Detection Impact of Synthetic Defective Wafer Map Images Generated by Enhanced-CycleGAN

Generative networks are increasingly being used to achieve data augmentation. This paper involves a comprehensive study of the impact of defective wafer map data augmentation by Enhanced-CycleGAN generative network on their detection. First, the closeness of generated images by Enhanced-CycleGAN with real images is examined as compared with the generative networks of Generative Adversarial Network, Variational Auto Encoder, Diffusion, and CycleGAN. Five quantitative metrics of Peak Signal-to-Noise Ratio, Structural Similarity Index Metric, Visual Information Fidelity, Kernel Inception Distance, and Fréchet Inception Distance are employed to assess the similarity of generated synthetic images with real images. The results obtained show that Enhanced-CycleGAN generates synthetic images which resemble real images more closely compared to the other generative networks. Then, the impact of the data augmentation by Enhanced-CycleGAN on the detection of defective patterns is examined via three commonly used deep learning-based detection models of VGG16, ResNet50 and YOLO11. It is shown that the detection accuracy is improved by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 \%}$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 8 \%}$</tex> depending on the detection model used.