Skip-MemGANs: An Ensemble Generative Adversarial Network Based on Skip Connection and Memory Module for Wafer Defect Detection

To realize the automatic detection of wafer surface defects, we propose Skip-MemGANs unsupervised detection network, which is an ensemble generative adversarial network that automatically detects defects by the difference between the target image and the reconstructed image.The network is composed of three generators and three discriminators. Each generator uses encoder-decoder convolutional neural network with two layers of skip connection and memory module to capture multi-scale input image features. These generators are randomly paired with discriminators, and receive feedback from the three discriminators, while the discriminators receive reconstructed samples from the three generators.Compared with a single GAN, the ensemble GAN can better simulate the distribution of normal data in the high-dimensional image space.We evaluate the single GAN model, GAN ensemble model and other basic models. The results show that our proposed Skip-MemGANs network outperforms other models in wafer defect detection task, the AUC value reached 0.956.