Integrated circuit defect classification based on multi-layer attention mechanisms

In the integrated circuit (IC) manufacturing process, defects directly impact the final product yield. Integrated circuit defects are characterized by a wide variety of defect types and complex circuit structures. We proposes a defect detection model based on multi-layer attention mechanisms, which enables the detection and classification of common defects in etching processes. First, we use a pre-trained backbone to extract features from different layers. Then, we perform feature encoding and fusion across these different layers. Finally, we utilize an end-to-end decoder to determine the location and type of defects. Compared to similar methods, our method shows a significant improvement in accuracy across different types of defects and requires fewer training samples. Some types of defects have already met the application requirements, and our approach incurs lower training costs when dealing with new types of defects, necessitating only fine-tuning of the model rather than retraining the entire network.