本研究針對半導體製造中晶圓缺陷檢測這一關鍵問題展開探討。隨著電路積 體密度的增加和晶圓設計複雜性的增加，晶圓缺陷變得更加普遍。每個晶圓缺陷的發生都是源於某些製造流程的特定異常行為，一套識別晶圓缺陷的系統，有助於發現半導體製造中的異常製程並採取相應措施加以解決。在半導體製造流程中，準確檢測並識別晶圓上的各種缺陷至關重要，但傳統基於卷積神經網絡（CNN）的方法存在一些固有的缺陷，例如高計算量、過擬合和對特定型態缺陷的處理不足。因此，本研究旨在提出一種新型的卷積神經網絡，將深度可分離卷積、多尺度注意力和可變形卷積等技術結合，以提升模型效率、改善缺陷處理能力並增強模型泛化能力。透過這些方法的綜合應用，我們期望能夠有效解決晶圓缺陷檢測中的挑戰，提高檢測準確性和效率，從而促進半導體製造流程的持續優化與提升。;This study focuses on the crucial issue of wafer defect detection in semiconductor manufacturing. With the increase in circuit integration density and the complexity of wafer design, wafer defects have become more prevalent. Each wafer defect originates from specific abnormal behaviors in the manufacturing process. Accurate detection and identification of various defects on wafers are essential in semiconductor manufacturing. However, traditional convolutional neural network (CNN)-based methods suffer from inherent drawbacks such as high computational complexity, overfitting, and inadequate handling of specific types of defects. Therefore, this study aims to propose a novel convolutional neural network that integrates techniques such as depth-wise separable convolution, multi-scale attention, and deformable convolution to improve model efficiency, enhance defect processing capabilities, and strengthen model generalization ability. Through the comprehensive application of these methods, we expect to effectively address the challenges in wafer defect detection, enhance detection accuracy and efficiency, and promote continuous optimization and improvement of semiconductor manufacturing processes.
