Fast And Accurate Target Extraction For High-resolution SAR Imagery

The techniques of fast and accurate target detection for high-resolution SAR imagery are investigated in this thesis.Classical region and edge-based target detection methods usually adopt bottom-up computation mechanism,thus they depend heavily on low-level features of an image,and there is no chance to correct when the computation error propagated from low-level to high-level,especially in processing inhomogeneous SAR images.In order to resolve this problem,a combinational method based on active contour models(ACMs)that utilize high-level information,visual saliency detection methods that can extract targets quickly,and the deep neural networks with unsupervised learning abilities,is proposed to cope with target detection for high-resolution and large-scale SAR imagery.According to this procedure,this thesis thoroughly studies the robustness of the classical ACMs to SAR imagery,the convex optimization for the energy functionals of ACMs,the fast target detection based on saliency detection and active contour model(ACM)for large-scale SAR imagery,the adaptivity of the multi-scale saliency method,usupervised feature learning of the deep neural network for SAR image chips discrimination and classification.The main work includes the following aspects.(1)The robustness of the classical ACMs to SAR imagery.In this thesis,well-known ACMs are reviewed,and their limitations with respect to target detection in SAR imagery are discussed through theoretical analysis and experimental research;A new ratio distance that measures the similarity between two speckle-image patches is defined using the probability density functions inside and outside the contours;The energy functional of the Chan and Vese(CV)model is then modified based on this ratio distance;Lastly,the modified CV(MCV)and the region-scalable fitting(RSF)models are combined with linear weights in the same manner as the local and global intensity fitting(LGIF)model,hence the proposed model is named the modified LGIF(MLGIF)model.The MLGIF model makes the global intensity fitting(GIF)and the local intensity fitting(LIF)forces complementary to each other during the contour evolution.(2)The convex optimization for the energy functionals of ACMs.The general energy of classical ACMs is given in this thesis,the minimization methods for these ACMs are also analyzed,and the basic reason why the contours of classical ACMs tend to fall into local minima in SAR image segmentation is pointed out,i.e.,their functionals are all non-convex.Then,the convex optimization methods are summarized and analyzed theoretically,and their global minimum solution is proved to exist.Lastly,a convex functional to optimize the MLGIF model,i.e.,the global minimization of the MLGIF(GMLGIF)model,is proposed and proved to achieve a global minimum resolution,and the algorithm of Chambolle's dual formulation is adopted as the numerical minimization of the GMLGIF model,this allows us a fast global minimization of the proposed energy.(3)The fast target detection based on saliency detection and ACM for large-scale SAR imagery.Aiming at the low computational efficiency of the GMLGIF model in target detection for large-scale SAR imagery,an algorithm(i.e.,Algorithm 4.3)based on the GMLGIF model and the spectral residual(SR)approach is proposed to resolve this problem.Then,a novel SAR ATR scheme based on the joint use of the Algorithm 4.3 and other discrimination and classification methods is proposed.The advantages of our SAR ATR(automatic target recognition)scheme are as follows: First,the SR approach makes our scheme more efficient in salient regions preselection.Then the GMLGIF model that makes it more robust to multiplicative speckled noise,is proposed and utilized to cope with the salient regions,so the candidate target chips are obtained.The cumulative sum of a multi-scale lacunarity(CSML)feature has an advantage over the classical lacunarity that only adopted one scale,because the CSML feature increases the differences among the lacunarity features of different chips.In addition,an auxiliary shape context feature,which is combined with the Affine-invariant Fourier descriptor(AIFD),makes the classification more accurate,and this is also a creative endeavor in target classification based on shape features.(4)The fast target detection based on saliency detection and ACM for large-scale SAR imagery.Since the SR approach can only detect small salient regions and very depends on manual threshold setting,and the computational cost of the GMLGIF model is very high when coping with a large-scale SAR image,a new target detection algorithm(i.e.,Algorithm 5.2),which is based on a multi-scale saliency detection approach and the global minimization of the improved LGIF(GILGIF)model adopting split Bregman(SBGILGIF)method,is proposed to address the large-scale SAR imagery rapidly and adaptively.(5)Unsupervised feature learning of the deep neural network for SAR image chips discrimination and classification.Since both the SAR ATR scheme(i.e.,Fig.4.11)mentioned above and the Algorithm 5.2 are very dependent on the manual feature design in salient chips selection,discrimination and classification,a fast target detection scheme,in which an unsupervised deep network is combined with the multi-scale saliency method mentioned above,is proposed to cope with the large-scale imagery automatically,and it does not rely on manual feature design any more.