Research On Data Augmentation Model For Human Motion Micro-Doppler Images

In recent years,our country has begun to become the deep aging society slowly.The issue of the elderly has gradually become an important social problem.In the aspect of the elderly monitoring,radar detection has great advantages over other sensors.However,the microDoppler images generated by radar are easily affected by other static objects or electromagnetic signals.So it is necessary to establish a large-scale quality evaluation database to screen the quality of spectrograms.Nevertheless,the process of subjective evaluation of the spectrograms is complicated and time-consuming,and it is difficult to satisfy the data quantity requirements of deep learning.Therefore,in order to solve this problem,the data augmentation model for human motion micro-Doppler images is deeply studied in this thesis.The main contents include:(1)A Human Motion Micro-Doppler Feature Attention(HMMDFA)mechanism was proposed based on the characteristics of human motion micro-Doppler and the rules of human motion in this thesis.This attention mechanism extracts the channel feature,Doppler frequency feature of human movement,energy feature of human movement and circulation feature of human movement by four different submodules.From the perspective of information extraction,HMMDFA can not only extract the channel and spatial information of the image,but also extract the phase information of the image when compared with the existing attention mechanism.In order to prove the performance of HMMDFA,several experiments were carried out in this thesis.(2)A HMMDFA-EDIT model was proposed by combining HMMDFA with the existing Exemplar-Domain aware Image-to-image Translator(EDIT)model in this thesis.Subsequently,the architecture parameters of HMMDFA was optimized and adjusted to further optimize the HMMDFA-EDIT model.The image generation quality of the optimized model was further improved.In order to verify the data augmentation performance of this model,a comparison experiment is conducted with the existing data augmentation models.The experimental results demonstrate that the proposed HMMDFA outperforms the existing attention mechanisms(Squeeze-and-Excitation attention and Coordinate attention)in both data augmentation tasks and image quality classification tasks.Not only the generated image quality is improved,but also the model parameter and computational complexity of the optimized HMMDFA-EDIT is not significantly increased compared with that of the original network EDIT.Compared with other data augmentation models,the generated image quality also achieved the best performance.