基于改进生成对抗网络的晶圆表面缺陷检测

Wafer Surface Defect Detection Based on Improved Generative Adversarial Network

对于晶圆表面缺陷检测来说,缺陷样本存在着样本数量不足,缺陷表现形式多样的问题。为解决此类问题,提出了一种基于改进生成对抗网络的晶圆表面缺陷检测模型。该模型首先在GANomaly模型的基础上引入了跳层连接,并引入CBAM注意力机制,用以更好地关注图像重要区域,其次引入记忆模块以约束潜在空间的表示,最后在原模型架构上新增一个自编码器架构判别器,以确保训练更稳定,更容易收敛到最佳平衡点。实验结果表明,该模型能够准确分辨具有缺陷的晶圆样本,检测精度达到了0.985,相较于GANomaly算法提升了6.7%。对于Mvtec AD数据集,检测精度达到了0.79,相较于GANomaly算法提升了3%。

For wafer surface defect detection,there are problems that the number of defect samples is insufficient and the defects are in various forms.In order to solve such problems,a wafer surface defect detection model based on improved Generative Adversarial Network is proposed.Firstly,the model introduces layer hopping connection on the basis of the GANomaly model,and introduces CBAM Attention Mechanism to better focus on the important regions of the image.Secondly,it introduces a memory module to constrain the representation of the potential space.Finally,it adds a new autoencoder architecture discriminator on the original model architecture to ensure that the training is more stable and it is easier to converge to the optimal equilibrium point.The experimental results show that the model is able to accurately distinguish wafer samples with defects,and the detection accuracy reaches 0.985,which is improved by 6.7%compared to the GANomaly algorithm.For the Mvtec AD dataset,the detection accuracy reaches 0.79,which is improved by 3%compared to the GANomaly algorithm.