IEEE Project Abstract

An automatic defect classification (ADC) systemidentifies and classifies wafer surface defects using scanningelectron microscope images. By classifying defects, manufacturerscan determine whether the wafer can be repaired and proceed tothe next fabrication step. Current ADC systems have high defectdetection performance. However, the classification power is poor.In most work sites, defect classification is performed manuallyusing the naked eye, which is unreliable. This study proposes anADC method based on deep learning that automatically classifiesvarious types of wafer surface damage. In contrast to conventionalADC methods, which apply a series of image recognition andmachine learning techniques to find features for defectclassification, the proposed model adopts a single convolutionalneural network (CNN) model that can extract effective features fordefect classification without using additional feature extractionalgorithms. Moreover, the proposed method can identify defectclasses not seen during training by comparing the CNN features ofthe unseen classes with those of the trained classes. Experimentswith real datasets verified that the proposed ADC method achieveshigh defect classification performance.