摘要
为求解电容层析成像中的非线性病态反问题,提出了一种多尺度密集连接网络(MD-Net)模型。该模型由多尺度特征融合模块和密集连接块组成,以通过融合多尺度特征进一步提高图像的重建精度。通过MATLAB仿真实验平台构建了流型数据集,利用密集连接网络的非线性映射能力,完成训练集的学习与训练,并利用测试集进行训练效果评价。在此基础上进行了静态实验。仿真与静态实验结果均表明:与LBP、Landweber迭代算法和其他深度学习方法相比,该方法的重建精度最高、抗噪能力强,并具有良好的泛化能力。
In order to solve the nonlinear ill-posed inverse problem in electrical capacitance tomography(ECT),a multiscale dense connection network(multi-scale densely connected network,MD-Net)model is proposed.The model consists of a multiscale feature fusion module and a densely connected block to further improve the reconstruction accuracy of images by fusing multiscale features.A flow-type data set is constructed by the MATLAB simulation experiment platform,and the learning and training of the training set are completed by using the nonlinear mapping ability of the densely connected network.The training effect is evaluated by using the test set.Static experiments are conducted on this basis.The simulation and static experiments results show that the method has the highest reconstruction accuracy,good noise immunity,and generalization ability compared with LBP,Landweber iterative algorithm,and other deep learning methods.
作者
张立峰
常恩健
ZHANG Lifeng;CHANG Enjian(Department of Automation,North China Electric Power University,Baoding,Hebei 071003,China)
出处
《计量学报》
CSCD
北大核心
2024年第5期678-684,共7页
Acta Metrologica Sinica
基金
国家自然科学基金(61973115)。
关键词
两相流测量
电容层析成像
图像重建
深度学习
密集连接网络
two-phase flow measurement
electrical capacitance tomography
image reconstruction
deep learning
densely connected network
