Research On Methods For Object Recognition From High-Resolution Remote Sensing Images In Complex Scene And Their Applications

Along with the rapid development of space technology,communication technology and information technology,people could use many sensors such as Optical,SAR,Hyper spectrum and Infrared to obtain a variety of remote sensing images with high space resolution and high spectrum resolution.Compared with low-resolution images,the high-resolution remote sensing images contain more structural and textural information of man-made objects.In order to detect these objects,various kinds of information should be extracted and merged for describing multiple objects categories.Further more,accurate objects detection and recognition from massive remote sensing data puts forward higher demands for military target interpretation system.In this context,based on the typical problems in target recognition in high resolution images,we carry out the relative research.The main innovation of this dissertation is as follows:1.As single feature and equal-weight multiple features can not give a comprehensive description to objects,the proposed method constructs new shape kernel,feature point kernel and appearance kernel.It uses a generalized way to merge these kernels to get a better presentation for targets.Moreover,the method compines the shape priors with deep boltzmann machine,producing a new energy function with edge,region and contour information.Experimental results show that the proposed method can yield better performance in object detection and extraction tasks compared with traditional single-kernel classifier and other combination methods.2.To recognize target with complex structure accurately,we proposed a target recognition method based on multi-scale sparse reconstruction of surface primitives.Which resolve the problem that it's difficult to identify the targets by using the overall shape in the complex scenes of remote sensing images.On the basis of the use of multi-scale surface primitives,we make similarity measure problem of high-dimensional targets be equivalent to the optimization problem based on the reconstruction by using the new similarity measure.This measure makes the most use of the target's overall shape,avoiding the problem that the difficulty of extracting the overall shape of the target leads to reduction of the target recognition accuracy.3.To make full use of information in complex remote sensing scene,we proposed a method of the target automatic interpretation based on consistent semantic space model.This method introduces the priori knowledge of data,as well as models the meaningful targets and spatial relations.It achieves feature detection and identification of the target by calculating the spatial geometry consistency quantitatively and the obtained semantic parse trees of the scene.Forming an effective high-dimensional data by introducing a variety of features of objects in conditional random statistical models,which improves the accuracy of object characterization and provides a good foundation for extracting semantic knowledge of information.