Ship Detection Technology In High Resolusion Remote Sensing Image Using Deep Learning

Ship detection is a traditional task for country's coastal security.Our country has a very long seacoast;remote sensing method could bestially enhance the defense of inshore area and the management of ocean shipping.At present,in pace with the raise of the resolution of optics,the spatial resolution of remote sensing also grows up rapidly.It is not only information's increasing,but also a question about how to process the high resolution image.As the richer detail and more complex texture,it is more difficult to analyze the image and more cost is needed for manual work.Also manual design features are not good enough to this situation.There is a pressing need to a new method which can use the big data effectively,and get a more precise result in ship detection.This thesis embarks from the multi-scale features of remote sensing image,focus on using deep learning to extract the target of ship,include the following three parts: Detect the bounding box of ship using hierarchical CNN,detect the ship by using superpixel method with CNN and use superpixel representation to segment the target.The details are as follows:First,we consider the normal CNN doesn't use the features of all layers efficiently,and the disadvantage of low location accuracy only use high layer.We research how different layers influence the ship detection.We design a single convolution network with high location accuracy and use transfer learning design a CNN which has high Classification accuracy.Combine the advantage of the two networks;propose a hierarchical CNN method which can gain a better result than the normal CNN.Second,we aim at the disadvantage of sliding window that its computation is complexity and it is difficult to put into effect.We combine the superpixel with CNN to detect the target.By analyzing the object proposal method;we finally use a single superpixel to get the possible ship target,and use CNN to judge the possible target accurately.This method combined both the priorities of the accurate estimation of supervised method and the fast calculated procedure of the unsupervised method;Put forward a relative high precise and fast method below no loss of resolution.Finally,we focus on the question to extract the target's pixels by sliding the window one by one.Use the neighbor pixels to distinguish if the central pixel is in target,and achieve the effect of target segmentation method.Also,we research to use superpixel's center to present an area,combine superpixel and CNN to get a faster ship extraction method.