DiffusionEMIS: Diffusion Model for 3-D Electromagnetic Inverse Scattering

We propose DiffusionEMIS, a new paradigm that formulates the 3-D electromagnetic inverse scattering (EMIS) as a denoising diffusion process from a 4-D noise distribution to the specific 3-D scatterer point cloud with reasonable electrical parameters at each point under the scattering field as a condition to enrich the research of 3-D EMIS community while avoiding the computational burden of 3-D meshes. During the training process, scatterer point clouds diffuse from ground-truth shapes to random distributions, and the model learns to reverse this noising process. During the inference process, the model progressively refines the randomly generated shape to the output result. We design a certain 3-D scattering system to construct two datasets: 3DEMIS-Mixed National Institute of Standards and Technology datab (MNIST), where the samples are nonuniform 3-D handwritten digits, and 3DEMIS-ShapeNet: An information-rich 3-D model repository (SHAPENET), where the samples are uniform 3-D objects. The extensive evaluations on 3DEMIS-MNIST and 3DEMIS-SHAPENET show that our DiffusionEMIS achieves favorable performance and great stability which means 3-D EMIS can be solved by a generative way. Our model outperforms previous common EMIS methods, such as the born iterative method (BIM) by a large margin and has extremely robust noise resistance. To evaluate the generalization performance of DiffusionEMIS, we construct the dataset 3DEMIS-English Mixed National Institute of Standards and Technology database (EMNIST) consisting of nonuniform 3-D handwritten letters for testing the model trained on 3DEMIS-MNIST, the test results indicate a great consistency with the ground truths. We also conducted tests on real data to further demonstrate the effectiveness of our method.