用于Mini/Micro-LED芯片缺陷检测的全局特征压缩卷积神经网络  

Global feature compression convolutional neural network for defect detection in Mini/Micro-LED chips

作　　者：田心如 褚洁 蔡觉平[1] 温凯林 王宇翔 Tian Xinru;Chu Jie;Cai Jueping;Wen Kailin;Wang Yuxiang(College of Microelectronics,Xidian University,Xi′an 710071,China;Suzhou Honghu Qiji Electronic Technology Co.,Ltd,Suzhou 215000,China)

机构地区：[1]西安电子科技大学微电子学院,西安710071 [2]苏州鸿鹄骐骥电子科技有限公司,苏州215000

出　　处：《仪器仪表学报》2024年第8期174-184,共11页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金面上项目(62274123);陕西省自然科学基础研究计划项目(2024JC-YBQN-0615);中央高校基本科研业务费专项资金项目(XJSJ23054);陕西省博士后项目(2023BSHEDZZ173)资助。

摘　　要：微型发光二极管(Mini/Micro-LED)是下一代显示技术。随着Mini/Micro-LED芯片物理尺寸的微小化,制造良品率下降、集成度激增,Mini/Micro LED芯片的快速、精确检测成为工业生产的关键。然而由于芯片尺寸小、分布密集,单个目标的特征信息占比不足,且工业检测要求检测算法速度快、易部署,Mini/Micro-LED芯片缺陷检测仍面临巨大挑战。针对这些问题,设计了一种压缩注意力细节-语义互补卷积神经网络(CADSC-CNN)。在特征融合网络加入基于自注意力机制的编码器结构,更容易获取全局信息,对小目标的特征信息进行补充;同时对自注意力进行压缩操作减少模型的参数量,提高检测速率。此外,通过工业相机采集的Mini/Micro-LED数据集验证该方法的有效性。实验表明,该方法的平均精度均值(mAP)达到了95.6%,速度为100.6 fps。Mini/Micro-LED represents the next generation of display technology.As the physical size of Mini/Micro-LEDchips becomes smaller,fabrication yields have decreased while the degree of integration has significantly increased.Consequently,fast and accurate inspection of Mini/Micro-LED chips is crucial for industrial production.However,inspecting Mini/Micro-LED chips remains challenging due to their small size and dense distribution.The limited feature information from individual targets and the need for fast,easily deployable inspection algorithms add to these challenges.To address these issues,we designed a compressed attention detail-semantic complementary convolutional neural network(CADSC-CNN).By incorporating an encoder structure based on a self-attention mechanism into the feature fusion network,it becomes easier to acquire global information to complement the features of small targets.Additionally,the compression operation of self-attention reduces the model′s parameter count,thereby improving the detection rate.We validated the effectiveness of this method using a Mini/Micro-LED dataset collected by an industrial camera.Experiments demonstrated that this method achieves a mean average precision(mAP)rate of 95.6%and a speed of 100.6 frames per second.

关 键 词：缺陷检测 Mini/Micro-LED 卷积神经网络 自注意力 

分 类 号：TH89[机械工程—仪器科学与技术] TP391[机械工程—精密仪器及机械]