Wafer Defect Recognition for Industrial Inspection: FCS-VMamba Model and Experimental Validation

Area of Science:

• Computer Vision
• Materials Science
• Semiconductor Manufacturing

Background:

• Semiconductor wafer defect classification is vital for chip yield and reliability.
• Challenges include weak imaging, detail loss, complex backgrounds, and edge device constraints.
• Existing Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) struggle with long-range dependencies and deployment costs.

Purpose of the Study:

• To develop a parameter-efficient model for industrial wafer defect recognition.
• To address limitations of traditional models in handling complex imaging scenarios.
• To leverage the Visual State Space Model (VMamba) architecture for improved performance.

Main Methods:

• Proposed FCS-VMamba, a domain-adapted model based on the VMamba architecture.
• Integrated Frequency Attention (FA) for enhanced feature extraction.
• Incorporated Cross-Layer Cross-Attention (CLCA) and Saliency Feature Suppression (SFS) modules.
• Utilized VMamba for global contextual modeling with linear computational complexity.

Main Results:

• FCS-VMamba achieved 86.06% macro-precision.
• Achieved 87.91% Top-1 accuracy.
• Model has only 1.2 million parameters, demonstrating parameter efficiency.

Conclusions:

• FCS-VMamba offers a practical and efficient solution for industrial wafer defect recognition.
• The model effectively addresses challenges in complex imaging and edge deployment.
• Demonstrates the potential of VMamba-based architectures in specialized industrial applications.