不确定性引导的芯片空洞分割网络  

Uncertainty⁃guided cavity segmentation network for chip inspection

作　　者：侯云 HOU Yun(Agile and Intelligent Computing Key Laboratory of Sichuan Province,Southwest China Institute of Electronic Technology,Chengdu 610065,China)

机构地区：[1]西南电子技术研究所敏捷智能计算四川省重点实验室,四川成都610065

出　　处：《现代电子技术》2025年第9期86-92,共7页Modern Electronics Technique

基　　金：国家重点研发计划(2023YFB4707200);国家自然科学基金项目(52175031);四川省自然科学基金项目(2023NSFSC0497)。

摘　　要：焊缝中存在的空洞缺陷严重影响芯片的气密性,这使得芯片检测成为智能制造过程中必不可少的步骤。空洞的多尺度形状以及表面灰度分布不均匀现象给基于深度学习的目标分割模型带来具大的挑战。为克服上述问题,文中提出一种基于不确定性引导的芯片空洞分割网络。该网络设计一种多尺度特征提取模块,提高模型对微小空洞的表征能力;并提出一种不确定引导模块,加强模型对空洞边缘的学习能力;此外,网络还引入通道注意力机制,自适应地调整模型整体感受野,增强特定语义的特征表示。为了验证算法的有效性,在所收集的空洞数据集上进行实验。实验结果证明:基于不确定性引导的芯片空洞分割网络对微小空洞的分割效果具有明显的提升;相比于基础U-Net,网络在空洞数据集上提高了约3.4%IoU和5.0%DICE值。Cavities that exist in welds can affect the air-tightness of chips seriously,which makes chip inspection be a crucial step in intelligent manufacturing.However,the multiscale shapes and uneven grayscale distribution of cavities pose great challenges to the deep-learning-based object segmentation models.To tackle these issues,this paper develops an uncertainty-guided cavity segmentation network(UGCSNet)for chip inspection.In the UGCSNet,a multi-scale feature extraction module is designed to improve the model representation capabilities for tiny-scaled cavities,and an uncertainty guidance module is proposed to enhance the model learning ability for the cavity edges.In addition,a channel attention mechanism is introduced to adaptively adjust the receptive field sizes of model and enhance the feature representation of specific semantics.In order to verify the effectiveness of the algorithm UGCSNet,experiments are carried out on the collected cavity datasets.Experimental results have shown that the proposed UGCSNet can improve the segmentation results of tiny cavities significantly.In comparison with the baseline U-Net,the UGCSNet achieves an improvement of IoU(intersection over union)for 3.4%and DICE for 5.0%.

关 键 词：深度学习 图像分割 芯片检测 注意力机制 不确定性 空洞 

分 类 号：TN307-34[电子电信—物理电子学]