Research Of Inshore Ship Target Detection Based On Optical Remote Sensing Imagery

In recent years,due to the advent of the remote sensing big data epoch,target detection and recognition algorithms using on optical remote sensing images has been applied in various fields.In the traditional algorithm,the approach of manually extracting the target feature by hand is far from satisfying the demand for real-time performance under big data.So,we utilize computer to help us complete the task of feature extraction and object detection efficiently.As for the deep learning,it is an algorithm that utilizes complex multiple layers network to extract higher-feature by computer.Based on the deep learning algorithm,our algorithms transfer the network model used in optical natural scene to apply in remote sensing scene by transfer learning in this paper.Meanwhile,we adjust and optimizes the network to tackle problems studied in this paper.The main research work of this paper includes the following three points:(1)This paper,starting with the target detection based on the bounding box,reconstructs the ship target detection algorithm model by using random forest and faster region convolutional neural network and try to tackle the problem of the unbalanced data by replacing the classifier.(2)The second part begins to do the research in the pixel-based semantic segmentation algorithm.Above all,we utilize the full convolutional network to semantically segment the remote sensing image,and Markov random field algorithm is used to eliminate false alarms.And then,the suspected target region is obtained.Next,we combine the traditional change detection algorithm to extract changed information in the interested region.Finally,combining these two results,we propose a port area change detection algorithm model based on the full convolutional network saliency.(3)This chapter uses bounding box-based target detection algorithm discussed in the second chapter of this paper as the backbone framework and combines semantic segmentation algorithm.The third part of this paper starts with the method of instance segmentation.And we use mask regionbased convolutional network to achieve the ship target detection and segmentation.Meanwhile,our network adjusts and optimizes the backbone network architecture and accelerates the computation speed by model compression.At last,the feasibility of the method was verified by experiments.