Electrical resistance tomography image reconstruction based on one-dimensional multi-branch convolutional neural network combined with attention mechanism

Electrical resistance tomography (ERT) is an important branch of process tomography (PT), which has been developed for decades. Image reconstruction is a critical step in ERT, where the object of reconstruction is the conductivity distribution of the measured field. Traditional algorithms cannot accurately establish the mapping between the measured voltage and conductivity distribution. With the development of machine learning, the convolutional neural network (CNN) has become a new image reconstruction method. Specific results have been achieved in ERT image reconstruction using CNNs. This study proposes a one-dimensional multi-branch convolutional neural network (1D-MBCNN) for ERT image reconstruction, which could retain the 1D spatial structure of the measured voltage and adaptively and efficiently extract feature information. COMSOL software and the PyTorch framework are used to build the dataset and train the neural network model, respectively. The advantages of the multi-branch structure and the effectiveness of the attention mechanism in ERT image reconstruction are verified by RIE and CC. We also evaluated the practicality of this method in the ERT system. Based on the results of different experiments, the method proposed in this paper has good imaging accuracy, noise resistance, generalization ability, and robustness.     

An image reconstruction algorithm based on total variation with adaptive mesh refinement for ECT

In this paper, a robust image reconstruction algorithm for electrical capacitance tomography (ECT) is proposed. The key feature of the algorithm is the use of adaptive mesh refinement based on total variation (TV) in solving the inverse problem. It keeps the edge preserving and scale-dependent properties of total variation regularization, and enhances the distinguishability by using adaptive mesh refinement. This strategy improves the spatial resolution efficiently with less calculation and is less underdetermined than uniform refinement. Simulation and experimental results show that the algorithm performs better than both standard Tikhonov regularization and the conventional total variation method.     

基于编码-解码对称神经网络的高分辨率图像重构机理研究

针对目前许多图像重构算法存在重构出来的图像不清晰、分辨率低等问题,提出了一种基于编码-解码对称神经网络的高分辨率图像重构算法。首先将图像进行压缩获取低分辨率图像,然后将低分辨率图像作为输入图像经过编码-解码对称神经网络,并利用其中的卷积神经网络进行编码得到特征图像,最后再利用反卷积神经网络进行解码实现图像的细节恢复。实验结果表明,经过基于编码-解码对称神经网络重构出来的图像比之前的低分辨率图像更加清晰,图像的分辨率得到了提高。     

Image reconstruction algorithm of electromagnetic tomography based on fractional Kalman filter

Aiming at the problem of low quality of image reconstruction of electromagnetic tomography (EMT), in this paper, an image reconstruction algorithm of EMT based on fractional Kalman filter (FKF) is proposed. Firstly, the principle of EMT and the principle of state equation of FKF are expound respectively. FKF is often used in the state estimation of nonlinear systems. There is a nonlinear relationship between the object field distribution and the sensor signal in the EMT. Therefore, according to this feature, FKF is applied to the image reconstruction algorithm of EMT. The image reconstruction process of EMT is regarded as the state estimation process of FKF, the normalized measurement voltage is taken as the observation value, and the sensitivity matrix is taken as the measurement matrix. To establish the nonlinear state estimation equation of the FKF and a priori estimation error covariance equation in the EMT, the gray value of image obtained by LBP is used as the initial value of the state estimation, a prior estimation state vector and a priori estimation error covariance matrix are obtained by prediction update, the Kalman filter gain and the posterior estimation error covariance matrix are obtained by the correction feedback process. After repeated iterations, the final state vector, i.e. reconstructed image of EMT is obtained. Finally, simulation experiments are carried out for seven different flow patterns. The results show that the image error and correlation coefficients of the reconstructed image of this algorithm are better than traditional algorithms such as LBP, Landweber, Kalman filter, and have better anti-noise effect than Kalman filter. Therefore, the image reconstruction algorithm of FKF is a new method and means to study the image reconstruction of EMT.     

基于径向基函数神经网络的超分辨率图像重建

为了突破成像极限,经济可行地获取高质量的卫星图像,提出了一种基于径向基神经网络的超分辨率图像重建算法。以径向基神经网络为基础,依据卫星图像退化模型获取网络训练所需的学习样本图像,采用向量映射的方式加速网络收敛。其中,径向基函数的中心、宽度及网络的隐含层数、连接权值是决定径向基神经网络的关键参数,直接关系到网络的重建性能。采用最近邻聚类算法,动态地建立起基函数的中心及宽度,自适应地确定网络的隐含层数及连接权值。建立起的径向基函数神经网络显著地提高了图像重建性能和网络收敛速度(221s即可收敛)。仿真实验和泛化实验表明,训练好的径向基神经网络可以有效地进行卫星图像的超分辨率重建,效率高,误差小。     

基于改进共轭梯度法的ERT图像重建

针对电阻层析成像(ERT)图像重建中灵敏度矩阵的病态特性导致共轭梯度法的收敛率低的问题,提出了改进的共轭梯度算法,ERT图像重建前先对数据进行归一化预处理,将解空间映射到Krylov子空间中,再通过共轭梯度法求解低维子空间中的反问题。分别利用共轭梯度法、预处理共轭梯度法和改进共轭梯度法对典型的气水两相流模型做了仿真实验。实验结果表明,改进共轭梯度法能够提高重建图像的质量,并且相对于其他算法,降低了计算时间。     

基于MCMC方法的电容成像图像重构算法

研究基于概率统计的电容成像图像重构算法,以马尔科夫随机场的方式给出介电常数分布的先验概率,利用电容成像(electrical capacitance tomography,ECT)线性模型得到似然函数,通过马尔科夫链蒙特卡罗(Markov chain Monte Carlo,MCMC)方法对介电常数分布的后验概率密度进行采样,马尔科夫链的转移核利用Metropolis-Hastings方法得到,结合嵌套迭代提高计算效率。仿真结果表明,嵌套迭代-MCMC方法在正则化参数设置合适的条件下,可以得到较好的图像质量,基于MCMC方法图像重构算法为解决ECT图像重构问题提供一种新思路。     

模型参数对卷积神经网络电容层析成像图像重建的影响

卷积神经网络凭借其较强的非线性拟合能力,在电容层析成像图像重建中逐渐得到应用。本文针对卷积神经网络模型超参数调节问题,研究了模型参数对卷积神经网络电容层析成像图像重建的影响。首先,通过数值方法构建了包含80 000组随机流型与40 000组典型流型的"电容矩阵-介质分布"数据集;然后,通过该数据集中的训练集对不同超参数的卷积神经网络模型进行训练和验证,并系统研究了网络初始化、网格密度、卷积核数、全连接层神经元数以及隐藏层结构等超参数对图像重建精度的影响;接着,利用额外生成的12 000组数据作为测试集对各网络模型性能进行评价;最后通过静态实验,对不同网络模型的图像重建效果进行了比较和分析。结果表明:网络隐藏层结构对图像重建精度影响较大,而网络初始化、网格密度、卷积核数以及全连接层神经元数等超参数对重建精度影响较小。     

一种超视场零件影像测量仪的研制

针对超过摄像机视场范围的待测零件,必须进行多次拍摄,然后对系列图像进行拼接以形成完整的零件图像。为了实现图像获取和拼接的自动化,设计了X-Y-Z三轴运动控制平台对定焦相机进行精确定位,讨论了仪器的配置、基于归一互相关测度的模板匹配算法的图像拼接技术以及算法优化策略。     

CT特征重建技术研究

传统的CT重建技术在检测物体内部细小缺陷时,具有噪声难以滤除、分辨率低、对比度差等特点,滤波器设计和实现具有难度.文中提出一种特征CT重建技术,能有效改善上述缺陷,它是以突现构件特征为目的的CT重建算法,重建图像的目的不是达到整体图像质量最好,而是从突现重建图像特征出发,达到重建图像作用.文中首先介绍了二维及三维特征CT重建算法理论,并通过详细分析、实验结果与参数计算的对比,验证了特征CT重建方法在高频信息、抗噪能力、滤波器设计及实现速度方面的明显优势.     

基于遗传算法的组合ERT图像重建算法研究

针对目前电阻层析成像图像重建算法存在成像精度较低的问题,以及为了满足应用于多相流领域的精度要求,提出一种基于遗传算法的组合算法,将线性反投影算法、修正的牛顿-拉夫逊类算法与区间剖分引入遗传算法种群初始化操作中,同时为了改善单纯遗传算法局部搜索能力差与未成熟收敛的问题,将粒子群算法引入遗传算法变异操作中。实验结果表明组合算法效果明显优于线性反投影算法,修正的牛顿-拉夫逊类算法,有效克服了遗传算法早熟收敛现象,提高了成像精度。     

Fast reconstruction of computerized tomography images based on the cross-entropy method

Computerized tomography (CT) has been applied to multi-phase flow measurement in recent years. Image reconstruction of CT often involves repeatedly solving large-dimensional matrix equations, which are computationally expensive, especially for the case of on-line flow regime identification. In this paper, a minimum cross-entropy (MCE) reconstruction based on wavelet multi-resolution processing, i.e., an MRMCE method, is proposed for fast reconstruction of CT images. Each row of the system’s matrix is transformed by 1-D wavelet decomposition. A regularized MCE solution is obtained using the simultaneous multiplicative algebraic reconstruction technique (SMART) at a coarse resolution level, where important information of the reconstructed image is contained. Then the solution in the finest resolution is obtained by inverse fast wavelet transformation (IFWT). Both computer simulation and experimental work were carried out for oil-gas two-phase flow regimes. Results obtained indicate that the MRMCE method improves the resolution of the reconstructed images and dramatically reduces the computation time compared with the traditional linear back-projection (LBP), MCE and algebraic reconstruction technique (ART) methods. Furthermore, the new method can also be used to accurately estimate the local time-averaged void fraction of dynamic two-phase flow. It is suitable for on-line multi-phase flow measurement.     

基于改进卷积神经网络的燃气调压器故障识别研究

针对传统故障识别方法不仅过分依赖专家经验对故障特征进行提取且识别准确率不高的问题,在深度学习理论基础上,提出了一种将一维卷积神经网络与SVM分类器相结合的改进深度卷积神经网络,实现调压器“端到端”的故障识别。首先,介绍了传统卷积神经网络结构;其次,将改进后的一维卷积神经网络与SVM相结合,提出了基于1-MsCNN-SVM算法的调压器故障识别模型,并对模型的组成部分进行了介绍;然后,通过对比实验确定了模型的卷积核长度和卷积层组数;最后,为验证模型的有效性,基于燃气调压器故障数据集,开展了燃气调压器故障识别研究。研究结果表明,改进后的1-MsCNN-SVM算法故障识别准确率高达99.20%,模型具有较好的分类准确率。     

A novel method of plotting sensor sensitivity field and image reconstruction algorithm for electrical capacitance tomography system

Sensor sensitivity field in electrical capacitance tomography is affected by the distribution of multiphase medium, which is the peculiarity of soft field. This brings great difficulty for image reconstruction. To improve the quality of image reconstruction, it is important to analyze the distribution of the sensitivity field. In this article, using the finite element method, we expound a kind of novel plotting pattern to field, which is the distribution of sensitivity field through computer simulation. From experiments and results of sensitivity field analysis, a novel method of image reconstruction based on genetic algorithms is presented. The finite element model is correct and simulation result is fine by adopting unequal interval plotting patterns. At the same time, the result of image reconstruction has high precision. __________ Translated from Chinese Journal of Scientific Instrument, 2005, 26(3) (in Chinese)     

结合粒子群优化和综合评价的脉冲耦合神经网络图像自动分割

为了解决脉冲耦合神经网络(Pulse Coupled Neural Network,PCNN)在图像分割中多参数设定以及评价准则单一的问题,提出了一种结合粒子群优化算法(Particle Swarm Optimization,PSO)和综合评价准则的PCNN图像自动分割方法。采用单调递增阈值搜索策略的PCNN改进模型,将PSO优化原理与由交叉熵参数,边缘匹配度和噪点控制度共同构成的综合评价相结合,以综合评价作为粒子的适应度函数,自动寻优获取PCNN图像分割模型的目标时间常数,连接系数以及迭代次数n,从而实现全参数自适应的PCNN图像分割。实验结果表明算法在保证PCNN运行效率下对不同类型图像都能进行正确完整的分割并兼顾纹理细节的保留。从实验数据可以看到,本文算法在综合评价和通用综合指标上均优于其他对比算法,综合评价平均优于其他算法10.5%。客观评价结果与视觉主观评价相一致,分割较理想,算法具有较高的鲁棒性。     

基于概率神经网络的尿样颜色识别

基于概率神经网络,提出一种与尿液反应后尿试纸的颜色识别方法。针对颜色色空间转换的非线性复杂关系,获取标准阈值颜色色度值,进行归一化处理后,建立基于概率神经网络的尿样颜色识别模型。实验结果表明,用概率神经网络进行尿样颜色识别是可行而有效的。与颜色色差评价方法作比较,该方法无须进行色空间转换,只利用设备原有RGB颜色空间的RGB值即可实现,更易于操作。     

基于加权奇异值分解截断共轭梯度的电容层析图像重建

针对电容层析成像技术(ECT)中的软场效应和病态问题,提出了一种基于加权奇异值分解(SVD)截断共轭梯度的电容层析(ECT)图像重建算法。阐述了电容层析成像工作原理,提出了12电极ECT系统的测量方法。在分析灵敏度矩阵的奇异值分解理论的基础上,推导出了加权SVD截断共轭梯度的数学模型,并利用Tikhonov方法进行正则化加权处理。最后,分析了算法的收敛性,并将其应用于电容层析成像系统的图像重建中。实验结果表明,对于层流,截断共轭梯度算法的平均误差能达到27.54%,全部流型平均迭代步数达到13步,与LBP、Landweber和CG算法比较,该算法具有成像效果好,成像速度快,易于实现等特点。     

A comparison of 3D cone-beam Computed Tomography (CT) image reconstruction performance on homogeneous multi-core processor and on other processors

Computed Tomography (CT) has become an important noninvasive tool in diagnostic medicine. In fact, the use of cone-beam CT is growing in the clinical area due to its ability to provide 3D volumetric information. This study aims to parallelize an analytical 3D image reconstruction algorithm and then implement it on a commercial homogeneous multi-core processor. It is indicated from our results that the homogeneous multi-core processor actually achieved faster reconstruction than the heterogeneous processor-based or the FPGA-based implementation. Our results demonstrated that the shared cache mechanism associated with the homogeneous processor would represent a significant benefit to performing high-speed CT reconstruction or other High Performance Computing (HPC) applications. In addition, while the Graphic Processing Unit (GPU) is considered the most powerful processor nowadays, our experimental results further indicated that the joint use of the multi-core CPU and GPU could enhance the processing power as well as improve the computational efficiency.     

Image reconstruction of electrical capacitance tomography based on optimal simulated annealing algorithm using orthogonal test method

Electrical capacitance tomography (ECT) is a visualization measurement method for two-phase flow. Imaging permittivity distributions using electrical capacitance tomography has always been one of the most significant issues studied by scholars, and the algorithm will have a great impact on the accuracy of image reconstruction result. This paper applies simulated annealing (SA) algorithm to image reconstruction in ECT. However, some parameters of SA algorithm need to be optimized in order to obtain better reconstructed images in ECT. The influence of different parameter values in SA algorithm for image reconstruction in ECT is studied, and a set of optimal parameters of the SA algorithm is obtained based on the orthogonal experimental design method in this paper. At the same time, simulation and static experiments are conducted. Reconstructed images by SA algorithm with optimized parameter are compared with the linear back projection (LBP) and Landweber iterative algorithms. The results show that better images can be obtained for typical oil-gas two-phase flow using SA algorithm. The quality and shape fidelity of reconstructed image for the central object are obviously improved.     

基于模糊神经网络 PID 的开关磁阻电机控制系统研究

针对开关磁阻电机存在的转矩脉动大、噪声大、速度不稳定等问题,对开关磁阻电机的启动、运行、调速等方面进行了研究,提出了一种基于模糊神经网络PID的控制方法,将模糊控制理论与BP神经网络相结合,构成了模糊BP神经网络,根据系统误差,误差的变化,以及误差变化的变化实时调整PID控制参数,使电机在整个转速范围内获得了最优的PID参数。实验采用DSP作为控制核心,不对称逆变桥作为功率变换器,驱动一台2 k W的开关磁阻电机运行。研究结果表明,该方法大大改善了开关磁阻电机控制系统的动、静态性能,控制精度高、转矩脉动小,对干扰有较高的鲁棒性。