Defect detection system
A method for training a defect-detection model for inspecting an inspection object includes generating defective inspection object data and defective material parameter data, generating at least one pair of a defective image and a non-defective image corresponding to the defective image based on inspection object data, the defective inspection object data, the defective material parameter data, and scene parameter data, comparing the defective image with the non-defective image, generating a visibility score of the at least one defect based on the comparison of the defective image with the non-defective image, classifying the visibility score as first, second, and third categories based on the visibility score and, based on the classification of the visibility score of the at least one defect, performing one or more training-data determination actions.
1 . A method for training a defect-detection model for inspecting an inspection object, the method comprising:
generating, by one or more processors, first data indicative of a defective inspection object and second data indicative of a defective material parameter;
generating, by the one or more processors, at least one pair of a defective object image and a non-defective object image based on inspection object data, the first data, the second data, and scene parameter data, wherein the defective object image of the at least one pair includes a first image portion with at least one defect of an inspection object, and a second image portion without the at least one defect of the inspection object,
comparing, by the one or more processors, the defective object image with the non-defective object image,
generating, by the one or more processors, a visibility score of the at least one defect based on the comparison of the defective object image with the non-defective object image,
classifying, by the one or more processors, the visibility score as:
(i) a first category based on the visibility score being equal to or greater than a first visibility threshold value;
(ii) a second category based on the visibility score being equal to or greater than a second visibility threshold value and less than the first visibility threshold value; or
(iii) a third category based on the visibility score being less than the second visibility threshold value; and
based on the classification of the visibility score of the at least one defect, performing, by the one or more processors, one or more training-data determination actions, wherein the one or more training-data determination action comprises using data regarding the at least one defect for training a defect-detection model and excluding the data regarding the at least one defect from training the defect-detection model.
2 . The method of claim 1 , wherein the non-defective object image of the at least one pair includes a third image portion corresponding to the first image portion of the defective object image and without the at least one defect of the inspection object, and a fourth image portion that is a same as the second image portion of the defective object image.
3 . The method of claim 1 , comprising receiving, by the one or more processors, inspection object data, defect data from a defect library, and material data from a material library,
wherein generating the first data and the second data comprises generating the first data and the second data based on the inspection object data, the defect data, and the material data.
4 . The method of claim 1 , wherein comparing the defective object image with the non-defective object image comprises comparing the defective object image and the non-defective object image within a ground-truth defect segmentation mask using a pixel-wise difference metric.
5 . The method of claim 1 , wherein the at least one defect comprises a plurality of defects,
wherein generating the visibility score of the at least one defect comprises generating the visibility score for each of the plurality of defects.
6 . The method of claim 1 , wherein using the data regarding the at least one defect for training the defect-detection model comprises using the data regarding the at least one defect for training the defect-detection model based on the visibility score being classified into the first category.
7 . The method of claim 1 , wherein excluding the data regarding the at least one defect from training the defect-detection model comprises excluding the data regarding the at least one defect from training the defect-detection model based on the visibility score being classified into the third category.
8 . The method of claim 1 , wherein the one or more training-data determination actions comprise masking the at least one defect from loss computation for training the defect detection model based on the visibility score being classified into the second category.
9 . A system for training a defect-detection model for inspecting an inspection object, the system comprising:
memory;
one or more processors in communication with the memory, wherein the one or more processors are configured to:
generate first data indicative of a defective inspection object and second data indicative of a defective material parameter;
generate at least one pair of a defective object image and a non-defective object image based on inspection object data, the first data, the second data, and scene parameter data, wherein the defective object image of the at least one pair includes a first image portion with at least one defect of an inspection object, and a second image portion without the at least one defect of the inspection object,
compare the defective object image with the non-defective object image,
generate a visibility score of the at least one defect based on the comparison of the defective object image with the non-defective object image,
classify the visibility score as:
(i) a first category based on the visibility score being equal to or greater than a first visibility threshold value;
(ii) a second category based on the visibility score being equal to or greater than a second visibility threshold value and less than the first visibility threshold value; or
(iii) a third category based on the visibility score being less than the second visibility threshold value; and
based on the classification of the visibility score of the at least one defect, perform one or more training-data determination actions, wherein the one or more training-data determination action comprises using data regarding the at least one defect for training a defect-detection model and excluding the data regarding the at least one defect from training the defect-detection model.
10 . The system of claim 9 , wherein the non-defective object image of the at least one pair includes a third image portion corresponding to the first image portion of the defective object image and without the at least one defect of the inspection object, and a fourth image portion that is a same as the second image portion of the defective object image.
11 . The system of claim 9 , wherein the one or more processors are configured to receive inspection object data, defect data from a defect library, and material data from a material library,
wherein generating the first data and the second data comprises generating the first data and the second data based on the inspection object data, the defect data, and the material data.
12 . The system of claim 9 , wherein comparing the defective object image with the non-defective object image comprises comparing the defective object image and the non-defective object image within a ground-truth defect segmentation mask using a pixel-wise difference metric.
13 . The system of claim 9 , wherein the at least one defect comprises a plurality of defects,
wherein generating the visibility score of the at least one defect comprises generating the visibility score for each of the plurality of defects.
14 . The system of claim 9 , wherein using the data regarding the at least one defect for training the defect-detection model comprises using the data regarding the at least one defect for training the defect-detection model based on the visibility score being classified into the first category.
15 . The system of claim 9 , wherein excluding the data regarding the at least one defect from training the defect-detection model comprises excluding the data regarding the at least one defect from training the defect-detection model based on the visibility score being classified into the third category.
16 . The system of claim 9 , wherein the one or more training-data determination actions comprise masking the at least one defect from loss computation for training the defect detection model based on the visibility score being classified into the second category.
17 . One or more non-transitory computer-readable storage media including instructions that, when executed by one or more processors, cause the one or more processors to:
generate first data indicative of a defective inspection object and second data indicative of a defective material parameter;
generate at least one pair of a defective object image and a non-defective object image based on inspection object data, the first data, the second data, and scene parameter data, wherein the defective object image of the at least one pair includes a first image portion with at least one defect of an inspection object, and a second image portion without the at least one defect of the inspection object,
compare the defective object image with the non-defective object image,
generate a visibility score of the at least one defect based on the comparison of the defective object image with the non-defective object image,
classify the visibility score as:
(i) a first category based on the visibility score being equal to or greater than a first visibility threshold value;
(ii) a second category based on the visibility score being equal to or greater than a second visibility threshold value and less than the first visibility threshold value; or
(iii) a third category based on the visibility score being less than the second visibility threshold value; and
based on the classification of the visibility score of the at least one defect, perform one or more training-data determination actions, wherein the one or more training-data determination action comprises using data regarding the at least one defect for training a defect-detection model and excluding the data regarding the at least one defect from training the defect-detection model.
18 . The one or more non-transitory computer-readable storage media of claim 17 , wherein the non-defective object image of the at least one pair includes a third image portion corresponding to the first image portion of the defective object image and without the at least one defect of the inspection object, and a fourth image portion that is a same as the second image portion of the defective object image.
19 . The one or more non-transitory computer-readable storage media of claim 17 , wherein comparing the defective object image with the non-defective object image comprises comparing the defective object image and the non-defective object image within a ground-truth defect segmentation mask using a pixel-wise difference metric.
20 . The one or more non-transitory computer-readable storage media of claim 17 , wherein the one or more training-data determination actions comprise masking the at least one defect from loss computation for training the defect detection model based on the visibility score being classified into the second category.