基于YOLO-MCG的PCB缺陷检测算法  

PCB Defect Detection Algorithm Based on YOLO-MCG

作　　者：胡兰兰 邓超 HU Lanlan;DENG Chao(School of Physics and Electronic Information Engineering,Henan Polytechnic University)

机构地区：[1]河南理工大学物理与电子信息学院

出　　处：《仪表技术与传感器》2024年第4期100-106,共7页Instrument Technique and Sensor

基　　金：河南省科技攻关项目(232102210100);河南理工大学基本科研业务费基础研究项目(B类)(NSFRF230601)。

摘　　要：为解决PCB缺陷检测中存在检测精度低,模型尺寸庞大的问题,文中提出基于YOLO-MCG的PCB缺陷检测算法。首先提出多尺度加权通道融合网络缩减模型体积,扩增小目标数据。接着提出混合空间金字塔卷积替换主干网络中SPP结构,扩大深层特征图感受野,增强模型语义信息特征提取性能。最后构建轻量化CG-CSP模块替换主干网络最深层CSP结构,减少网络参数,提高网络过滤冗余背景信息能力。实验结果表明,YOLO-MCG算法在8.13 MB的模型尺寸下平均精度均值为97.72%,与改进前模型比较,mAP提升3.77%,模型尺寸缩减69.89%,有效降低模型复杂度,提升缺陷检测效果。For the problems of low detection accuracy and large model size in PCB defect detection,a PCB defect detection algorithm based on YOLO-MCG was proposed.First,a multi scale weighted channel fusion network was presented to reduce the model volume and amplify small target data.Afterwards,the mixed space pyramid convolution was proposed to replace the SPP structure in the backbone network,which expanded the receptive field of deep feature maps and enhanced the performance of the model semantic information feature extraction.Eventually,a lightweight CG-CSP module was constructed to replace the deepest CSP structure in the backbone network,which reduced the network parameters and improved the network s ability to filter redundant background information.Experimental results show that the YOLO-MCG algorithm obtains an average mean precision of 97.72%with a model size of 8.13 MB.Compared with the pre-improved model,mAP is increased by 3.77%and the model size is reduced by 69.89%,which effectively reduces the model complexity and improves the defect detection effect.

关 键 词：PCB缺陷检测 小目标 混合空间金字塔卷积 轻量化 注意力机制 感受野 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]