集成电路芯片封装缺陷视觉检测与分类方法研究

Integrated circuit (IC) chip packaging defects will directly affect chip’s operation efficiency and life span. However, existing theoretical research on packaging defects detection and classification is insufficient to be applied to industrial manufacturing, which demands for high real-time capability, accuracy, and stability. To address this issue, this project will do research on methods of defect image interpolation, image segmentation, defect feature extraction and classification in IC chip packaging defects vision detection and classification, and the application to IC packaging & testing factories. Four main points are included in this project. ①The general edge-directed adaptive image interpolation algorithm, in which the edge direction threshold is adjusted adaptively, is proposed to separate strong edges from weak ones, protect weak edges and texture regions. ②The multilevel image segmentation algorithm for defect images is proposed, based on the establishment of the segmentation evaluation function, the improvement of convergence efficiency, segmentation accuracy and stability over the threshold selection and optimization strategy of GSA (gravitational search algorithm). ③Effect defect feature extraction and classification method is proposed, based on SVM-RFE (support vector machine recursive feature elimination) and the balance between classification accuracy and feature number, to reduce feature redundancy. ④The IC chip packaging defects vision detection and classification system is established and applied to actual production. The outcome of this project will provide theoretical and technical support for the application of IC chip packaging defects vision detection and classification, and the development of key high-end equipment.

集成电路芯片封装缺陷直接影响芯片运行效能及寿命，但现有封装缺陷检测与分类的理论研究尚不足以应用到对实时性、准确性及稳定性要求较高的工业制造场合。鉴于此，本项目拟研究封装缺陷视觉检测与分类中的缺陷图像插值、分割及缺陷特征提取与分类方法，应用于半导体封装测试企业。拟研究：①提出通用边缘导向自适应图像插值算法，自适应调整边缘方向判定阈值，具备强弱边缘区分及弱边缘、纹理区域保护能力；②建立缺陷图像多阈值分割评价函数，研究基于引力搜索算法的阈值选取与寻优机制以提高收敛效率、分割精度与稳定性，提出基于智能优化算法的缺陷图像多阈值分割算法；③基于支持向量机递归特征消除，平衡分类准确率与特征数量，降低特征冗余度，提出有效的缺陷特征提取与分类方法；④建立封装缺陷视觉检测与分类系统，应用于企业生产实际。本项目研究成果将为集成电路芯片封装缺陷视觉检测与分类的应用及关键高端设备的研制提供理论与技术支撑。

集成电路芯片封装缺陷视觉检测是集成电路芯片封装测试过程中的重要环节，其核心技术及关键高端装备的研制，是当前我国芯片封测产业面临的一项“卡脖子”问题。本项目针对集成电路芯片封装缺陷视觉检测与分类中的光源设计与优化、芯片图像快速校正与字符区域定位、缺陷图像插值、封装缺陷分类识别等关键技术展开理论研究，探索研究成果的应用与推广。①设计搭建芯片图像自动化采集与检测系统平台，实现连续、高精度芯片图像采集；②针对当前光源照度优化存在照度效果评价因素单一、照度优化方法通用性不足等问题，提出基于改进樽海鞘算法的视觉检测系统LED光源照度优化方法，获取视觉检测系统具有最优照度效果时的光源位姿信息参数，实现视觉检测系统最优照明；③针对芯片图像倾斜与小尺寸芯片激光打印时的字符倾斜、位置错误等缺陷，提出基于改进Harris角点检测的芯片图像快速校正方法，与基于改进凸包检测的芯片图像字符区域定位方法，提高芯片图像倾斜校正与字符区域定位效率；④针对传统定向双三次卷积插值算法在芯片缺陷图像插值过程中存在的边缘退化效应，提出自适应定向双三次卷积插值算法，在图像局部窗口强弱边缘及边缘方向判定过程中引入Otsu算法，用于自适应区分强边缘像素与弱边缘及纹理区域像素，有效保护芯片缺陷图像的弱边缘与纹理区域；⑤提出基于改进Faster R-CNN的芯片封装缺陷识别方法，应用传统数据扩增与改进DCGAN实现缺陷样本数据增强，改进主干网络，应用模型迁移微调网络结构加快模型训练，有效提升芯片封装缺陷分类识别的准确率与速度；⑥设计系统MFC程序与人机交互界面，分析芯片图像采集和检测各环节时间效率，针对芯片高端测试分选装备的机械结构、力矩控制、视觉检测算法及控制系统等集成电路封测行业痛点进行重点攻关，研制适用于多类型封装芯片的高端测试分选装备。