基于RT-YOLO-V5的芯片外观缺陷检测  

Chip appearance defect detection based on RT-YOLO-V5

作　　者：郭翠娟[1,2] 王妍 刘净月 席雨 徐伟[1,2] 王坦 GUO Cuijuan;WANG Yan;LIU Jingyue;XI Yu;XU Wei;WANG Tan(School of Electronics and Information Engineering,Tiangong University,Tianjin 300387,China;Tianjin Key Labo-ratory of Optoelectronic Detection Technology and System,Tiangong University,Tianjin 300387,China;Microelectr-onic Device Reliability Laboratory Research and Test Center of Defense Technology for Aerospace Science and Industry,Beijing 100854,China)

机构地区：[1]天津工业大学电子与信息工程学院,天津300387 [2]天津工业大学天津光电探测技术与系统重点实验室,天津300387 [3]中国航天科工防御技术研究院微电子器件可靠性实验室研究与试验中心,北京100854

出　　处：《天津工业大学学报》2024年第3期50-57,共8页Journal of Tiangong University

基　　金：中国博士后科学基金面上基金资助项目(2019M661013);天津市科技计划资助项目(20YDTPJC01090,22YDTPJC00090)。

摘　　要：针对传统的人工芯片检测方法效率低、过分依赖人为操作且误检率高等产生的问题,提出了一种基于ResCBS模块与增加微检测层(Tiny-scale detection layer)的RT-YOLO-V5检测方法用于检测芯片外观缺陷。首先搭建了图像采集系统,并制作了芯片外观缺陷检测数据集。为解决芯片外观缺陷形状不规则、大小不统一、位置不确定带来的检测精度低等问题,在CBS模块中增加短连接,融合输入输出的特征信息,减少信息损失,优化推理速度;其次,增加一个微小尺度的检测层,提高模型对微小目标的特征提取能力。实验结果表明:使用改进后的网络对芯片外观缺陷进行检测,平均精度(mAP)达到95.5%,相对于原始网络提升了5.7%;除此之外,改进后的RT-YOLO-V5在先验框损失(Box_loss)与小目标缺陷的检测精度上都得到了一定的提升。Aiming at the problems caused by traditional manual chip detection,with low efficiency,excessive dependence on human operation and high misdetection rate,an RT-YOLO-V5 detection method was proposed to detect chip appearance defects based on the Res-CBS module and an additional Tiny-scale detection layer.First of all,an image acquisition system was built,and a chip appearance defect detection dataset was produced.Due to the defects are irregular in shape,inconsistent in size and uncertain in location,the performance of YOLO-V5 network can no longer meet the detection requirements.A short connection was added to the CBS module,fusing the feature information of input and output,reducing the information loss and optimizing the inference speed.In addition,a tiny-scale detection layer is added as well,to improve the feature extraction capability of the model for tiny targets.The experimental results show that using the improved network for chip appearance defect detection,mAP reached 95.5%,which was a 5.7%improvement compared to the original network.In addition,the improved RT-YOLO-V5 has gained some improvement in both Box_loss and the accuracy of tiny-scale defect detection.

关 键 词：YOLO-V5 芯片 缺陷检测 特征融合 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]