Research On Machine Learning Algorithms For Solving Electromagnetic Inverse Scattering Problems

Electromagnetic scattering and inverse scattering problems have a wide range of engineering applications in microwave remote sensing and biomedical fields.There are inherent nonlinear relationships between the electromagnetic parameters of the target and the measured scattering field,besides,the equations for solving inverse scattering problems are ill-posedness,especially in the strong scatterer inversion tasks.Some optimization methods and iterative methods are usually applied to figure out the inverse scattering problems at present.But some limits exist in these methods.Noise has a bad effect on the inversion results.The approximation methods perform well only in weak scatterer inversion tasks,low resolutions appear when the contrasts of the scatterers are high or their electrical sizes are large.The iterative methods also suffer some difficulties such as high computational complexity and not suitable for the real-time reconstruction.For the purpose of improving the inversion accuracy,ensuring the real-time inversion and enlarge the application domain of the algorithm,this article adapt deep learning algorithms to fit nonlinear relationships in inverse scattering problems.The study of the algorithms contains the following parts mainly:1.Solving ISPs based on supervised learning networkThis part adopts a machine learning model with high resolution,simple logical and fast computing speed to solve 2-D electromagnetic inverse scattering problems.The network includes a feature extraction module,an imaging enhancing module and a recognition module,the electromagnetic parameters of the target can be reconstructed by the scattering field data after training.Experimental results show that the method is fast in calculation and with higher resolution in reconstructing strong scatterers.2.Solving ISPs based on BP approximation + supervised learning networkBased on the calculated Green's operators,the Back Propagation(BP)algorithm assume the induced current is proportional to the scattering field.Generating the preimaging results by BP approximation can effectively reduce the number of physics parameters,moreover,release the burden of the neural network.Besides,we can also reduce the complexity of network as the result of less parameters needed to be fitted.3.Solving ISPs based on Deep Q Network(DQN)The DQN learn and formulate the strategy to select and combine some smallscaled convolutional neural networks for different inversion situations.We design a serious of small-scaled but specialized CNN modules for solving ISPs,and use different modules collaboratively according to the learned strategy.In comparison with the existed deep learning algorithms for solving ISPs,the proposed method is promising in solving inverse scattering quantitative imaging problems in real time.Moreover,the dynamical CNN chain has better performance and wider application areas in solving ISPs.