Image Reconstruction of Electrical Capacitance Tomography Based on ADMM-Net
Image Reconstruction of Electrical Capacitance Tomography Based on ADMM-Net
复制标题
DOI：
10.1109/jsen.2023.3288910
复制
发表时间：
2023-08
影响因子：
4.3
通讯作者：
Dongchen Lu;Lifeng Zhang
中科院分区：
文献类型：
--
作者：
Dongchen Lu;Lifeng Zhang
本文提出了一种基于压缩感知原理的交替方向乘子法（ADMMs）算法与神经网络相结合的方法，以提高图像重建的质量和速度，同时解决参数选择问题。首先，采用ADMM算法求解稀疏电容层析成像（ECT）模型，得到相应的迭代过程。然后，将迭代过程依次构造为重构层、卷积层、非线性激活层和乘法器更新层，以形成ADMM-Net网络模型。最后，有限BFGS（L-BFGS）算法被用来优化网络的参数，通过端到端的训练。与单一的网络模型相比，该方法将数学推理过程封装成网络的形式进行图像重建，增强了网络的可解释性。同时，采用网络优化的方法对参数进行调整，避免了传统ADMM算法参数选择的不确定性。为了验证该方法的有效性，进行了仿真和静态实验，并与常用的Landweber算法、Tikhonov正则化算法、共轭梯度（CG）算法和迭代硬阈值（IHT）算法进行了比较。实验结果表明，该方法在重建精度、收敛速度和鲁棒性方面均优于其他五种算法。
This article proposes a method that combines the alternating direction method of multipliers (ADMMs) algorithm with neural networks based on the compressed sensing principle to improve the quality and speed of image reconstruction while solving the parameter selection problem. First, the ADMM algorithm is used to solve the sparse electrical capacitance tomography (ECT) model and obtain the corresponding iterative process. Then, the iterative process is sequentially constructed as a reconstruction layer, convolution layer, nonlinear activation layer, and multiplier update layer to form the ADMM-Net network model. Finally, the limited-BFGS (L-BFGS) algorithm is used to optimize the parameters of the network through end-to-end training. Compared with a single network model, this method encapsulates the mathematical reasoning process in the form of a network for image reconstruction, enhancing the interpretability of the network. At the same time, using network optimization to adjust the parameters avoids the uncertainty of traditional ADMM algorithm parameter selection. To verify the effectiveness of this method, simulations, static experiments, and comparisons with the commonly used Landweber algorithm, Tikhonov regularization algorithm, conjugate gradient (CG) algorithm, and iterative hard thresholding (IHT) algorithm are conducted. The results show that the proposed method outperforms the other five algorithms in terms of reconstruction accuracy, convergence speed, and robustness.