A Novel Gan-Based Data Augmentation Algorithm for Semiconductor Defect Inspection

A deep learning solution is proposed for the problem of object inspection in semiconductor images. Supervised learning method approaches require large annotated semiconductor datasets, which are often difficult to obtain. Therefore, we develop a new deep convolutional generative adversarial network (DCGAN)) to generate simulated data. Real image data and generated image data are used to train the residual network (ResNet) defect inspection network. Compared to training with the original dataset, using the synthetic dataset resulted in a 3.12% improvement in the accuracy of local defect detection. The total defect inspection accuracy also improves from 93.75% to 95.31%.