基于VAE生成网络的EMT金属缺陷数据集的扩展与验证  

Extension and verification of EMT metal defect data set based on VAE generation network

作　　者：王浩哲 王烨辰 张亚茹 张辰轩 WANG Haozhe;WANG Yechen;ZHANG Yaru;ZHANG Chenxuan(Institute of Electrical Tomography of Tianjin Polytechnic University,Tianjin 300000,China)

机构地区：[1]天津工业大学电学层析成像研究所,天津300000

出　　处：《电子设计工程》2023年第11期25-29,共5页Electronic Design Engineering

基　　金：国家级大学生创新创业训练计划项目(202110058028)。

摘　　要：新兴的图像处理神经网络需要大量优质的训练数据训练才能发挥良好的效果,为了解决电磁层析成像(ElectroMagnetic Tomography,EMT)技术生成图像质量不高且不便于大量获取的问题,使用COMSOL Multiphysics软件仿真,建立EMT系统模型生成缺陷的真值图像,并利用共轭梯度算法生成重建图像与其对应,组合输入到变分自编码器(Variational Auto-Encoder,VAE)生成网络中,进行数据增强制成大量数据集,并对VAE生成图像进行结构相似性比对。实验结果表明,VAE生成图像的结构相似度(Structural Similarity,SSIM)与利用共轭梯度算法重建图像的结构相似度非常接近,相对误差控制在10%以内,证明了VAE生成网络应用于扩展神经网络数据集的合理性。The emerging image processing neural network needs a large number of high⁃quality training data to achieve good results.In order to solve the problem that the image quality of ElectroMagnetic Tomography(EMT)is not high and it is not easy to obtain a large number of images,COMSOL Multiphysics software is used to simulate and establish the EMT system model to generate the truth image of the defect and use the conjugate gradient algorithm to generate the reconstructed image corresponding to it.Combined data is put into the Variational Auto⁃Encoder(VAE)generation network for data enhancement to produce a large number of data sets,and compare the structural similarity of the images generated by VAE.The experimental results show that the Structural Similarity(SSIM)of the VAE generated image is very close to that of the conjugate gradient reconstructed image,and the relative error is controlled within 10%,which proves the rationality of the application of VAE generation network to the extended neural network data set.

关 键 词：EMT COMSOL仿真 共轭梯度算法 VAE生成网络 

分 类 号：TN99[电子电信—信号与信息处理]