Accurate Detection Method of Si3N4 Wafer Fuzzy Defects Embedded with Hybrid Cross-Attention Mechanism and Feature Gradient Histogram

This study systematically addresses critical challenges associated with blurred edges of wafer defects, including multi-dimensional feature aggregation, abrupt gradient decline, and hierarchical information loss. To tackle these issues, an innovative, precise segmentation method is proposed based on a dual cross-attention mechanism and a feature gradient histogram. Through in-depth analysis of edge-blurring characteristics in wafer defects, a multi-scale embedding matrix equation was formulated to optimize the contour extraction process. Additionally, a multi-level encoder architecture was implemented to enhance the efficiency of edge contour information extraction. To address boundary information loss during segmentation, a boundary gradient optimization model was constructed using multi-scale differential equations, enabling accurate fitting of boundary gradients through feature recombination vectors. Experimental results demonstrate the effectiveness of the proposed wafer defect segmentation approach. The method achieves an average accuracy of 97.51%, with mean intersection-over-union (mIoU) scores exceeding 89% across three distinct types of wafer defect detection tasks. By effectively mitigating the adverse effects of edge blur on segmentation precision, this approach offers a comprehensive solution for wafer defect detection. The contributions of this research not only enhance the accuracy and reliability of defect identification but also provide robust technical support for improving product quality and manufacturing efficiency in high-end semiconductor industries. These advancements hold significant practical value in promoting the high-quality development of the semiconductor sector.