CONVOLUTIONAL NEURAL NETWORK FOR WAFER SURFACE DEFECT CLASSIFICATION AND THE DETECTION OF UNKNOWN DEFECT CLASS

Abstract

An automatic defect classification (ADC) system identifies and classifies wafer surface defects using scanning electron microscope images. By classifying defects, manufacturers can determine whether the wafer can be repaired and proceed to the next fabrication step. Current ADC systems have high defect detection performance. However, the classification power is poor. In most work sites, defect classification is performed manually using the naked eye, which is unreliable. This study proposes an ADC method based on deep learning that automatically classifies various types of wafer surface damage. In contrast to conventional ADC methods, which apply a series of image recognition and machine learning techniques to find features for defect classification, the proposed model adopts a single convolutional neural network (CNN) model that can extract effective features for defect classification without using additional feature extraction algorithms. Moreover, the proposed method can identify defect classes not seen during training by comparing the CNN features of the unseen classes with those of the trained classes. Experiments with real datasets verified that the proposed ADC method achieves high defect classification performance.

Existing System

Support vector machine (SVM) and stacked autoencoder (SAE).

Proposed System

This paper proposed a CNN-based ADC method. The proposed method has two contributions. First, the method is specifically designed so that the CNN architecture achieves high classification performance for known defect classes. Second, the method is able to classify unknown defects as an ‘Unknown’ class without retraining the CNN by using a k-nearest neighbors (k-NN) algorithm that is implemented in the feature space created by the CNN training result.

Architecture

BLOCK DIAGRAM