The application of VGGNet based on the improved multi-head attention mechanism in wafer inspection

Accurate detection of defects in wafer images can aid engineers in promptly and precisely identifying abnormal fault factors in the wafer production process. Deep learning is increasingly applied to wafer defect detection due to its strong feature extraction capabilities. However, these methods often require large, well-labeled datasets, which are difficult to obtain in practice. This study addresses dataset imbalance by proposing an improved deep learning model that integrates an enhanced multi-head attention mechanism with a strengthened VGGNet. This approach enables richer feature extraction from imbalanced datasets, improving wafer defect recognition accuracy. By utilizing the improved multi-head attention mechanism, we map the features of input wafer images into multi-dimensional subspaces, enhancing the model's representation capacity and improving its generalization ability. Specifically, we incorporate residual connections into the fully connected layers of the traditional VGGNet, addressing the gradient vanishing problem in deep VGGNet networks.Experimental results on the real dataset WM811K demonstrate that this algorithm significantly enhances the robustness of wafer image detection, and its performance on imbalanced datasets surpasses existing algorithms.