基于深度学习的PCB缺陷检测  

PCB defect detection based on deep learning

作　　者：陈建豪 徐洁 汪志锋[1] CHEN Jianhao;XU Jie;WANG Zhifeng(School of Intelligent Manufacturing and Control Engineering,Shanghai Second University of Technology,Shanghai 201209,China)

机构地区：[1]上海第二工业大学智能制造与控制工程学院,上海201209

出　　处：《现代电子技术》2025年第8期7-12,共6页Modern Electronics Technique

摘　　要：针对当前印刷电路板(PCB)缺陷检测存在的检测精度低、速度慢等问题,设计一种基于改进YOLOv7的WiseYOLOv7算法。首先,在原有self-attention的基础上加入焦点调制网络,将不同粒度级别的空间特征汇总为调制器,以自适应的方式注入查询操作中,省去大量交互和聚合操作,从而使得模型轻量化;其次,利用RCSOSA模块减少特征图的通道数量,同时增强相邻层特征不同通道间的信息交流,提高模型对PCB小目标缺陷的特征提取能力和数据处理效率;最后,选用动态非单调焦点机制的Wise-IoU损失函数来加强对高质量锚盒的选取,优化目标检测器的性能。与YOLOv7基础算法相比,改进算法将平均精度由92.0%提高至96.1%,提高了4.1%,检测时间由21.9 ms缩短到17.9 ms,改进算法在精度和速度上都有明显提升。In allusion to the current printed circuit board(PCB)defect detection problems such as low detection accuracy and slow speed,a Wise-YOLOv7 algorithm based on improved YOLOv7 is designed.A focus modulation network is added on the basis of the original self-attention,and spatial features of different granularity levels are aggregated into modulators and injected into the query operation by means of adaptive way,which can eliminate a large number of interaction and aggregation operations,thus making the model lightweight.The RCSOSA module is used to reduce the number of channels of the feature map,and the information exchange between different channels of adjacent layer features is enhanced,so as to improve the feature extraction capability of the model for small target defects of PCB and the data processing efficiency.The Wise-IoU loss function with dynamic non-monotonic focus mechanism is selected to enhance the selection of high-quality anchor boxes and optimize the performance of the target detector.In comparison with the YOLOv7 base algorithm,the improved algorithm can increase the average accuracy from 92.0%to 96.1%,which is an improvement of 4.1%,and can shorten the detection time from 21.9 ms to 17.9 ms.The improved algorithm has a significant improvement in both accuracy and speed.

关 键 词：印刷电路板 缺陷检测 YOLOv7算法 深度学习 焦点调制网络 RCSOSA模块 Wise-IoU损失函数 

分 类 号：TN41-34[电子电信—微电子学与固体电子学] TP391.41[自动化与计算机技术—计算机应用技术]