Automatic Ship Detection In Optical Remote Sensing Images Based On Deep Learning

The rapid development of remote sensing technology has enabled researchers to easily obtain a large number of high-quality,high spatial resolution optical remote sensing images.These high spatial resolution optical remote sensing images contain rich target detail information and complex target spatial structure information,which make it possible to realize large-scale scene understanding and semantic information extraction,and play an important role in national economy and national defense construction.Among them,object detection based on optical remote sensing images has been extensively studied by scholars.However,due to some detection difficulties of the ship itself,these object detection algorithms are still unable to obtain satisfactory results in ship detection.Deep learning has strong feature learning ability and feature expression ability.It can extract complex image global information and object context information.In this thesis,we study the automatic ship detection based on deep learning of optical remote sensing images.The specific research works and main contributions are as follows:(1)This thesis summarizes the research status of deep learning,optical remote sensing image object detection and optical remote sensing image ship detection in recent years.The components of a convolutional neural network which is widely used in the field of computer vision,including convolutional layer,downsampling layer and classification layer,are briefly analyzed.The network implementation and structure of three typical convolutional neural networks: VGGNet,GoogleNet and ResNet are introduced.At the same time,two types of landmark object detection frameworks based on convolutional neural networks are introduced in detail: single-stage target detection framework and two-stage target detection framework.(2)In view of the characteristics of the ship's large scale changes and the small objects in the optical remote sensing image,a multi-scale coverage strategy combining image pyramid,feature pyramid and receptive wild pyramid is proposed.At the same time,the key issues involved in image pyramids,feature pyramids and receptive wild pyramids are explored in detail.For image pyramid,the use of image pyramid(whether the training,testing phase is used alone or in combination)and the selection of nonmaximum suppression algorithms are explored.For feature pyramid,network structure design issues including feature pyramids,candidate anchors and the generating of scales are explored.For the receptive field pyramid,the scale problem of the receptive field is explored.In addition,the joint use of these three types of pyramids is also explored.(3)Aiming at difficult problems such as dense arrangement of ships,arbitrary direction and extreme aspect ratio of ships in optical remote sensing images,a detection scheme based on rotating bounding box is proposed.Firstly,a highly customized rotational bounding box is generated by the rotational region proposal network to roughly detect the ships,then the deep features of the region of interest is obtained by pooling the region of interest based on the minimum circumscribed rectangle,and then multiscale and multi-scale pooling of region of interest are executed to further optimize the characteristics of the region of interest,and finally we use the regression detection network to accurately detect the ships.The scheme uses the rotational bounding box to locate the ships,which effectively avoids the problem of large overlap of dense targets.At the same time,the problem of arbitrary orientation of the ships is solved by the regression of the angle variable.Through multi-scale,multi-regional area of interest pooling,it effectively avoids the interference of oil pollution and islands.A large number of qualitative and quantitative experimental results on the largescale optical remote sensing image dataset DOTA show the effectiveness and advancement of the proposed algorithm compared with the current mainstream algorithms.