Research On SAR Images Simulation Method Based On Neural Network

Since the advent of synthetic aperture radar(SAR),the research on SAR images has been deeply concerned by relevant scholars in the electromagnetic field.Limited by the acquisition method of real SAR images,electromagnetic simulation is generally used to obtain a large number of simulated SAR images for research use.But the traditional electromagnetic simulation method often contains two deficiencies.First,the algorithm is difficult to describe the various uncertain randomness in the environment and the approximate processing that may exist in the solution process,which will lead to the low similarity and large feature differences between simulated SAR images and measured data.The second one is the electromagnetic simulation requires a large amount of time and computing resources for each calculation and simulation.The above two deficiencies greatly limit the further application of simulated SAR images.In recent years,with the continuous breakthroughs in the field of artificial intelligence,neural network technology has also developed rapidly.The trained neural network has good effect and high efficiency when extracting and processing data features,therefore,it is widely used in the field of image processing and simulation generation.Based on its application characteristics,neural networks is used to carry out relevant research on the above two deficiencies between simulated and measured SAR images.The main contents of the work in this paper are summarized as follows:(1)The characteristic differences between simulated SAR images and measurd SAR images are systematically compared and analyzed in this paper.At the same time,the basic structure of the nearal network and the training algorithm are also briefly derived and demonstrated in this paper,which lays the foundation of the algorithm models for the work in this paper.(2)In order to solve the problem of large characteristic difference between simulated data and measured data,the Cycle GAN is adopted in this paper to perform feature processing on simulated data.The Cycle GAN model takes sinulated data as its input,and after feature processing the input data,it can output SAR images with high fidelity.Relevant experimental results show that after the Cycle GAN's processing,the recognition accuracy of simulated data can be increased from 29.89% to 80.33%,which greatly improves the similarity between simulated data and measured data.(3)In order to solve the problem that electromagnetic simulation consumes a lot of calculation and time resources,WGAN-gp model and LSGAN model are used in this paper to directly generate SAR images.The final experimental results show that these two network models can shorten the simulation time to about 11% of the shooting and ouncing ray method,which greatly improves the simulation efficiency of SAR images.In summary,based on the practical problems existing in electromagnetic simulation SAR images,different neural network models are adopted to solve the corresponding problems in this paper,which has great theoretical research and practical application significance.