Ship Feature Extraction And Fusion In Multiple Remote Sensing Images

Integrating information from multiple remote sensing images can improve the reliability and accuracy of target interpretation. In this sense, multiple remote sensing image fusion is a focus of attention in military remote sensing field. Warships are kinds of most important military targets. Due to their mobile and relocatable attribution, it is needed and also possible to use multiple images to acquire more complete information of this kind of targets. In order to detect and classify ships in a marine battlefield, the basic technique of target characteristic analysis, feature extraction and some key techniques for image fusion applications, including multi-target association and fusion detection, are studied systematically in this thesis.Typical remote imaging distortions are analyzed firstly. For sea area optical images, a Convex Hull of Extended Centroids (CH-EC) target global invariant feature extraction method is proposed based on affine geometry theory. This method utilizes the invariant properties of affine geometry to calculate the affine invariant, which has a more fast process speed. At the same time, this method finds the CH-EC to construct uniformly distributed sequential triangle regions, which enhances the stability of feature regions greatly.For a ship target in a complex background of optical images, a MS-Gabor local invariant feature extraction method based on Scale-Space theory is proposed. This method uses the band-pass characteristic and multiple channels characteristic of Gabor filter to find invariant feature points in target images, which is more intuitional and more accordant with vision perceptive model, hence make the invariant feature is more robust under the change of illumination and the disturbance of noise and background. At the same time, based on Scale-Space theory, this method designed Multiple Scale Gabor filter banks, therefore the feature points are scale invariant, which makes the invariant feature extraction method more adaptable to imaging geometric distortion.One of the absolutely necessary pre-condition of multiple remote sensing images fusion is target association, which is to determine if the information from two or more images are related to the same target and should be fused together. Based on the result of image invariant feature extraction, a novel multiple targets association method based on Association Cost Matrix optimization is proposed. Firstly, this method constructs ACM based on the dissimilarities of image invariant feature matching between target pairs from two images respectively, which avoids the bottleneck that the time-dependent kinematic parameters cannot be estimated from sparse remote sensing images. Secondly, the method modeled ACM as a specific object function, and then under the restricted rules of association, the simulated annealing algorithm is introduced to accelerate the process of estimating global optimal ACM, which can distinguish the ambiguous relations among multi-target pairs and has good association performance in dense targets scenarios. Considering the time information of sequential remote sensing images, a multiple feature fusion tracking method is also proposed under the global optimal association frame. The kernel of this method is combining the complementary results of kinematic feature matching and image feature matching to improve the accuracy of multi-target tracking effectively.Multiple remote sensing image fusion can not only use redundant information, but also use complementary imformation among images to improve the performace of target interpretation. The target fusion detection technique based on heterogeneous remote sensing images, SAR (Synthetic Aperture Radar) and optical images, is studied. Firstly, the general frame of auto target detection in remote sensing images is concluded, and then a Standard Deviation Map (STDM) based optical image target detection method is proposed. This method uses local statistic to characterize the object that is obvious different from its neighboring area, so it performes robust detection in optical images. Secondly, after analysed the different target characteristics in SAR and optical images, two target fusion detection methods based on weighted M-distance fusion algorithm and D-S evidential theory fusion algorithm are proposed. The two methods can make full use of complementary information between SAR and optical images at feature level and decision level respectively, both reducing the false alarms effectively.