.png){kind=icon}
.png){kind=icon}
Return
ResNeSt Wafer Map Defect Pattern Recognition Based on the Multi-attention Mechanism and Enhanced Activation Function
.png){kind=icon}
PRE
AI
DOI：10.1007/s10836-026-06229-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Wafer manufacturing plays an important role in semiconductor industry, which requires high precision and complicated processes, and defects are inevitable in manufacturing. Wafer map defects show different image features due to their formation, the shape, size, and distribution position of defects are random, which limits the effectiveness of traditional machine learning detection. In order to improve the accuracy of wafer map defect pattern recognition and improve the ability to identify wafer map defects, this paper conducts research on wafer map defect pattern recognition based on the ResNeSt method with the multi-attention mechanism and enhanced activation function. The Hswish-ResNeSt50 network is generated by modifying the activation function in the ResNeSt network, which improves the accuracy of wafer map defect recognition. By introducing the CBAM and the self-attention mechanism between the convolutional layers of the Hswish-ResNeSt50 network, three methods for adding the CBAM and the self-attention mechanism are designed, namely: introducing CBAM between the network’s convolutional layers to design the Hswish-ResNeSt50-LC(LC) network; introducing the self-attention mechanism between the network’s convolutional layers to design the Hswish-ResNeSt50-LS(LS) network; introducing both CBAM and the self-attention mechanism between the network’s convolutional layers to design the Hswish-ResNeSt50-LCS(LCS) network. Experimental results show that the above methods can effectively improve the ability of the network to recognize wafer map features. Among them, Hswish-ResNeSt50-LC3, Hswish-ResNeSt50-LCS3 and Hswish-ResNeSt50-LCS4 have the best effect by introducing attention mechanism in different convolutional layers of Hswish-ResNeSt50 network. The accuracy of these three network reached 98.21%, 98.18% and 98.21% respectively.
Keywords:
Wafer map
Defect pattern recognition
ResNeSt50
Attention mechanism
Journal
J
IF:
0
Papers:
21
Citations:
0