 . The device of  claim 2 , wherein the operations further include:
 obtaining reference anomaly scores, wherein the reference anomaly scores include a respective plurality of reference anomaly scores for each detection area in the image; and   for each detection area, calculating a respective anomaly-score variation for the detection area based on the respective plurality of reference anomaly scores for the detection area,   wherein identifying the one or more outlier anomaly scores in the plurality of anomaly scores is based on the anomaly-score variations.   
     
     
         4 . The device of  claim 1 , wherein generating the respective adjustment factor for each of the one or more detection areas includes:
 smoothing the plurality of anomaly scores, wherein each anomaly score in the plurality of anomaly scores is smoothed based on the anomaly scores of the detection areas that are in the respective neighborhood of the corresponding detection area of the anomaly score, thereby generating smoothed anomaly scores.   
     
     
         5 . The device of  claim 4 , wherein generating the respective adjustment factor for each of the one or more detection areas includes:
 changing the smoothed anomaly scores that are less than a threshold value to the threshold value.   
     
     
         6 . The device of  claim 5 , wherein adjusting the at least one anomaly score includes dividing the at least one anomaly score by the respective adjustment factor of the corresponding detection area. 
     
     
         7 . A method comprising:
 obtaining a plurality of anomaly scores, wherein each anomaly score of the plurality of anomaly scores corresponds to a respective detection area in an image;   generating respective adjustment factors for at least some of the detection areas, wherein each of the respective adjustment factors is generated based on the respective anomaly score of the corresponding detection area and on the respective anomaly scores of the detection areas that are in a respective neighborhood of the corresponding detection area; and   adjusting at least one anomaly score of the plurality of anomaly scores, the at last one anomaly score corresponding to a detection area of the one or more detection areas for which a respective adjustment factor was generated, based on the respective adjustment factor of the respective detection area that corresponds to the at least one anomaly score.   
     
     
         8 . The method of  claim 7 , wherein generating the respective adjustment factors for at least some of the detection areas includes:
 identifying one or more outlier anomaly scores in the plurality of anomaly scores; and   replacing the one or more outlier anomaly scores with one or more inlier anomaly scores.   
     
     
         9 . The method of  claim 8 , further comprising:
 obtaining reference anomaly scores, wherein the reference anomaly scores include a respective plurality of reference anomaly scores for each detection area in the image; and   for each detection area, calculating a respective anomaly-score variation for the detection area based on the respective plurality of reference anomaly scores for the detection area,   wherein identifying the one or more outlier anomaly scores in the plurality of anomaly scores is based on the anomaly-score variations.   
     
     
         10 . The method of  claim 9 , wherein generating the respective adjustment factors for at least some of the detection areas includes:
 after replacing the one or more outlier anomaly scores with the one or more inlier anomaly scores, smoothing the plurality of anomaly scores, wherein each anomaly score in the plurality of scores is smoothed based on the anomaly scores of the detection areas that are in the respective neighborhood of the corresponding detection area of the anomaly score, thereby generating smoothed anomaly scores.   
     
     
         11 . The method of  claim 10 , wherein generating the respective adjustment factors for at least some of the detection areas includes:
 changing the smoothed anomaly scores that are less than a threshold value to the threshold value.   
     
     
         12 . The method of  claim 11 , wherein the threshold value is one, and
 wherein adjusting the at least one anomaly score includes dividing the at least one anomaly score by the respective adjustment factor.   
     
     
         13 . One or more computer-readable storage media storing instructions that, when executed by one or more computing devices, cause the one or more computing devices to perform operations comprising:
 obtaining a plurality of anomaly scores, wherein each anomaly score of the plurality of anomaly scores corresponds to a respective detection area in an image;   generating respective adjustment factors for at least some of the detection areas, wherein each of the respective adjustment factors is generated based on the respective anomaly score of the corresponding detection area and on the respective anomaly scores of the detection areas that are in a respective neighborhood of the corresponding detection area; and   adjusting at least one anomaly score of the plurality of anomaly scores, the at last one anomaly score corresponding to a detection area of the one or more detection areas for which a respective adjustment factor was generated, based on the respective adjustment factor of the respective detection area that corresponds to the at least one anomaly score.   
     
     
         14 . The one or more computer-readable storage media of  claim 13 , wherein generating the respective adjustment factors for at least some of the detection areas includes:
 identifying one or more outlier anomaly scores in the plurality of anomaly scores; and   replacing the one or more outlier anomaly scores with one or more inlier anomaly scores.   
     
     
         15 . The one or more computer-readable storage media of  claim 14 , wherein the operations further comprise:
 obtaining reference anomaly scores, wherein the reference anomaly scores include a respective plurality of reference anomaly scores for each detection area in the image; and   for each detection area, calculating a respective anomaly-score variation for the detection area based on the respective plurality of reference anomaly scores for the detection area,   wherein identifying the one or more outlier anomaly scores in the plurality of anomaly scores is based on the anomaly-score variations.   
     
     
         16 . The one or more computer-readable storage media of  claim 14 , wherein generating the respective adjustment factors for at least some of the detection areas includes:
 after replacing the one or more outlier anomaly scores with the one or more inlier anomaly scores, smoothing the plurality of anomaly scores, wherein each anomaly score in the plurality of anomaly scores is smoothed based on the anomaly scores of the detection areas that are in the respective neighborhood of the corresponding detection area of the anomaly score, thereby generating smoothed anomaly scores.   
     
     
         17 . The one or more computer-readable storage media of  claim 16 , wherein generating the respective adjustment factors for at least some of the detection areas includes:
 changing the smoothed anomaly scores that are less than a threshold value to the threshold value.   
     
     
         18 . The one or more computer-readable storage media of  claim 17 , wherein the threshold value is one, and
 wherein adjusting the at least one anomaly score includes dividing the at least one anomaly score by the respective adjustment factor.