基于深度学习的印刷电路板要素CT图像检测技术研究

发布时间:2019-07-01 17:06
【摘要】:印刷电路板(Printed Circuit Board,PCB)实现了电子元器件间的电气连接,是电子产品中不可缺少的重要部件。在生产与使用过程中,PCB经常会出现焊盘破损、断路等问题,造成设备无法正常工作,因此快速检测与定位故障点对维护电子设备的正常工作具有重要意义。锥束CT(Computed Tomography)成像技术能够获取PCB内部结构的高分辨率三维图像,为非接触条件下无损检测PCB缺陷提供了一种新的技术手段。传统图像检测算法多使用象素级的底层特征,这类特征无语义层次,鲁棒性差,受噪声影响大。深度学习作为近阶段出现的一种特征提取技术,具有精简的模型结构与较强的特征表示能力,能够提取检测目标的高层语义信息,为解决基于锥束CT三维图像的PCB导线、过孔等电路要素检测问题提供了有力的理论工具。本文以PCB无损检测的实际应用需求为背景,以实现PCB过孔与导线的自动化检测为目标,研究基于深度学习的图像检测技术。主要研究内容包括深度模型的特征提取技术、基于深度学习的过孔与导线检测算法以及算法的软件实现。论文主要研究成果如下:(1)介绍了深度学习算法的历史、提出与发展现状。回顾了传统浅层网络模型的发展过程,总结了深层网络与浅层网络相比所具有的优势。分析了传统训练方法在训练深层网络时存在的数据获取、局部极值与梯度弥散等问题。重点介绍了深可信网络的构造与对比散度训练方法,并论述了深可信网络具有的特点。依据编码器与解码器的有无,将现有的深度模型分为生成型网络、区分型网络与解码型网络三类,并对这三类模型的典型结构、改进方法、优缺点做了分析与介绍。(2)针对经典深可信网络存在参数规模较大的问题,提出了一种基于伪标签训练的深度模型。并设计了应用伪标签训练降低模型参数数量的方法。该方法使用生成型网络模型提取训练数据的统计特征,使用主成分分析方法降低统计特征的维数,然后将降维后的特征作为对应数据的标签,再使用重新标记的数据训练一个传统的神经网络,通过调整隐含层神经元节点数可以显著减少神经网络参数规模。经过在四个机器学习标准测试数据集上的测试,本文设计的训练方法在不损失模型泛化性能的前提下,能够降低深可信网络规模至原始的40%左右。针对无监督训练为什么会有助于有监督学习这一问题,通过对伪标签训练实验结果的分析,提出了一种解释无监督学习原理的观点。认为无监督学习的输出相比人工数据标签提供了更加丰富的数据先验信息,伪标签能够更好地反应数据特征,这使模型的代价函数更加精细,无监督训练在代价函数的构造中起到正则化作用。(3)针对PCB的CT图像对比度低、噪声大、存在大量伪影的问题,提出了基于深度学习的PCB过孔与导线检测方法并进行了软件实现。该方法采用伪标签训练方法构造深度模型,通过在样本图像上进行训练,模型可以区分过孔、背景与12种形状的导线。所以基于深度学习的方法可以同时检测过孔与导线要素。对于过孔检测,可以利用模型输出结果直接进行判断。对于导线检测,本文根据导线形状移动滑动窗口,并以此跟踪导线轨迹,直至检测到导线端点为止。实验结果表明,基于深度学习的检测方法能够有效地克服CT图像对比度低、噪声大的问题,并具有一定抗伪影干扰能力。在检测正确率与效果上都要明显优于Hough变换算法。
[Abstract]:The printed circuit board (PCB) realizes the electrical connection among the electronic components, and is an indispensable part of the electronic product. In the process of production and use, the PCB often has the problems of pad breakage, open circuit and the like, which causes the equipment not to work normally, so the rapid detection and positioning failure point is of great significance to the normal operation of the maintenance electronic equipment. Cone-beam CT (CT) imaging technology can acquire a high-resolution three-dimensional image of the internal structure of a PCB, and provides a new technical means for non-destructive testing of PCB defects under non-contact conditions. The traditional image detection algorithm uses the low-level feature of the pixel level, which has no semantic level, poor robustness and large noise. As a feature extraction technology in the near-phase, the depth study has a thin model structure and a strong feature representation capability, can extract high-level semantic information of the detection target, and aims to solve the PCB lead based on the cone-beam CT three-dimensional image, The detection of the circuit elements such as vias provides a powerful theoretical tool. In this paper, based on the actual application requirement of the non-destructive testing of the PCB, this paper aims to realize the automatic detection of the PCB via and the lead, and studies the image detection technology based on depth learning. The main research contents include feature extraction technology of depth model, through-hole and wire detection algorithm based on depth learning, and software implementation of the algorithm. The main research results are as follows: (1) The history, development and development of the depth learning algorithm are introduced. The development of the traditional shallow network model is reviewed, and the advantages of the deep network and the shallow network are summarized. The data acquisition, local extremum and gradient dispersion of the traditional training method in the training of deep network are analyzed. In this paper, the structure of deep trusted network and the training method of contrast divergence are introduced, and the characteristics of the deep trusted network are also discussed. According to the existence of the encoder and the decoder, the existing depth model is divided into three types: the generation type network, the distinguishing type network and the decoding type network, and the typical structure, the improvement method and the advantages and disadvantages of the three types of models are analyzed and introduced. (2) A deep model based on pseudo-label training is proposed for the problem of large parameter scale in the classical deep trusted network. The method of using pseudo-label training to reduce the number of model parameters is also designed. The method uses the generation type network model to extract the statistical characteristics of the training data, reduces the dimension of the statistical feature by using the principal component analysis method, and then uses the characteristic of the reduced dimension as the label of the corresponding data, and then uses the re-marked data to train a traditional neural network, By adjusting the number of the neuron nodes of the hidden layer, the size of the neural network parameters can be significantly reduced. After the test on four machine learning standard test data sets, the designed training method can reduce the scale of the deep trusted network to the original 40% without losing the generalization performance of the model. Based on the analysis of the experimental results of the pseudo-label training, an idea is put forward to explain the principle of unsupervised learning. It is considered that the output of the non-supervised learning provides more abundant data prior information than the artificial data tag, and the pseudo-label can better reflect the data characteristics, which makes the cost function of the model more precise, and the non-supervised training plays a regularized role in the construction of the cost function. (3) Aiming at the problems of low contrast, large noise and large number of artifacts in the CT image of the PCB, the method for detecting the PCB via hole and the lead wire based on depth learning is put forward and the software implementation is carried out. The method constructs a depth model by using a pseudo-label training method, and the model can distinguish between the via, the background and the 12-shaped wires by training on the sample image. The method of depth learning can simultaneously detect the via and lead elements. For the via detection, it is possible to directly judge the result of the output of the model. For wire detection, this article moves the sliding window based on the wire shape and tracks the wire trace until the wire end point is detected. The experimental results show that the detection method based on depth learning can effectively overcome the problems of low contrast and large noise of the CT image, and has certain anti-artifact interference capability. And the detection accuracy and the effect are obviously better than the Hough transform algorithm.
【学位授予单位】:解放军信息工程大学
【学位级别】:硕士
【学位授予年份】:2015
【分类号】:TP391.41

【参考文献】

相关期刊论文 前2条

1 彭静;金亚秋;;Gauss比较函数与复杂遥感图像中线状目标的无偏自动检测[J];计算机辅助设计与图形学学报;2007年12期

2 赵波,孙即祥,张学庆,张翠平;使用Radon变换快速检测直线目标的多应用方法[J];无线电工程;2005年05期



本文编号:2508653

论文下载
论文发表
教材专著
专利申请


    下载步骤:
    1.微信扫码,备注编号 2508653.
    2.
    点击下载


    本文链接:http://www.bigengculture.com/shoufeilunwen/xixikjs/2508653.html