Customer Challenge

  • Automatically detect and classify the defects in SEM images
  • Combining detection and classification can identify the defect as well as correctly localize the defects.


  • Detect the defects
  • Classify the defects
  • Bin the defects that do not fall into standard bincodes into a new category “Other”


  • The initial prototype implements a deep learning network based on CNN to detect and classify defects present in scanning electron microscope images
  • 5 defect classes were present in the images. A new class ‘other’ was introduced to handle unseen defects
  • As the images available were not sufficient for developing the deep neural network, image augmentation techniques like flip, random crop and contrast enhancement was used


  • Automatic defect detection and classification
  • Identification of multiple types of defects from the image
  • Activation maps overlayed on images as heatmaps for better visualisation


The accuracy achieved is 79.18 % on validation data. The work is currently ongoing. 

Contact us