Research On Automatic Target Recognition For MMW Detection System With Deep Learning

As that the millimeter wave detection system could be operated under all time and all-weather conditions,and has the advantage over other systems in anti-noise and ranging coverage,it plays an important role in target detection and has been widely applied in military and civil area,such as short-range detection,accurate guidance,remote sensing and security check.Automatic target recognition(ATR)is one of the key technologies in the application of millimeter-wave detection system.However,due to the influence of factors,such as organization,parameters,noise interference,posture and aspect of targets,there are still many theoretical and technical problems in automatic target recognition for millimeter wave detection system.Thus,with the objective of providing a general theoretical guideline for building high-level feature representations for millimeter wave detection system,some research is studied based on deep learning theory in this dissertation.The main research work includes:(1)Deep discriminatory dictionary learning for MMW Radar HRRP based automatic ground target recognition.Taking the aspect-sensitivity and high correlation between HRRP in adjacent azimuth into consideration,we proposed a discriminatory dictionary learning algorithm based on aspect character.With utilizing constraints of correlation between atomic and aspect character,the proposed method can effectively improve the reconstruction accuracy of the HRRP samples.By taking the advantage of deep learning theory,we designed a stacked dictionary learning model based on the discriminatory dictionary,which can further improve the accuracy of MMW Radar HRRP based target recognition and strengthen the noise immunity and robustness,especial when the samples is limited and the SNR is low.(2)Deep denoising contractive auto-encoder for MMW Radar HRRP based automatic ground target recognition.Due to the fact that the aspect-sensitivity and ground clutter can greatly decrease the performance of conventional recognition algorithms,an improved denoising contractive auto-encoder is proposed by utilizing the constraints of contractility and reconstruction.The ability to establish robust feature representation is enhanced greatly by transforming the traditional artificial recognition framework into a data-driven autonomous learning framework.An end-to-end deep learning model is further designed based on denoising contractive auto-encoder,which can effectively enhance the ability of building robust feature representations from MMW Radar HRRP samples,especial in the condition of small SNR and small samples.(3)Deep collaborative auto-encoder for MMW Radar HRRP based automatic ground target recognition.By utilizing the adjacent aspect character to compensate the influence of aspect-sensitivity,an improved collaborative auto-encoder is proposed with the combination of collaborative filter(CF)theory,and a deep learning framework is designed based on collaborative auto-encoder.The ability of building high-level discriminatory feature representations is effectively enhanced with the aid of inner-class HRRP samples,and the recognition accuracy for MMW Radar HRRP in small SNR is greatly improved.(4)A feature fusion algorithm for MMW SAR based automatic target recognition.Due to the fact that the information of traditional SAR image is not sufficient for accurate target recognition,a complex-value domain deep denoising convolutional neural network is proposed to utilize the coherency matrix to predict the label of target.With the combination of convolutional layer and deconvolutional layer,the raw coherency matrix is projected into the high-level feature space,and the robust feature representation embedded in the raw data is constructed.By taking the advantage of boundary extraction of superpixel representation,we further designed a fusion model composed of denoising convolutional neural network and superpixel representation,which utilizes boundary information to correct the singular pixels and misclassified boundary pixels.The proposed can greatly improve the target recognition accuracy and enhance and robustness.(5)Deep denoising convolutional neural network for MMW In SAR based automatic target recognition.Due to the fact that the MMW In SAR imaging is seriously impact by the noise interference and information loss,an end-to-end deep denoising convolutional neural network is proposed by combining the denoising auto-encoder and convolutional neural network,which is designed to construct high-level feature representations for target image and generate target label.Via analyzing the mechanisms and the imaging principles of MMW In SAR,we further proposed a visibility matrix based deep denoising convolutional neural network for MMW In SAR automatic target recognition.Without information loss in the inverse imaging process and with the utilization of phase information,the proposed method can construct more discriminatory feature representations;generate higher recognition accuracy,especial in small SNR and small samples condition.