Analysis Of Scattering Characteristics Of Vehicle Radar Images Based On Complex Convolutional Network

As a key task in the field of SAR ATR,the analysis of scattering characteristics can simultaneously obtain the local geometric structure and electromagnetic scattering characteristics of the target,with the help of the amplitude and the phase in complex-valued images.Currently,scattering characteristics are often extracted via the attributed scattering center(ASC)model,which has the definite physical interpretability and the capability in deriving local scattering structure directly.However,the performance of the ASC model is usually affected by data preprocessing and the initial value of its hyperparameters.Furthermore,the complicated tuning steps of multivariable optimization are also time-consuming.To tackle with these problems,this article,via the MSTAR dataset,developed a framework for scattering characteristics analysis with the complex-valued neural network(CVNN).Major works of this thesis are as follows:(1)This paper developed an end-to-end robust CVNN framework by extending the autoencoder to the complex-valued domain,learning the robust representation of the scattering characteristics from the vehicle images.(2)The multi-scale structure and the residual structure were introduced to the CVNN,obtaining representations of the large distributed scattering structures and the small isolated scattering centers.(3)In order to suppress the clutter background,we extended the attention mechanism to the complex-valued domain,focusing on the targets’ structures with stronger backscattering.An improved reconstruction loss function was also proposed,in which the important scattering centers were given priority to improve the classification performance and accelerate network convergence.(4)A prediction module that directly obtained the location of the attributed scattering centers was built to replace the decoder in the proposed complex-valued autoencoder framework.Accordingly,the interpretability of the complex-valued model would be enhanced and the correlation between the latent variable features and the physical scattering structure of the target was established.Experimental results with the MSTAR dataset show that,the proposed model can directly learn robust representations of the scattering characteristics from vehicle patches,and has a comparative classification accuracy of 97.79% to that of those mainstream supervise models.In a time elapse test,it takes 173.4423 seconds for the traditional ASC model to process an image via a CPU,while only 6.971412 seconds for the proposed model to process 587 images,which can be shortened to 0.561102 seconds with a GPU.The multi-scale features extracted from the proposed model not only achieve good recognition performance,but also provide the strong interpretability in scattering characteristics.The proposed model has a simple structure and a lower operating cost,which provides a new idea for the development of SAR-ATR.