摘要：

半导体器件封装过程中出现的内部空洞缺陷会直接影响电子设备的性能。针对半导体器件X射线内部图像空洞缺陷中尺度不一、难标注、难定位及噪声干扰等问题, 提出半自动标注方法和基于U-Net的器件内部空洞缺陷检测方法。半自动标注方法使用阈值分割初步定位缺陷区域, 生成缺陷的外接矩形框, 然后人工对矩形框进行精细化修改和完善, 作为提示输入到分段任意模型(SAM)中, 得到高精度的分割结果。半自动标注方法能够节省标注时间且提高标签质量, 克服标注难题。针对经典U-Net方法泛化性较差的问题, 提出一种改进的U-Net方法(EFU-Net)。首先在编码器中引入边缘位置增强(EPE)模块, 通过结合Sobel滤波器和坐标注意力机制加强对图像边缘信息的感知, 有效整合位置信息, 以提高特征提取的准确性; 然后引入特征融合控制(FFC)模块替代传统的跳跃连接, 融合高层特征、低层特征和预测掩码3个特征, 并利用多层并行空洞卷积和注意力门控机制实现更有针对性和高质量的特征融合。在半导体器件数据集上的实验结果表明, EFU-Net的Dice系数和MIoU分别达到70.71%、77.23%, 与U-Net方法相比, 分别提升了14和7.71个百分点, 具有更好的分割性能。

关键词: 缺陷检测, 语义分割, 半自动标注, 边缘位置增强, 注意力机制, 特征融合控制

Abstract:

Internal void defects that occur during the packaging process of semiconductor devices directly affect the performance of electronic devices. Aiming at the problems of different sizes of void defects in internal X-ray images of semiconductor devices and the difficulty in labeling, positioning, and noise interference, a semi-automatic annotation method and a U-Net-based device internal void defect detection method are proposed. The semiautomatic annotation method uses threshold segmentation to initially locate the defect area, generate the bounding rectangle of the defect, and then manually modify and improve the rectangular frame, which is input into the Segment Anything Model (SAM) as a prompt to obtain high-precision segmentation results. Semiautomatic labeling methods can reduce labeling time, improve label quality, and overcome labeling problems. To address the poor generalization of the classic U-Net method, an improved U-Net method (EFU-Net) is proposed. First, an Edge and Position Enhancement (EPE) module is introduced into the encoder. By combining the Sobel filter and coordinate attention mechanism, it enhances the perception of image edge information and effectively integrates position information to improve the accuracy of feature extraction. The Feature Fusion Control (FFC) module is introduced to replace the traditional skip connection, fuse the three features of high-level features, low-level features, and prediction mask, and utilize multilayer parallel atrous convolution and an attention force gating mechanism to enable more targeted and high-quality feature fusion. The experimental results on the semiconductor device dataset show that the Dice coefficient and MIoU of EFA-NET reach 70.71% and 77.23%, respectively, which are improvements of 14 and 7.71, respectively, percentage points compared with the U-Net method and exhibit better segmentation performance.

Key words: defect detection, semantic segmentation, semi-automatic annotation, Edge Position Enhancement(EPE), attention mechanism, Feature Fusion Control(FFC)