Research On Sparse Feature-based SAR Image Processing And Applications

Synthetic Aperture Radar(SAR)is a high resolution and active coherent imaging radar working in the microwave frequency band with the advantage of all-weather,distances.It can adapt to bad environment,etc and plays a very important role in earth observation.With the progress of resolution and imaging way technology,SAR image data quantity of ballooning brings great pressure to data transmission and storage,coherent noise brings difficult to SAR image interpretation and application.From the characteristics and manifestation of SAR image target and the development of sparse signal analysis,it is obvious that sparse representation has broad prospects for SAR image processing and application.The research of SAR image processing and application method based on sparse feature can improve the level of SAR image processing and reconciliation,and promote the application of SAR image in civil military and military field.In this paper,starting from SAR image’s sparse characteristics,SAR image compression,speckle suppression,target classify,image fusion of SAR and infrared image based on dictionary learning and the sparse representation model are discussed in depth.The main research work of this paper are shown as follows:Chapter two: Fixed threshold choice of the existing algorithms limits reconstruction accuracy and speed for blind sparsity signal.Hence,the paper puts forward a new improved orthogonal matching pursuit algorithm which selects atoms and determines candidate set more quickly with the nonlinear decreasing threshold,adjusts the supporting set adaptively at each iteration in order to estimate the true supporting set,removes lower-energy atoms from the candidate set with regularized secondary screening,which improves the algorithm speed and ensures accuracy.The simulation results show that the improved algorithm can raise reconstruction accuracy by 8.5%and reconstruction speed by 9.2% under the same conditions.Chapter three: A contourlet domain SAR image de-speckling algorithm via self-snake diffusion and sparse representation theory is presented.For this algorithm,firstly,the contourlet transform is applied to the speckled SAR image,adjusts the directional number of each dimension to represent SAR image in the high dimensional space.Then,the low frequency subband without sparsity is filtered by self-snake diffusion and the filtered coefficient is regarded as the local average estimate of the low-frequency subband in the contourlet domain.Sparse representation optimization model of SAR image is presented for suppressing the speckle noise of the high frequency subbands with sparsity,and solves sparse coefficients of the high frequency subbands by using the improved orthogonal matching pursuit algorithm.Finally,the de-speckled image is reconstructed with all of the filtered subband coefficients by the inverse contourlet transform.The simulation results show that the presented algorithm has a better de-speckling performance with preserving the edge of the SAR image.Chapter four: In connection with SAR image target classify,(1)A method with increasing sub classifier diversity based on Ada Boost is presented.First,it trains weak classifiers with support vector machine by extracting two-dimensional linear discriminant analysis(2D-LDA)feature and generalized two-dimensional principal component analyze(G2DPCA)feature of the training samples.Then,it constitutes a strong classifier by combining weak classifiers with Ada Boost.M2 algorithm.Experimental results based on SAR images show that the presented method is better than those methods using a single classifier Ada Boost algorithm in terms of recognition accuracy and classify time;(2)A method based on extended maximum average correlation height(EMACH)and sparse representation is presented.For the algorithm,it trains samples and generates templates using EMACH filter,extracts the template’s G2 DPCA feature to form an over-complete dictionary,sparse representation coefficient of the test sample’s feature is computed,classification is realized according to the energy of coefficient.Experimental results show that the proposed method has lower complexity and short recognition time,it is a feasible and effective method for SAR image target recognition;(3)A method based on concatenate dictionaries and sparse representation is presented.For the algorithm,it constructs a cascade structure dictionary with the training sample images,solves the sparse coefficients with the cascade dictionary for the testing sample image,classify the SAR image target according to the reconstruction error and voting mechanism.The experimental results show that the proposed method can improve the speed of SAR image target classification.Chapter five: In connection with SAR image and infrared image fusion,(1)An adaptive weighted image fusion method which combines the idea of fuzzy theory based on Curvelet transform is presented.For the method,it defines the membership function with fuzzy logic variables,makes different weights to transform coefficients of different levels,and designs a kind of adaptive weighted image fusion strategy.Experimental results validate the reliability and credibility of this method in term of visual quality and objective evaluation,and it can effectively improve the fusion quality;(2)A novel image fusion method is proposed combining NSCT transform and sparse representation.For the method,firstly,the registered images are decomposed by NSCT transform to obtain the low frequency subband and a series high frequency subbands.Secondly,the low frequency subband with lower sparseness is disposed with region energy fusion strategy.For the high frequency subband,it constructs the over complete dictionary,solves sparse coefficient under the trained dictionary,and chooses the high frequency coefficients with the larger energy fusion rule.Lastly,the fusion image is obtained with the different frequency coefficients by the inverse NSCT transform.Experimental results show that the proposed method can retain a visual quality and objective evaluation index,and performs some related fusion approaches.