【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0.809 2和0.705 1，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0.712 8和0.697 4，准确度为0.834 1，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。更多还原

【Abstract】 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 0.809 2 and 0.705 1, respectively, outperforming existing mainstream methods by 4.52% and 1.74%. The accuracy and recall rates are 0.712 8 and 0.697 4, respectively, with an overall accuracy of 0.834 1, 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.更多还原