Research On SAR ATR Based On HRRP Time-frequency Domain Sparse Representation

Synthetic aperture radar (Synthetic Aperture Radar, SAR) is a new microwaveimaging radar with a huge virtual antenna radar formed by flight platform movement. Ithas a large swath data acquisition capability with high resolution in all-weather, all time.It has been widely used in the national economy and national defense domain. And, ithas become a high resolution earth observation sensor with fast development in manycountries of the world. In many SAR data application domains, such as spacereconnaissance and information access, marine target surveillance and battlefieldprecision strike assessment, SAR automatic target recognition (Automatic TargetRecognition, ATR) is one of the most critical core technology.This paper introduces the current research situation of SAR ATR firstly, pointingout challenges and difficulties in the research area. One of the challenges is the poorSAR image quality, such as blurry image and defocused image due to non-cooperativetarget motion or low signal-to-noise ratio (Signal-to-Noise Ratio, SNR). Thosedefocused or distortion images make the SAR target features extraction and recognitionbecome much more difficult in the image domain. However, SAR ATR based on targethigh resolution range profile (High Resolution Range Profile, HRRP) can overcomethese difficulties. In time-frequency domain, SAR target scattering features can berevealed and extracted effectively in spite of the complex electromagnetic scatteringmechanism. Several time-frequency analysis methods are studied in this paper. It isfound that Gabor atom matching pursuit (Matching Pursuit, MP) time-frequencyanalysis method can represent various scattering characteristics of HRRP adaptively.The MP method has advantages of high time-frequency resolution, no cross-terminterference and noise suppression. It is found that target scattering features in the MPtime-frequency map of HRRP is sparse and localized.In the light of the sparse distribution of HRRP scattering features intime-frequency map, we study to model and characterize the HRRP time-frequency mapusing the sparse representation theory in the paper. The principle of ATR using sparserepresentation and dictionary learning algorithms is described. The fixed dictionarylearning and adaptive dictionary learning algorithms are analyzed in detail. Especially,the label consistent K-SVD (Label Consistent K-SVD, LC-KSVD) algorithm is studiedin depth, which is an effective dictionary learning algorithm for classification. On the basis of the above, we propose a novel method for SAR ATR based on HRRPtime-frequency map LC-KSVD. Finally, ATR experiments are performed using publicSAR datasets released by moving and stationary target acquisition and recognition(Moving and Stationary Target Acquisition and Recognition, MSTAR) programsponsored by Defense Advanced Research Projects Agency (Defense AdvancedResearch Projects Agency, DARPA) of USA. The experimental results show that theproposed method has a good classification performance, and the method is superior tothose methods using two-dimensional SAR image in the case of low SNR or defocusedimage.