Physical Model-Deep Network Hybrid Radar Imaging And Related In-Wall Anomalous Medium Detection

The wall penetrating radar technique has been widely used in in-wall target and anomalous medium detection,localization,identification and imaging since the electromagnetic waves has the characteristics that it can penetrate non-conducting medium.Thus,it has been widely used in security check,building construction and military and it has become the hot spot of current research.However,the attenuation and dispersion of electromagnetic waves by the wall will greatly weaken the intensity of the target echo signal,which brings great challenges to the detection and imaging of the target in the wall.Moreover,the refraction and reflection of electromagnetic waves on the surface and inside of the wall will cause multipath ghost images in the reconstructed image,which lead to image blurred.Thus,based on a full understanding of the theory of the in-wall detection radar and the basic theory of radar imaging,this paper proposes the following research contents for the imaging of anomalous media in the wall combining with the physical model of the radar signal and the deep network.Firstly,this paper introduces the signal physical model and scenario of the in-wall detection radar and elaborate the relationship among the radar signal,the wall and the target from the perspective of the radar signal composition.In addition,the propagation path of the electromagnetic waves from the antenna to the target and the calculation method of propagation delay are analyzed,and the back projection method,low-rank sparse representation and compressed sensing are used to image the target in the wall.The latter two methods obtain the optimal parameter values by optimizing the objective error function,and then provide theoretical guidance for the subsequent imaging network design.Secondly,the detection and imaging of anomalous media in walls combined with deep learning networks are investigated.Based on the theoretical basis of compressed sensing,a deep learning imaging network driven by physical model is proposed.First,the learned convolutional neural network prior terms are used to replace the regularization terms in compressed sensing to extract high-dimensional characteristics of the radar echo signal.Second,the physical model of the radar signal is integrated into the data consistency item to ensure the consistency of the algorithm with the measured value.Finally,the compressed sensing iterative solution framework is unrolled to yield a deep learning imaging network,and the performance of target imaging is analyzed based on gprMax simulation data.The advantage of this method is that it overcomes the limitations of traditional compressed sensing for different scenarios and the shortcomings of target detail reconstruction.Third,a hand-held wall detection radar imaging system and a large-wall target imaging system are designed for the scene of wall detection.The handheld imaging system extends the two-dimensional back projection to three-dimensional space,and uses the relationship between the focus of a special ellipsoid and the chord length to match the propagation delay.This method can effectively reduce the matching time of propagation delay.In the large-wall imaging system,this paper proposes a large-wall imaging network STW-Net based on full convolution neural network.The network extracts the redundant information of each adjacent sub image of the wall,and then completes the two-dimensional imaging of the wall.Finally,the standard experimental environment of large-wall imaging system is built to verify the effectiveness and reliability of target imaging based on the measured data.Experimental results demonstrate that the STW-Net has higher reconstruction accuracy and detail restoration degree comparing with the traditional image schemes.