Research On Ship Recognition Using Optical Remote Sensing Data

Ship is an important carrier in the sea,so ship automatic detection and recognition is of great significance.Especially with the deve lopment of the op tical satellite remote sensing technology,how to automatically detect and recognize the ship-target from big data of optical remote sensing images is a challenging task.A ccording to the practical needs of ship-tar get recognition,this thesis studies on the ship-tar get recognition based on the optical rem ote sensing images the resolution of which is over 4m.The following are the main contents:(1)Aiming at the pro blem of ship tar get segmentation,a Chan-V ese(CV)image segmentation model based on local inform ation entropy-driven is proposed.For this model,the calculation of the local field and the information entropy based on the curve as the weights in the target energy function of the inside and outside of the curve are expressed,which raises the speed and autonomy of evolution.The init ial curve extraction m ethod based on visual saliency detection mechanism is introduced and a two-dimension Otsu threshold segmentation algorithm based on local box filtering techn ique is proposed,which reduces the algorithm 's time complexity.The Hough circle detection algo rithm is improved,in which the image edge points are judged by the given set of inequalities,which reduces the computing amount.(2)In order to selec t the most influential attributes,three kinds of attribute reduc tion algorithms are studied,which are the attribut e reduction algorithm based on the infor mation entropy,the semi supervised algorithm for attribute reduction based on rough set theory and the attribute reduction algorithm based on m ulti-objetive evolutionary method,and the sem i supervised algorithm for attribute reduction based on the dom inance neighborhood rough set theory is proposed.The neighborhood m utual information is used in the attribute reduction algorithm based on the inform ation entropy,and the available resources are richened.The preference-ordered attribute se t(dominance-based set)is defined in the s emi supervised algorithm for attribute reduction based on th e dominance neighborhood rough set theory,and the computing time of the algorithm is reduced.The co mpound crossover operation and k-nearest neighbor rank assignm ent strategy are com bined in the multi-objective evolution attribute reduction algorithm,and the conve rgence speed is accelerated.The sim ulation experiment shows that each of the three methods has its own advantages and disadvantages.(3)For the considering of designing high-perf ormance classifiers,two methods and four models are studied.A manifold space-based affinity propagation method is introduced and a defined distance measuring based on m anifold similarity is presented,then the classification accuracy of classifier is im proved;a info rmation entropy-based HDR m ethod is proposed,entropy is introduced to judge whether or not sample nodes stop splitting,which enables the splitting procedure to complete automatically according to the change of information entropy,the subjectivity of the m ethod is enhan ced;the variational inference-based dynam ic probability generative model is proposed,where th e distance measure based on manifold similarity is introduced,and the num ber of the neighborhood nodes participating in computing is restricted,in which a high perfor mance can be acquired;a finite mixture model based on inverted Dirichlet distribution is im proved,in its initialization procedure,box filtering technique based on integral image is utilized to compute the mean of closer centers,which develops the computational efficiency to a certain degree;a co-training model based on dominance-based neighborhood rough sets theory is improved,in this paper dominance-based reduction measurement is defined,which can simplify the computing process,and also reduce the number of iterations;an ac tive learning model based on m anifold structure is finally improved where the sample's discrepancy is redefined by manifold similarity which increases the quality of classification.With respect to the above methods,theoretical analysis and experimental results indicate the algorithms and models are effective to ship recognition.And it is valuable to consult and promote the research on ship recognition using optical remote sensing data in future.