基于目标检测的电路板表面缺陷检测方法研究  

Research on Surface Defect Detection Method of Circuit Board Based on Target Detection

作　　者：宋旭东[1] 孙永博 宋亮[2] 刘君和 SONG Xudong;SUN Yongbo;SONG Liang;LIU Junhe(School of Rail Intlligence Engineering,Dalian Jiaotong University,Dalian 116028,China;Continuing Education College,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学轨道智能工程学院,辽宁大连116028 [2]大连交通大学继续教育学院,辽宁大连116028

出　　处：《大连交通大学学报》2025年第2期113-120,共8页Journal of Dalian Jiaotong University

基　　金：辽宁省科技厅计划项目(2022JH2/101300268);辽宁省教育科学“十四五”规划2022年度课题(JG22DB131)。

摘　　要：针对目前电路板缺陷检测方法效率低、容易出现误检和漏检等问题,提出一种基于改进YOLOX的电路板表面缺陷检测网络。首先,提出一种改进的特征融合网络,将主干网络输出的较浅层特征图引人金字塔自下而上的路径,并作为新增的第四个预测特征图,用于解决因缺陷尺寸较小导致的漏检问题。其次,将CBAM注意力模块引入YOLOX的主干网络,以抑制图像中的几余信息,对重要特征进行强化。最后,改进YOLOX的回归损失,使用SIoU替换原损失函数,考虑标注框和预测框的角度关系,增强网络的收敛速度和对微小缺陷的定位能力。试验结果表明,改进网络在电路板数据集上的mAP为99.70%,与原YOLOX相比提高了2.28%,对缺陷的检测精度有较好提升,且检测速度满足电路板缺陷诊断实时性要求。同时,对比其他主流目标检测网络均表现出明显精度优势,证明了改进算法的有效性,该方法对工业生产中电路板的质量把控和降低检测成本有重要意义。Aiming at the inefficient defect detection method of current circuit board and easy to misdetection and leakage and other issues,a surface defect detection network of circuit board is proposed based on improved YOLOX.An improved feature fusion network is proposed to introduce the shallower feature map output from the backbone network into the bottom-up path of the pyramid and as an added fourth predictive feature map for solving the leakage detection problem caused by the small size of defects.The CBAM attention module is introduced into the backbone network of the YOLOX to suppress redundant information in the image and to enhance important features.Finally,the regression loss of the YOLOX is improved by replacing the original loss function with SIoU,which takes the angular relationship between the labeled and predicted frames into account to enhance the convergence speed of the network and its ability to locate tiny defects.The experiment results show that the mAP of the improved network on the circuit board dataset is 99.7%,which is 2.28%higher than that of the original YOLOX,which provides a better improvement in the detection accuracy of the defects.The detection speed meets the real-time requirements for circuit board defect diagnosis.Compared with other mainstream target detection networks,significant accuracy advantages are obtained,proving the effectiveness of the improved algorithm for industrial production of circuit board quality control and reducing the cost of detection.

关 键 词：电路板 缺陷检测 YOLOX 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]