Research On ISAR-based Fast Imaging And Recognition For Maneuvering Targets

Inverse synthetic aperture radar(ISAR)imaging and recognition plays a vital role in military and civilian fields such as space warning,and environmental monitoring.In real-world scenarios,due to the non-cooperative maneuverity of the target and the complexity of the electromagnetic environment,the difficulty of balancing the imaging quality and the computational cost is aggravated,and the limitation of traditional recognition methods are more prominent.It is a bottleneck to meet the high-accuracy and real-time performances simultaneously.To tackle this problem,it is necessary to conduct more depth study on fast and efficient signal processing,and further explore the utilization of multi-dimensional information of the target.Aiming at this issue,we conduct key technical research on ISAR imaging and recognition of maneuvering targets under complex conditions,with the expectation of improving the precision and real-time performance,simultaneously.We first improve the feature quality implied in the pre-input information of recognition by optimizing the accuracy and complexity of the imaging processing algorithms;then reduce the strong dependence of imaging quality and artificial feature engineering in recognition processing by using deep learning to mine multi-dimensional feature information.The thesis mainly includes the following parts of research content.(1)ISAR system modeling,general algorithm study and optimization design.Based on the classic 2-D turntable model and the synthetic aperture principle,we first analyze the echo signal and its diversity characteristics,and study the characteristics of range profiles(RP)and cross-range profiles to establish the theoretical foundation for subsequent research.Then we propose the idea of achieving high quality and near-real-time imaging and fast recognition by improving compensation precision to solve the image blurring caused by translational movement,and enhancing the profiles' aggregation to improve the image defocus caused by azimuth movement.(2)Study on parameter estimation and motion compensation based on RP correlation.First,we improve the precision and complexity of general parameter estimation method by utilizing the correlation between RPs.Second,specific to radial uniformly moving target,we improve range alignment precision by employing fractional Fourier transform(Fr FT)to generate high resolution range profile(HRRP),and perform complexity optimization on searching the Fr FT matched-order.Furthermore,specific to radial non-uniformly moving target,high-order parameter estimation method HSACM is developed by modeling the parameter estimation as least square(LS)problem.Experimental analysis verified that,the proposed high-order parameter estimation method significantly outperforms other methods both on precision and computational complexity.(3)Study on high-order signal processing and fast imaging of maneuvering target.Aiming at the inherent shortcomings of Fr FT on compressing high-order range echoes and the time-varying Doppler caused by non-uniformly rotation,a new low-complexity signal processing technique called power-weighted Fourier transform(PWFT),is proposed.Then by utilizing its spectral sharpening characteristic on radar signal,high-order translational compensation and azimuth focusing based on PWFT is studied,based on which,the fast imaging algorithm PWFT-RD is proposed.Furthermore,an irregular window function is designed to reduce the trailing ghost in maneuvering target imaging.Simulations and experiments verified the effectiveness of PWFT in improving the imaging focus performance,and the PWFT-RD algorithm can achieve high quality and near-real-time imaging of maneuvering target.(4)Study on recognition network based on deep learning of multi-dimensional features.Specific to three typical representations of ISAR echo data: raw data,1-D RP,and 2-D ISAR image,analyze their enhancing or suppressing effects on the original multi-dimension features.Then an end-to-end recognition network based on the convolutional neural network is designed,with the aim of exploring multi-dimension information of the target by utilizing deep learning.Based on the framework,three single representation-driven recognition networks are designed and analyzed.Then faced to different scenario requirements,multi-stream fusion networks driven by multi-dimension representations are designed by implementing data enhancement and structure optimization.Comparing experiments validate the effectiveness of the proposed networks,and the proposed multi-stream fusion recognition network fully utilized the complementary advantages of multi-presentations of the samples,which outperform bettter than existing recognition schemes in specific scenarios such as high target maneuverability and extremely low signal-to-noise ratio.