Wafer Defect Image Generation Method Based on Improved Styleganv3 Network

This paper takes a look at training a generator model based on a limited dataset that can fit the distribution of the original dataset, improving the reconstruction ability of wafer datasets. High-fidelity wafer defect image generation remains challenging due to limited real data and poor physical authenticity of existing methods. We propose an enhanced StyleGANv3 framework with two key innovations: (1) a Heterogeneous Kernel Fusion Unit (HKFU) enabling multi-scale defect feature refinement via spatiotemporal attention and dynamic gating; (2) a Dynamic Adaptive Attention Module (DAAM) adaptively boosting discriminator sensitivity. Experiments on Mixtype-WM38 and MIR-WM811K datasets demonstrate state-of-the-art performance, achieving FID scores of 25.20 and 28.70 alongside SDS values of 36.00 and 35.45. The proposed method in this article helps alleviate the problem of limited datasets and makes an important contribution to data preparation for downstream classification and detection tasks.