Abstract: To solve the problems of low accuracy,slow speed and large number of model parameters in the current Printed Circuit Board(PCB)defect detection network,a Lightweight Weighting Novel Network(LWN-Net)based on improved YOLOv3 was proposed.To solve the excessive number of backbone network(Darknet53)parameters in YOLOv3,a lightweight feature augmentation network was proposed as feature extraction network for the model.Considering that the detection accuracy would be reduced caused by imbalance of semantic information and location information in the process of feature extraction,the weight aggregation distribution mechanism was introduced to eliminate imbalance and improve the feature extraction ability of the model.A novel feature pyramid network was proposed to enhance the network's ability to extract detailed information and reduce information redundancy.To speed up the convergence of the model and improve the detection accuracy,the regression loss function SIoU was added to the network training.The result showed that the model size was compressed by 87.5% by comparing with YOLOv3,but the detection speed was increased by 8.32 frames,the prediction accuracy and recall rate were increased by 0.88% and 1.6%.The proposed network provided a more efficient method for PCB defect detection problem.

Key words: printed circuit board defect detection, YOLOv3, lightweight, SIoU loss function

摘要： 针对现阶段印刷电路板缺陷检测任务中网络精度低、速度慢、模型参数量大的问题,提出基于改进YOLOv3的轻量级权重新型网络(LWN-Net),并提出轻量级特征增强网络作为模型的特征提取网络,解决YOLOv3中主干网络Darknet53参数量过多的问题。考虑到特征提取过程中语义信息和位置信息不平衡会导致检测精度降低,构建权重聚合分配机制消除不平衡,以提高模型特征提取能力。提出新型特征金字塔网络,增强网络对细节信息的提取能力并降低信息冗余度。采用回归损失函数SIoU加快模型的收敛速度并提高检测精度。结果表明,相比YOLOv3,LWN-Net网络的模型规模压缩了87.5%,而检测速度提升了8.32帧·s-1,预测精度和召回率分别提升了0.88%和1.6%。该网络的提出为印刷电路板的缺陷检测问题提供了一种更高效的方法。

关键词: PCB缺陷检测, YOLOv3, 轻量级, SIoU损失函数