基于MSAF-DeSTSeg的微弱缺陷分割算法

Weak Defect Segmentation Algorithm Based on MSAF-DeSTSeg

摘要:

晶圆表面质量影响产品寿命,需及时检测异常以降低成本,但异常样本难收集且形态多样,有监督学习算法受限,为此提出了一种基于MSAF-DeSTSeg的晶圆表面异常检测算法；利用知识蒸馏DeSTSeg网络分割晶圆异常区域,引入多尺度技术用于特征输出阶段防止出现特征丢失,在网络的分割头位置设计可变形卷积空间金字塔池化模块,增强复杂异常感知并抑制背景干扰；技术创新点包括多尺度特征融合技术和可变形卷积快速空间金字塔池化模块的应用；实验结果表明,在晶粒数据集上,改进后的模型在图像AUC、平均像素精度和实例像素精度上分别达到了97.79%、73.06%和71.77%的准确率；经实际应用,该算法在无缺陷样本下满足晶圆异常检测需求,性能优于原模型。

Abstract:

The quality of wafer surfaces directly impacts product lifespan, necessitating timely anomaly detection to reduce costs. However, anomaly samples are difficult to collect and exhibit diverse morphologies, limiting the effectiveness of supervised learning algorithms. To address this, an MSAF-DeSTSeg-based algorithm for wafer surface anomaly detection is proposed. This approach utilizes the knowledge distillation DeSTSeg network to segment anomaly regions on wafers, incorporating multi-scale technology during the feature output stage to prevent feature loss. A deformable convolutional spatial pyramid pooling module is designed at the segmentation head of the network to enhance the perception of complex anomalies and suppress background interference. The technical innovations include the application of multi-scale feature fusion technology and the deformable convolutional fast spatial pyramid pooling module. Experimental results indicate that on a grain dataset, the improved model achieves accuracy rates of 97.7.