基于改进CycleGAN网络的电磁层析成像算法OA

In order to solve the problem that reconstructed images in electromagnetic tomography(EMT)are prone to arte-facts due to the highly nonlinear and ill-posed of the inverse problem,a deep learning structure namely cycle generative adversarial network with attention mechanism(CycleGAN-AM)is proposed for EMT image recon-struction.The network consists of two GAN networks,which can capture more nonlinear features through dual generators and dual discriminators.The use of global attention mechanism(GAM)in the generator to learn chan-nel dependencies improves the accuracy and interpretability of CycleGAN-AM.The performance of the algorithm proposed in this paper is evaluated by simulation and metal detection experiments.The imaging results show that CycleGAN-AM is able to accurately recover the boundaries of the objects under test and can effectively recon-struct the objects under new conductivity distributions(objects of varying sizes/numbers)and noise interference.Compared with traditional intelligent learning methods,CycleGAN-AM can improve the imaging accuracy by more than 10%.