5 December 2025 in Measurement Science and Technology
Wafer defect detection is critical to semiconductor manufacturing, where precise identification of defect types is a key to optimizing process and enhancing efficiency. To address challenges such as low resolution, complex defect distributions, and high-density in wafer images, this paper proposes a novel density-guided dual-stream vision transformer model (DG-ViT). DG-ViT innovatively integrates a Semantic-guided Global transformer (SGGT) and a density-aware graph convolutional network (DAGCN). SGGT captures long-range semantic correlations of defect through global semantic attention, while DAGCN deeply characterizes local structural features via density-driven topological modeling. An adaptive density-aware feature masking leverages adaptive density heatmaps to filter noise, significantly enhancing the clarity of defect representations. Together, these components empower DG-ViT with profound analytical capabilities for complex defect patterns. Experiments on two large datasets demonstrate that DG-ViT outperforms mainstream models, markedly improving defect classification accuracy and robustness in complex and low-resolution scenarios, achieving accuracy of 94.81% on WM-811 K and 98.54% on Mixed-WM38.
8 February 2026 in Expert Systems with Applications