Key points are not available for this paper at this time.
Frequency estimation is a pivotal process in many signal-processing applications. Generating radar range profiles for linear frequency modulated radar systems is such a case where spectral analysis is used to estimate target ranges. Conventional methods like fast Fourier transform (FFT) are the golden standard in frequency estimation, despite its Rayleigh resolution limit and high sidelobe levels. To address such limitations this paper introduces HRFreqNet; a deep neural network (DNN) architecture for high-resolution frequency estimation from 1D complex time domain data consisting of multiple frequency components. Our deep learning (DL) architecture consists of an auto-encoder block to improve signal-to-noise ratio (SNR), a frequency estimation block to learn frequency transformations to generate pseudo frequency representations(FR), and finally, a 1D-UNET block to reconstruct high-resolution FR. Experimental results on synthetically generated data show enhanced performance in terms of resolution, estimation accuracy, and ability to suppress noise. Achieved range profiles are also sparser with lower sidelobe levels. The proposed HRFreqNet is evaluated over both synthetic and experimental real-world radar data and it is observed that accurate, sparse, high-resolution range profiles are obtained compared to existing approaches.
Biswas et al. (Mon,) studied this question.