Ship Detection And Instance Segmentation In Optical Remote Sensing Image Based On Deep Learning

Ship detection in optical remote sensing image is an ensential step for a wide range of applications such as military applications,civil applications,whose main tasks are detection,location in wide range field fastly,above all,classifier and segmentation for these objects base on our aim.Therefore,the research in this area has great significance both for military and civilian purposes.It is a challenging task due to the different scales and appearances of the ship.On the other hand,object detection task in optical remote sensing image has improved remarkably in recent years due to the achieved advances in deep convolutional neural networks(DCNNs).But,the application is much less in ship detection.So,we implenment this in ship detection in this thesis.Most of the proposed methods depend on two approaches,one namely: a region proposal stage and a classifier stage such as RCNN series,which called two-stage method;another namely: generate anchors and classication through regression once called one-stage method,such as YOLO and SSD.In this thesis,a uniform two-stage model for interpretation process of the ship image based on the based on the complex scenes of remote sensing ship images from two directions of detection and segmentation are proposed based on Faster-RCNN and MaskRCNN respectively.Backone,Region Proposal Network(RPN)and Fast-RCNN are included in Faster-RCNN.Mask-RCNN includes Backone(Res Net101 + FPN),RPN and mask branch.The majority of work in this thesis are as follows:(1)At first,we instroduce the status of the ship detection based on tradional methods and deep learing method respectively.(2)Explain the basic concepts of deep learning and the commonly used deep convolutional neural network architecture briefly.(3)Data preprocessing: Statistics for samples in the dataset,removing the negative samples,make the dataset into the COCO format and Poscal_VOC format dataset respectively,dividing the dataset into a training set,a veliidation set,a test set and data augmentation for the training set.(4)Ship detection scheme based on Faster-RCNN,an optimized anchor frame design scheme is proposed: the anchor ratios and the anchor scales are designed as [16,32,64],[0.2,1/3.0,1,3.,5] respectively;and non-maximum Suppression(NMS)changed to soft-NMS.in the part of the RPN and Fast-RCNN respectively.(5)Ship detection scheme based on Mask-RCNN,an optimized anchor frame design scheme is proposed: the anchor ratios and the anchor scales are designed as [16,32,64,128,256],[1/3.0,1,3] respectively;and non-maximum Suppression(NMS)changed to soft-NMS.in the part of the RPN and Fast-RCNN respectively.(6)And at last,we summary the work in this thesis and looking farward the development of the ship detection in high-resolution optical remote sensing image based on the deep convolution neural networks in the feature.