Research On Enhanced Imaging And Target Recognition Based On Terahertz Radar And Deep Learning

Different from microwave wave and infrared wave,terahertz(Thz)wave possesses distinctive scattering properties and plays an important role in communication,detection,early warning and so on,which become a new strategic commanding point of military science and technology competition.Compared with microwave radar,terahertz radar has a significant advantage in the field of radar imaging due to its higher bandwidth,higher carrier frequency and higher spatial resolution.To be specific,the imaging results in the terahertz band can obtain more abundant target information,such as the contour,structure and material properties,which is conducive to the target recognition and image interpretation.However,in the terahertz band,traditional radar imaging models and imaging methods also face some new problems and challenges.For example,rough surface leads to a sharp increase in the number of scattered centers from the target,which is difficult to describe this scattering characteristics by point scattering model and can not reconstruct the extended structure completely.In the traditional radar imaging method,the image quality in the terahertz band is degraded due to the sidelobe of multiple scattering centers.Under this background,the paper focus on enhanced imaging and target recognition with terahertz radar.In terms of enhanced imaging,this paper focus on image quality improvement and target structure inversion.In target recognition,this paper studies the target recognition method based on feature fusion.In terms of image quality improvement,the research object is chosen as 3D image quality improvement.Aiming at the disadvantages on the time needed for enhanced imaging and computational burden of sparse regularization and compress sensing,this paper focuses on the fast way to promote the image quality,and puts forward the image quality improvement based on lightweight network.The basic framework are as follows:the original imaging results firstly are established within the given space and known imaging geometry.Then the expected imaging results are reconstructed according to the prior knowledge of imaging.Finally,limited to the training dataset and network parameters,an end-to-end 3D radar enhanced imaging network is built.Numerical results and electromagnetic simulation verify the effectiveness of the proposed method.Compared with the sparse regularization,the enhanced imaging time is improved by two orders of magnitude,and real-time enhanced imaging can be realized.In terms of image quality,the optimal image quality is achieved compared with the sparse regularization and the existing methods based on deep learning.In terms of the target structure inversion,the research object is chosen as the reconstruction of the extended objects.Considering these shortcomings on these strict imaging conditions of the methods based on large rotating angle and multiple views and the poor performance of complex target structure,this paper focuses on the detection and reconstruction of barrel.These traditional methods,which relies on recognition and matching of phase features of different scattering centers,are transformed into the target detection based on tank shape priori and transfer learning.The basic framework is as follows: the paper firstly analyzes the imaging characteristics of the extended object and the typical tank,and get the reason why it disappear and confirm corresponding the location in image domain.Then,based on the these prior knowledge of tank imaging,the multi-scale detection network is established.Finally,two corresponding reconstruction methods are determined according to detection results.The electromagnetic simulation and experimental results verify that the proposed method can achieve fast and accurate detection and reconstruction.Four typical target detection networks,namely Faster RCNN,Yolo V3,Retina Net and Center Net,are compared to achieve the best detection performance under the same conditions.In terms of target recognition based on feature fusion,the research object is chosen as the high resolution range proile(HRRP).The stability of target recognition system usually face two situations: firstly,the scattering characteristics of complex targets along the change of observation angle is large,which directly affects the feature distribution of HRRP.Secondly,the environment of the recognized systems varies greatly,and target recognition under low SNR measures the accuracy and stability.Therefore,this paper focuses on the stable target recognition under low SNR,and proposes a multi-view feature fusion method based on denoising feature enhancement.The basic framework is as follows: the denoised network is firstly designed to reduce noise and improve signal to noise ratio.Then the feature from multiple views is fused to get higher order features.Finally,these higher order features are utilized for target recognition.It should be noted that the experimental data are originated from electromagnetic computational simulations of eight different aircraft targets.Experimental results show that feature fusion can achieve more stable and higher recognition accuracy by multi-view feature fusion.When the signal noise ratio(SNR)is lower than 0d B,the proposed method can improve the recognition accuracy by about 5% compared with the traditional recognition method.To conclude,this paper provides some new ideas and methods for some existing problems of radar enhanced imaging and target structure inversion.For target recognition based on feature fusion,a new framework of feature fusion is provided,which is a new solution of efficient and stable recognition system in practical work.