基于圆卷积神经网络的粘连导电粒子检测  

Detection of conductive multi-particles based on circular convolutional neural network

作　　者：刘子龙 罗晨[1,2] 周怡君 贾磊[1,2] LIU Zilong;LUO Chen;ZHOU Yijun;JIA Lei(College of Mechanical Engineering,Southeast University,Nanjing 211189,China;Wuxi Shangshi-finevision Technology Co.,Ltd,Wuxi 214174,China)

机构地区：[1]东南大学机械工程学院,江苏南京211189 [2]无锡尚实电子科技有限公司,江苏无锡214174

出　　处：《光学精密工程》2024年第11期1788-1800,共13页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.51975119,No.52375487);江苏省重点研发计划资助项目(No.BE2023041)。

摘　　要：为了提高粘连导电粒子检测的精度和稳定性,提高评价指标的客观性和与实际生产需求的适配度,提出了基于圆卷积神经网络的粘连粒子检测。首先提出了更适合粒子检测的圆卷积,修改可变形卷积的采样策略,限制采样点偏移量x,y坐标的自由度,增加尺寸控制参数作为补偿。然后,基于U-MultiNet网络架构将圆卷积替代原有卷积形式,并增加注意力机制,通过标签图计算自注意力,以此作为权重修改损失函数和标签图。最后,提出可重复性和可再现性指标综合评价算法的精度和稳定性。实验结果表明,本文方法的可重复性和可再现性分别为0.8092和0.7051,相比现有主流方法提高了4.52%和1.74%;精确度和召回率分别为0.7128和0.6974,准确度为0.8341,比现有主流方法高1.68%。相比于现有主流方法,该方法对于粘连干扰的粒子检测效果有明显提升,可以满足工业上对粒子检测精度、稳定性和实时性的要求。To enhance the accuracy and stability of conductive particle detection and to meet actual production demands,a multi-particle detection method based on a simplified deformable convolutional(circular convolutional)neural network is proposed.First,an appropriate model and network are chosen based on the characteristics of the detection task and target.Then,a deformable convolution sampling strategy is introduced and modified to restrict the sampling point offset,with added size control parameters.A circular convolution,more suitable for particle detection,replaces some convolutional layers of the original network.Additionally,an attention mechanism is introduced to calculate self-attention through label graphs,which serve as weight modification loss functions and label graphs.Finally,a comprehensive evaluation algorithm for the accuracy and stability of repeatability and reproducibility indicators is proposed.The results show that the repeatability and reproducibility indicators of our method are 0.8092 and 0.7051,respectively,outperforming existing mainstream methods by 4.52%and 1.74%.The accuracy and recall rates are 0.7128 and 0.6974,respectively,with an overall accuracy of 0.8341,surpassing existing methods by 1.68%.Compared to existing mainstream methods,our approach significantly improves the particle detection performance under adhesion interference,meeting industrial requirements for accuracy,stability,and real-time processing.

关 键 词：机器视觉 深度学习 神经网络 导电粒子检测 圆卷积 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]