Classification of Semiconductor Defects Using Few Shot Learning with Sythetic Minority Oversampling Technique

Nowadays, advanced technologies for semiconductor processing are rapidly increasing and leading to higher prospect of surface defects that must be promptly detected. The complex manufacturing process of semiconductor devices results in defects with diverse shapes and textures, hence, the existing Support Vector Machine (SVM) is utilized. However, SVM led to overfitting, because it was trained on a specific dataset and not generalized well to other datasets. Therefore, this research proposes an advanced Few Shot Learning (FSL) with Sythetic Minority Oversampling Technique (SMOTE) to classify the semiconductor defects. Initially, the input data is gathered from Wafer Maps (WM-811k) dataset, a complete sources of semiconductor wafer maps image. Then, SMOTE is used for preprocessing, as it generates synthetic images by adding new images between existing minority class images and neighbors. After that, the features are extracted with the help of Scale Invariant Feature Transform (SIFT) and finally, FSL is introduced for classification. Here, FSL takes only few samples from input images to classify the various types of semiconductor defects. From the results, the proposed FSL-SMOTE provided better results in terms of accuracy, precision, recall, and F1 score of 98.90%, 97.01%, 97.83%, and 98.52% respectively when compared with existing Convolutional Neural Networks (CNN).