Defect detection of semiconductor wafers using neural networks

Abstract

In semiconductor manufacturing process, detecting defects in wafers and determining types of the defects are important. Since manual defect inspection takes much time and has limited duration time, researchers have tried to develop automatic inspection systems. In this thesis, we implemented 3 models: convolutional neural network model, the modified two-stage model, and the semantic segmentation model. CNN model determines whether the image has defect or not. The two-stage model finds the defect regions and classifies the regions. The semantic segmentation model performs pixel-wise classification of the image. We found that CNN model determines whether the semiconductor images have defect or not well. Furthermore, we found that both of the two-stage model and the semantic segmentation model perform pixel-wise classification on semiconductor images. The two models showed almost similar performance but the semantic segmentation model showed a slightly better performance than the two-stage model.