基于多尺度自适应特征聚合网络的ECT图像重建

ECT image reconstruction based on multi-scale adaptive feature aggregation network

针对深层卷积神经网络在电容层析成像图像重建过程中存在电容特征提取尺度单一、中间层特征利用率不高等问题,提出了一种多尺度自适应特征聚合网络模型。首先,利用堆叠的增强型选择核卷积模块设计了一种特征增强模块(FEM),并通过串联多个FEM自适应地提取电容向量多个尺度的特征信息,极大地减少了使用普通卷积所带来的伪影现象;其次,引入了一种特征聚合机制,采用长短残差连接加强了远近特征信息的相关性,解决了网络中间层特征利用不充分的问题。实验结果表明,与传统算法及卷积神经网络算法相比,所提方法在主观视觉效果和客观评价指标上都具有更好的性能,图像相关系数最高达到0.9629,图像相对误差降低至0.0530。

To address the problems of single capacitance feature extraction scale and low utilization of intermediate layer features in the image reconstruction process of electrical capacitance tomography based on deep convolution neural network,a multi-scale adaptive feature aggregation network model is proposed for electrical capacitance tomography image reconstruction.Firstly,a feature enhancement module(FEM)is designed by using stacked enhanced selection kernel convolutional module,which adaptively extracts feature information from multiple scales of the capacitance vector by concatenating multiple FEM.The artifacts caused by using ordinary convolution is reduced.Secondly,a feature aggregation mechanism is introduced,which uses long and short residual connections to enhance the correlation of far and near feature information.The problem of insufficient utilization of middle layer features in the network is solved.Compared with traditional algorithms and CNN algorithm,the experimental results show that the proposed method has better performance in subjective visual effects and objective evaluation indicators,with the highest image correlation coefficient reaching 0.9629 and the relative error of the image reduced to 0.0530.