一种增强型稠密连接网络模型应用于ECT图像重建  

Enhanced dense connectivity network model applied to ECT image reconstruction

作　　者：马敏 孙妮 MA Min;SUN Ni(Civil Aviation University of China,College of Electronic Information and Automation,Tianjin 300300,China)

机构地区：[1]中国民航大学电子信息与自动化学院,天津300300

出　　处：《振动与冲击》2024年第10期82-88,共7页Journal of Vibration and Shock

基　　金：国家基金面上项目(61871379)。

摘　　要：针对传统神经网络在电容层析成像图像重建中,存在难以较好地实现电容特征张量的底层位置特征和顶层语义特征融合的问题,提出了一种增强型稠密连接网络模型(enhanced dense connectivity network,EDC-net)。首先,训练全连接神经网络获得初步介电常数分布,利用全连接神经网络的输出特征图作为补偿型U-net网络的输入。其次,搭建补偿型U-net网络,在编码和解码器间添加类DenseNet结构的稠密跳跃连接机制,以保留大量的底层位置特征信息,减少模型多个输出节点的特征损失;同时利用多尺度密集空洞卷积模块替代补偿型U-net中普通卷积,扩大模型的感受野并丰富多尺度信息。最后采用高效通道注意力机制模块实现子解码器节点的输出特征跨通道交互,增强模型对重要信息的关注度,提高模型的非线性拟合能力。试验结果表明,与Landweber迭代算法、U-net算法对比,基于EDC-net算法的重建图像分辨率高,成像边缘清晰,且更具鲁棒性。In order to solve the problem that traditional neural networks cannot integrate the bottom position features and the top semantic features of the capacitive feature tensor well in ECT image reconstruction,an enhanced dense connection network model was proposed.First,the initial dielectric constant distribution was obtained by training the Fully Connected Neural Network,and the output characteristic map of the fully connected neural network was used as the input of the compensated U-net network.Secondly,a compensated U-net network was built,and a DenseNet-like dense jump connection mechanism was added between the encoder and decoder to retain a large amount of underlying location feature information and reduce the feature loss of multiple output nodes of the model.At the same time,the multi-scale dense cavity convolutional module was used to replace the ordinary convolution in the compensated U-net to enlarge the receptive field of the model and enrich the multi-scale information.Finally,an efficient channel attention mechanism module was used to realize the cross-channel interaction of the output features of sub-decoder nodes,which enhances the model's attention to important information and improves the nonlinear fitting ability of the model.The experimental results show that the reconstructed images based on this algorithm have higher resolution,clearer imaging edges,and more robustness than the Landweber iterative algorithm and the U-net algorithm.

关 键 词：电容层析成像 图像重建 稠密跳跃连接 特征跨通道交互 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]