摘要
由于人工智能技术以及深度学习的不断进步,目标检测有了更多的应用。因为在印刷电路板(Printed Circuit Board)的制作和运输保存过程中许多因素都可能会使电路板产生不同类型的瑕疵,这些瑕疵会影响到电子设备的性能。目前PCB瑕疵检测任务都是通过人工检验的方法去完成,但是人工瑕疵检测一直存在诸多弊端。因此我们尝试将基于机器学习的目标检测技术引入到印刷电路板瑕疵检测中。本文是把PCB瑕疵检测问题转化为目标分类问题,利用梯度直方图(Histogram of Gradient)特征作为分类特征并使用支持向量机(Support Vector Machine)作为分类器。本文从两个层面对传统基于SVM的目标检测算法进行改进。首先,本文提出了一种负样本提取方式,通过增加样本数量以及引入随机性的方式提升了算法的检测精度。第二,我们发现SVM检测过程中的搜索框大小会显著影响检测精度。因此本文利用训练集中的目标框大小作为检测框大小。基于这两种改进方法,本文算法可以实现在检测过程中显著提升检测精度并降低误检率的目的。我们将算法在主流PCB检测数据集,即HRIPCB上进行验证,得到了37.0%的检测准确率。本文算法可以广泛应用于实际PCB瑕疵检测任务中,帮助节省人工成本。
Due to the artificial intelligence technology and deep learning progress,target detection has more applications.It is inevitable that Printed Circuit Board(PCB)will have different types of defects in the process of manufacture,transportation and preservation,which will affect the performance of electronic equipment.At present,PCB defect detection is done by manual inspection,but manual defect detection always has many drawbacks.Therefore,we try to introduce the target detection technology based on machine learning into PCB defect detection.In this paper,PCB defect detection problem is transformed into target classification problem,using Histogram of Gradient as classification feature and Support Vector Machine as classifier.This paper improves the traditional SVM-based target detection algorithm from two aspects.Firstly,this paper proposes a negative sample extraction method,which improves the detection accuracy of the algorithm by increasing the number of samples and introducing randomness.Secondly,we find that the size of the search box will significantly affect the detection accuracy during SVM detection.Therefore,this paper uses the size of the target frame in the training set as the size of the detection frame.Based on these two improved methods,the proposed algorithm can significantly improve the detection accuracy and reduce the false detection rate in the detection process.We verify the algorithm on the mainstream PCB detection dataset,namely HRIPCB,and get the detection accuracy of 37.0%.Our proposed method can be widely used in actual PCB defect detection tasks,which helps to save labor costs.
作者
唐佳泉
Tang Jiaquan(Hubei Institute of Engineering,School of Physics and Electronic Information,Xiaogan Hubei,432000)
出处
《电子测试》
2021年第18期36-39,共4页
Electronic Test