基于残差自注意力连接的深度电学层析成像方法  

Electrical tomography imaging method based on Deep CNN with residual self-attention skip connection

作　　者：王子辰 陈晓艳[1] 王倩 王迪 谢娜 Wang Zichen;Chen Xiaoyan;Wang Qian;Wang Di;Xie Na(College of Electronic Information and Automation,Tianjin University of Science&Technology,Tianjin 300222,China)

机构地区：[1]天津科技大学电子信息与自动化学院,天津300222

出　　处：《仪器仪表学报》2023年第5期288-301,共14页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61301246)项目资助。

摘　　要：针对电学层析重建(ET)的“软场”特性和逆问题求解的病态性所造成的边界伪影和空间分辨率低的问题,本文提出一种基于迭代展开的预重建模块和改进的注意力深度U形卷积神经网络(CNN)的深度成像方法。其中,预重建模块是由牛顿—拉夫逊迭代算法得到的4层反卷积神经网络;深度U形CNN模块中,在特征提取和重建模块中加入残差连接,用于缓解深度CNN模型中的梯度消失问题,同时引入自注意力跳跃连接实现对全局特征和局部特征的抽象融合,使模型更好地表达图像重建问题的非线性特征。重建结果表明空间分辨率高,内含物边界清晰,重建相对误差为0.10,相关系数为0.93,说明本方法可以有效改善ET图像的质量,为无损测量与检测可视化提出了一种可靠方法。The boundary artifacts and low-spatial resolution in reconstruction due to the‘soft-field'and the ill-posed nature of the inverse problems imaging with electrical tomography(ET)are considered.This article designs a novel deep learning-based ET image reconstruction framework consisted of an unrolling iteration pre-reconstructor and a modified attention-based deep convolutional neural network(CNN)postprocessor.Specifically,the pre-reconstructor,a four-layers deconvolution network,is unrolled by the Newton-Raphson algorithm.The U-Net is the backbone of the post-processor and two carefully designed feature connections are introduced.Firstly,the residual connection is added to the feature extraction and image reconstruction block which could alleviate the reverse gradient vanishing problems.Secondly,the residual self-attention skip connections are proposed which could better fuse the global and local information.These above-mentioned strategies can better express the nonlinear characteristics of ET inverse problems.The visual results show that the reconstruction using the proposed methods has higher spatial resolution and more clear shape representation(i.e.,sharper boundary features and clear medium distributions).The quantity results(RE=0.10 and CC=0.93 in test performance)indicate that the proposed method could improve the imaging results effectively.A reliable method for nondestructive measurement and visualization is promoted.

关 键 词：电学层析成像 预重建模块 深度卷积神经网络 残差连接 自注意力机制 深度学习 

分 类 号：R318[医药卫生—生物医学工程] TH701[医药卫生—基础医学]