SAR Image Classsification And Recognition Based On Compressive Sensing

Synthetic Aperture Radar (SAR) is one of equipments with all-day work. it hasadvantages of high resolution and strong penetrating ability. The technologies of targetrecognition and terrain classification for SAR play an important role in national defense.However, As an ultra-wideband application, SAR is limited by Nyquist sample theory,and need too high sample rate and too large data volume. Compressive sensing theorytakes the analog signal into the digital signal with a method different from Nyquistsample theory, it indicates that sparse or compressible signals can be recovered withoverwhelming probability by less measurements than what Nyquist requires.This paper does some attempting study on compressive sensing RADAR systemapplication, compressive sensing theory is applied to SAR target recognition field andSAR terrain classification field. The main work includes the following three aspects:(1) A algorithm based on orthogonal neighborhood preserving projections andsparse representation for SAR image recognition is presented. It utilizes the fact that thedata in MSTAR database lies in the manifold and adopts orthogonal neighborhoodpreserving projections to feature extraction of target image and then places theframework of sparse representation for the final classification. The experiment resultshows a good performance.(2) A algorithm based on random measurements and orthogonal neighborhoodpreserving projections for SAR image recognition is presented. This paper does someattempting study on compressive sensing RADAR system application, with purpose ofclassification on measurements of different types of targets directly. Throughconstructing a dictionary to describe the characters on every class exactly, thosemeasurements for testing can be represented sparsely under this dictionary. After that,those measurements are recognized by a sparse representation classifier, and theexperiment shows a good performance for classification.(3) A algorithm based on two dimensional random measurements and supportvector machines for SAR terrain classification is proposed. Linear kernel SVM’sclassifier in the measurement domain has true accuracy close to the accuracy of the bestlinear threshold classifier in the data domain. so after one dimensional randommeasurements, a method of wavelet packet decomposition is adopted to extract features.Support vector machines is used for terrain classification. The classification result is not very ideal in the low sampling rate because of the limitation of one dimensional randommeasurements. And two dimensional random measurements takes advantages ofkeeping the image structure information and low computing complexity. So it solvesupper problem and this algorithm maintains a very high recognition rate.This work was supported by the National Natural Science Foundation of China(No.60971128); Huawei innovation research project (No. IRP-2011-03-04).