Scholarlink

<p>Wafer defect detection is pivotal in the semiconductor manufacturing process, directly influencing production efficiency and product quality. With the increase in integration and design complexity of semiconductor wafers, the variety and number of defects have also risen. Given the inefficiencies of manual inspection, recent advancements in state-space models (SSM), exemplified by Mamba, have demonstrated proficiency in modeling long-range dependencies while maintaining linear computational complexity. Inspired by this, we propose a dual-path fusion network(DPFNet), a VisionMamba-based approach tailored for image classification and wafer defect detection. DPFNet integrates state-space modeling with depthwise separable convolution (DSC) techniques, effectively capturing global sequence information and local detailed features. Utilizing PatchEmbedding, the input image is segmented into chunks, with the SSM module modeling dependencies between these sequences. Concurrently, the convolutional branch, incorporating DSC and channel attention modules, focuses on local feature extraction. Through a gating mechanism, these dual feature representations are adaptively fused, enabling a step-by-step extraction and processing of features from low to high levels via multilayer DPFBlocks. Our extensive experiments on the MixedWM38 dataset demonstrate DPFNet’s outstanding performance, achieving accuracy, precision, recall, and F1-score of 98.84%, 98.70%, 98.69%, and 98.69%, respectively. Compared to existing deep learning models, DPFNet exhibits superior accuracy and robustness, particularly in complex hybrid defect scenarios. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

State-space models and depth separable convolution for high-precision hybrid wafer defect detection

  • 期刊论文

  • 发表日期：

    2025年6月14日

查看全文
  • Haibin Zhang,
  • Tingting Lang,
  • Cheng Wang,
  • Wei Liu,
  • Shaosheng Tang,
  • Yanjun Ji

摘要

Wafer defect detection is pivotal in the semiconductor manufacturing process, directly influencing production efficiency and product quality. With the increase in integration and design complexity of semiconductor wafers, the variety and number of defects have also risen. Given the inefficiencies of manual inspection, recent advancements in state-space models (SSM), exemplified by Mamba, have demonstrated proficiency in modeling long-range dependencies while maintaining linear computational complexity. Inspired by this, we propose a dual-path fusion network(DPFNet), a VisionMamba-based approach tailored for image classification and wafer defect detection. DPFNet integrates state-space modeling with depthwise separable convolution (DSC) techniques, effectively capturing global sequence information and local detailed features. Utilizing PatchEmbedding, the input image is segmented into chunks, with the SSM module modeling dependencies between these sequences. Concurrently, the convolutional branch, incorporating DSC and channel attention modules, focuses on local feature extraction. Through a gating mechanism, these dual feature representations are adaptively fused, enabling a step-by-step extraction and processing of features from low to high levels via multilayer DPFBlocks. Our extensive experiments on the MixedWM38 dataset demonstrate DPFNet’s outstanding performance, achieving accuracy, precision, recall, and F1-score of 98.84%, 98.70%, 98.69%, and 98.69%, respectively. Compared to existing deep learning models, DPFNet exhibits superior accuracy and robustness, particularly in complex hybrid defect scenarios.