Defect inspection method and defect inspection system

A defect inspection system, comprising:

• a first connecting device, configured to connect at least one optical inspection tool;
• a second connecting device, configured to connect a defect review tool;
• a storage medium, configured to store images retrieved by the first connecting device;
• a processor, coupled to the first connecting device, the second connecting device and the storage medium, and configured to execute instructions to perform steps of:
• retrieving a plurality of candidate defect images from a plurality of inspection images obtained by the at least one optical inspection tool performing hot scans on at least one wafer by using the first connecting device and extracting a plurality of attributes from the inspection images;
• creating a random forest classifier including a plurality of decision trees for classifying the candidate defect images, wherein the decision trees are built with different subset of the attributes and the candidate defect images;
• retrieving a plurality of candidate defect images from one of the at least one optical inspection tool in runtime by using the first connecting device;
• applying the candidate defect images to the decision trees in the random forest classifier, classifying the candidate defect images into nuisance images and real defect images according to votes of the decision trees, and filtering out the nuisance images from the candidate defect images;
• and
• sampling the real defect images with the votes over a confidence value and transmitting the sampled real defect images to the defect review tool for microscopic review by using the second connecting device.

A defect inspection method, adapted to an electronic apparatus, the method comprising:

• retrieving a plurality of candidate defect images from a plurality of inspection images obtained by at least one optical inspection tool performing hot scans on at least one wafer and extracting a plurality of attributes from the inspection images;
• creating a first random forest classifier including a plurality of first decision trees for classifying the candidate defect images, wherein the first decision trees are built with different first subset of the attributes and the candidate defect images;
• creating a second random forest classifier including a plurality of second decision trees for classifying the candidate defect images, wherein the second decision trees are built with different second subset of the attributes and the candidate defect images, and the first subset of the attributes and has a complexity in a decision boundary higher than a complexity in a decision boundary of the second subset of the attributes;
• retrieving a plurality of candidate defect images from one of the at least one optical inspection tool in runtime;
• applying the candidate defect images to the first decision trees in the first random forest classifier, classifying the candidate defect images into nuisance images and real defect images according to votes of the first decision trees, and filtering out the nuisance images from the candidate defect images;
• applying the rest of the candidate defect images after filtering to the second decision trees in the second random forest classifier and classifying the candidate defect images into nuisance images and real defect images according to votes of the first decision trees, and filtering out the nuisance images from the candidate defect images;
• and
• sampling the real defect images with the votes over a confidence value for microscopic review.