Scholarlink

<p>Accurate classification of wafer-map defect patterns is critical for boosting yield and reducing cost in semiconductor fabrication. To address this task, we combine the advantage of Residual Networks (ResNet)—their ability to minimize information loss in deep networks—with long short-term memory (LSTM) for sequence modeling over spatial tokens. We present the Shortcut3-ResNet with Five Sets and Bidirectional LSTM with Additive Attention Network (SCRBLAA-Net). A Shortcut3-ResNet with Five Sets (SCR5) block first distills spatial features; the resulting vector is reshaped into a pseudo-sequence and analyzed by an attention-augmented bidirectional LSTM (Bi-LSTM), after which the refined sequence is fused back with the original spatial representation. On the WM-811K dataset with the “none” class removed, SCRBLAA-Net achieves a macro-F1 of 93.99%, outperforming the baseline SCR5 by 1.59 percentage points and exceeding the SCR5–LSTM and SCR5–Bi-LSTM variants by 0.78 and 0.21 percentage points, respectively.</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

Combining residual network and bidirectional long short-term memory with additive attention for wafer defect classification

  • 期刊论文

  • 发表日期：

    2025年11月15日

查看全文
  • Gyumin Kang,
  • Gyucheol Lee,
  • Younghoon Kim

摘要

Accurate classification of wafer-map defect patterns is critical for boosting yield and reducing cost in semiconductor fabrication. To address this task, we combine the advantage of Residual Networks (ResNet)—their ability to minimize information loss in deep networks—with long short-term memory (LSTM) for sequence modeling over spatial tokens. We present the Shortcut3-ResNet with Five Sets and Bidirectional LSTM with Additive Attention Network (SCRBLAA-Net). A Shortcut3-ResNet with Five Sets (SCR5) block first distills spatial features; the resulting vector is reshaped into a pseudo-sequence and analyzed by an attention-augmented bidirectional LSTM (Bi-LSTM), after which the refined sequence is fused back with the original spatial representation. On the WM-811K dataset with the “none” class removed, SCRBLAA-Net achieves a macro-F1 of 93.99%, outperforming the baseline SCR5 by 1.59 percentage points and exceeding the SCR5–LSTM and SCR5–Bi-LSTM variants by 0.78 and 0.21 percentage points, respectively.