融合谱归一化与残差机制的WGAN-GP逆散射成像算法OA

To address the problems of strong nonlinear ill-posed mapping and fidelity loss in electromagnetic Inverse Scattering Problem(ISP),an improved Wasserstein Generative Adversarial Network with Gradient Penalty(WGAN-GP)integrating spectral normalization and residual mechanisms is proposed.By introducing the Wasserstein distance and combining Spectral Normalization(SN)with Gradient Penalty(GP)to strictly constrain the Lipschitz continuity of the discriminator,the algorithm resolves the mode collapse problem inherent in traditional Generative Adversarial Networks(GAN).In terms of architectural design,a parallel dual-path residual ConvBlock generator is constructed and integrated with spatial positional encoding to enhance the model's feature extraction capability for the fine topology and inhomogeneous field quantities of scatterers.Experimental results demonstrate that the algorithm achieves stable convergence,with the average relative L2 error on the test set reduced to 0.005 3,achieving an order-of-magnitude improvement compared with the standard GAN.Compared with the baseline UNet,the proposed algorithm effectively overcomes the mean smoothing effect and exhibits significant advantages in topological fidelity,providing an efficient solution for high-fidelity electromagnetic imaging.