Research On SAR Image Target Recognition Via Compressed Sensing

Synthetic Aperture Radar (SAR) is an active microwave sensor, which can obtain twodimensional high resolution SAR image and provide a powerful support to earth observation andmilitary reconnaissance. SAR image automatic target recognition (ATR) is one of major source ofbattlefield information attracting scholars of domestic and foreign. Compressed Sensing (CS) is a newkind of signal processing theory and show a great performance in target recognition with quiet fewobservations. Based on the features of SAR image automatic target recognition and CompressedSensing theory, this thesis focuses on the combination of SAR target recognition and CompressedSensing. The main work is summarized as follows:1. The first chapter studied the basic methods of sparse representation: fixed orthogonal basis,unions of basis and over-complete dictionary. And then, the development as well as advantages anddisadvantages of these bases are analyzed. Finally, it compared the sparsity and reconstruction ofimages based fourier transform、discrete cosine transform and discrete wavelet transform, andsummed up the importance of signal sparse representation.2. Research on SAR image target recognition based on Shift Invariant Shearlet Transform (SIST).Firstly, SIST was applied to decompose SAR images into low frequency sub-band and directionalhigh frequency sub-bands. Then, fused low frequency and high frequency sub-band which obtainedby measurement function as target feature information, thus we obtain a better description for images.Simultaneously, by processing shearlet coefficients with threshold can denoise speckle noise in SARimage efficiently. Finally, simulation results based MSTAR database demonstrate that: combininglow-frequency and high-frequency component get a high level of recognition rate, proves thefeasibility of the proposed method.3. A hybrid method based gradient descent was proposed for optimizing the measurement matrix.Firstly, the optimization model of measurement matrix was established. Then, a greatly improvedgradient descent algorithm which combined chaos motion and momentum term is presented to takefull advantage of them. Based on randomness and ergodicity of chaotic motion, chaos factor wasintroduced to step size so that the step size was adaptive in iteration process; momentum term wasadded to improve the stability of the algorithm. Experimental results show that the method decreasesthe cross-correlation between measurement matrix and sparsity transform matrix and improves theperformance of measurement matrix. Based on this, it designs a target classification algorithm which extracted features of SAR image by measurement matrix directly and realized SAR image targetrecognition. Simulation results show that: this method improves the recognition rate of three typeswith variants; it is an effective method for SAR target recognition.4. An improved optimization method of measurement matrix based on gradient descent (GD)algorithm with variable step is researched. Firstly, gradient descent with variable step method is usedfor solving the established model. And then, cooling factor of simulated annealing (SA) is introducedto the learning rate, so that the step size was adaptive in iteration process. Comparison with theexisting methods, the proposed method has fast convergence and reduces the cross correlationbetween measurement matrix and sparse transform basis. Finally, target recognition rate withfeature extracted by measurement matrix is improved, proves the superiority and efficiency of themethod.