Coarse-to-Fine Network Based Interference Mitigation For Micro-Doppler Spectrogram

In recent years,deep learning technology has become more and more important in radar field.Generally,neural network has better processing capabilities for image data.In practice of technology,the radar signal with motion information is usually converted into the micro-Doppler spectrogram.Then spectrogram is put into the network for feature extraction and learning.Therefore,the quality of the spectrogram directly affects the effect of network.For example,a spectrogram with interference may cause the network to extract wrong features and mislead training.However,during the actual experiment,there is unpredictable and unavoidable interferences in the radar signal limited by environment and equipment.Traditional signal-based denoising methods often have little effect.The interference mitigation for micro-Doppler spectrogram is a very challenging task.In response to the above problems,this paper mainly focuses on the problem of interference mitigation combined with neural network algorithms for micro-Doppler spectrogram.The main innovations and work of this dissertation are as follows:Firstly,interference mitigation problems can be understood as image inpainting problems.Therefore,we mask the interference part and improve the coarse-to-fine network for restoration.As human movement captured by the radar has a certain periodicity,the network can use the surrounding pixels to generate and predict combined with the attention layer.At the same time advanced discriminator and loss function in the generative network are used to supervise training.In addition,the pre-training results of the ImageNet data set on the coarse-to-fine networks are used to initialize parameters in advance to improve the convergence speed of the network.2.Since coarse-to-fine network input requires masking the interference,this paper uses the Fast R-CNN network to automatically locate the interference waveform from the perspective of engineering practice.This model has carried out a number of tests on the simulation data set of motion capture,and turned out the method outperforms the traditional methods.At the same time,this method can also extract certain features from the interference waveform without adding a mask.and results verify the feasibility and effectiveness when the method faces various interferences.