基于CycleGAN改进的晶圆缺陷图数据增强方法

Wafer defect image data augmentation method based on improved CycleGAN

摘要:

针对某些晶圆图故障图案类型样本量较少导致的数据集不平衡问题进行研究，提出了一种基于CycleGAN改进的晶圆缺陷图数据增强方法；在生成器中引入ULSAM注意力机制，以增强跨通道特征提取能力；对判别器优化结构，以减少关键信息丢失，并在模型中采用 PReLU 激活函数以提升梯度稳定性与收敛速度；实验结果表明，所提方法在Donut类别SSIM提升0.203 4，FID降低84.95；基于增强前后的数据集分别训练晶圆缺陷分类模型， Donut 类和 Random 类的分类准确率分别提高 1.1% 和 7.2%；该方法在提升少数类缺陷图生成质量的同时，为提升晶圆缺陷分类检测模型的精度提供了数据支撑。

Abstract:

To address the data imbalance caused by the limited number of samples in certain wafer defect pattern categories, a defect image augmentation method based on an improved CycleGAN framework was proposed. The generator incorporates the ULSAM attention mechanism to enhance cross-channel feature extraction capability. The discriminator is structurally optimized to reduce critical information loss, and PReLU activation is employed to improve gradient stability and accelerate model convergence. Experimental results show that the proposed method increases the SSIM by 0.2034 and reduces the FID by 84.95 for the Donut category. Classification models trained on datasets before and after augmentation achieve accuracy improvements of 1.1% and 7.2% for the Donut and Random categories, respectively. The method enhances the generation quality of minority-class defect images and improves the classification accuracy of wafer defect detection models under imbalanced data conditions.