Low contrast non-referential defect detection
Defect detection techniques can be used for inspecting semiconductor devices, such as CMOS image sensors, during the manufacturing process. The defects can include common defects, such as scratches, dirt, etc., as well as low-contrast defects, such as watermarks. The detection technique may use a supervised machine learning network.
1. A method to detect a defect in a semiconductor device, the method comprising:
generating a plurality of binary slices of an image of the semiconductor device;
segmenting each binary slice of the plurality of binary slices into segments based on statistical contrast properties of the image;
inputting a subset of the segments into a machine learning network, wherein the subset of the segments include segments with a higher contrast profile as compared to other segments not included in the subset; and
receiving an output of the machine learning network identifying presence or absence of at least one defect in the image.
2. The method of claim 1 , wherein the machine learning network includes a multi-layered machine learning model, and wherein each binary slice corresponds to a different grayscale level.
3. The method of claim 1 , wherein the machine learning network is trained to detect defects using a supervised machine learning technique with at least one training image by:
generating a plurality of binary slices of the at least one training image to create training image slices;
segmenting each slice of the training image slices based on statistical contrast properties of the at least one training image to generate training segments;
receiving a defect label for the training segments with a known defect and receiving a no-defect label for one or more segments without the known defect;
inputting the training segments into the machine learning network; and
providing a reference output to the machine learning network based on the labels of the training segments,
wherein the machine learning network performs iterative operations until its output matches the reference output.
4. The method of claim 1 , wherein the defect is a low-contrast watermark defect.
5. The method of claim 1 , further comprising:
transforming the image to enhance at least one feature of the image; and
preprocessing the image, including filtering the image with a series of linear and non-linear filters.
6. The method of claim 1 , wherein the statistical contrast properties are defined by one or more of histogram, ranges, standard deviation, min-max, average, or ratio of dark to light pixels.
7. The method of claim 1 , wherein the method is performed using a graphic processing unit (GPU) and a computer processing unit (CPU).
8. The method of claim 1 , wherein the plurality of binary slices are generated using different thresholds corresponding to different grayscale levels, wherein each binary slice corresponds to a different grayscale level.
9. An inspection system comprising:
a detector to capture an image of a semiconductor device;
one or more processors of a machine; and
a memory storing instructions that, when executed by the one or more processors, cause the machine to perform operations comprising:
generating a plurality of binary slices of the image of the semiconductor device;
segmenting each binary slice of the plurality of binary slices into segments based on statistical contrast properties of the image;
inputting a subset of the segments into a machine learning network, wherein the subset of the segments include segments with a higher contrast profile as compared to other segments not included in the subset; and
receiving an output of the machine learning network identifying presence or absence of at least one defect in the image.
10. The inspection system of claim 9 , wherein the machine learning network includes a multi-layered machine learning model, and wherein each binary slice corresponds to a different grayscale level.
11. The inspection system of claim 9 , wherein the machine learning network is trained to detect defects using a supervised machine learning technique with at least one training image by:
generating a plurality of binary slices of the at least one training image to create training image slices;
segmenting each slice of the training image slices based on statistical contrast properties of the at least one training image to generate training segments;
receiving a defect label for the training segments with a known defect and receiving a no-defect label for one or more segments without the known defect;
inputting the training segments into the machine learning network; and
providing a reference output to the machine learning network based on the labels of the training segments,
wherein the machine learning network performs iterative operations until its output matches the reference output.
12. The inspection system of claim 9 , wherein the defect is a low-contrast watermark defect.
13. The inspection system of claim 9 , the operations further comprising:
transforming the image to enhance at least one feature of the image; and
preprocessing the image, including filtering the image with a series of linear and non-linear filters.
14. The inspection system of claim 9 , wherein the statistical contrast properties are defined by one or more of histogram, ranges, standard deviation, min-max, average, or ratio of dark to light pixels.
15. The inspection system of claim 9 , wherein the one or more processors include graphic processing unit (GPU).
16. The inspection system of claim 9 , wherein the plurality of binary slices are generated using different thresholds corresponding to different grayscale levels, wherein each binary slice corresponds to a different grayscale level.
17. A non-transitory machine-storage medium embodying instructions that, when executed by a machine, cause the machine to perform operations comprising:
generating a plurality of binary slices of an image of a semiconductor device;
segmenting each binary slice of the plurality of binary slices to create segments based on statistical contrast properties of the image;
inputting a subset of the segments into a machine learning network, wherein the subset of the segments include segments with a higher contrast profile as compared to other segments not included in the subset; and
receiving an output of the machine learning network identifying presence or absence of at least one defect in the image.
18. The non-transitory machine-storage medium of claim 17 , wherein the machine learning network includes a multi-layered machine learning model, and wherein each binary slice corresponds to a different grayscale level.
19. The non-transitory machine-storage medium of claim 17 , wherein the machine learning network is trained to detect defects using a supervised machine learning technique with at least one training image by:
generating a plurality of binary slices of the at least one training image to create training image slices;
segmenting each slice of the training image slices based on statistical contrast properties of the at least one training image to generate training segments;
receiving a defect label for the training segments with a known defect and receiving a no-defect label for one or more segments without the known defect;
inputting the training segments into the machine learning network; and
providing a reference output to the machine learning network based on the labels of the training segments,
wherein the machine learning network performs iterative operations until its output matches the reference output.
20. The non-transitory machine-storage medium of claim 17 , wherein the defect is a low-contrast watermark defect.
21. The non-transitory machine-storage medium of claim 17 , further comprising:
preprocessing the image, including filtering the image with a series of linear and non-linear filters.
22. The non-transitory machine-storage medium of claim 17 , wherein the statistical contrast properties are defined by one or more of histogram, ranges, standard deviation, min-max, average, or ratio of dark to light pixels.
23. The non-transitory machine-storage medium of claim 17 , wherein the plurality of binary slices are generated using different thresholds corresponding to different grayscale levels, wherein each binary slice corresponds to a different grayscale level.