SR-FABNet: Super-Resolution branch guided Fourier attention detection network for efficient optical inspection of nanoscale wafer defects

Abstract

In-line optical inspection of nanoscale defects in patterned wafers is crucial for yield management and control in advanced semiconductor production. Current industrial inspection methods primarily rely on die-to-die or die-to-database comparisons. However, due to the continuously shrinking process nodes, optical methods suffer from limited optical resolution and efficiency loss, presenting significant challenges. Therefore, there is an urgent need for efficient and precise optical inspection methods to detect nanoscale physical defects from low-resolution optical diffraction patterns. To address this gap, we propose a super-resolution (SR) branch guided Fourier attention detection network for efficient optical inspections of nanoscale wafer defects. The network establishes a SR branch to guide the utilization of high-resolution image information for defect detection. Moreover, we introduce a novel Fourier Attention Block (FAB) to enhance the model’s capability of discerning defects from sophisticatedly-designed background patterns by exploring the distribution of information in the frequency domain. Additionally, knowledge distillation strategy is also incorporated to improve the deployment efficiency and generalization capability of our model. A series of experiments show that the proposed method achieves an optimal mAP of 96.1 % for the task of patterned wafer defect detection, which is 2.4 % higher than that of YOLOv11s with only a slight inference time increase (10.0 ms to 10.8 ms), achieving a good balance between detection efficiency and accuracy. Meanwhile, the proposed algorithm also works well for complex scenarios even under limited data training. The proposed method provides a strong tool for intelligent and end-to-end defect inspection of patterned wafers.