A Method of Improving the Quality of SAR Target Electromagnetic Simulation Image Based on Scattering Feature Enhancement

‍ ‍At this stage， the deep learning algorithm usually faces the situation of partial sample missing of the measured data when recognizing the synthetic aperture radar （SAR） target. The use of electromagnetic simulation data for auxiliary recognition is one of the effective ways. However， there are inevitable differences between the simulated and measured data. The existing methods for improving the quality of the simulated image pay more attention to the similarity of the overall style between the simulated and measured data， and ignore the importance of the target scattering characteristics for recognition. To solve these problems， this paper proposed a method to improve the quality of SAR target electromagnetic simulation image based on scattering feature enhancement. This method improved the loss function under the framework of cycle generation adversarial networks （CycleGAN）. On the one hand， the least squares loss function was used to replace the cross entropy loss function to avoid the gradient disappearing and realize the iterative optimization of the target texture structure features； On the other hand， the MS-SSIM-L1 loss function was introduced to better retain the detail information and structure outline of the generated image， maintain the consistency of the overall structure of the target， and effectively avoid over-learning of the model. Based on four types of vehicle target simulation data set and MSTAR measured data set， by using the target contour， shadow contour and target intensity feature similarity index， it was verified that the method in this paper enhanced the scattering features such as target detail texture and structure contour. On this basis， the target classification and recognition experiment was carried out in combination with A-ConvNets network. Compared with original CycleGAN method， method in this paper improved the recognition accuracy under different sample missing conditions. Through feature visualization， it was shown that the generated image was closer to the target feature distribution of the measured image， which verifies the effectiveness of the method in this paper.