Research On Fine-grained Ship Classification Based On Remote Sensing Data

With the development of artificial intelligence technology and the advent of the big data era of remote sensing,the maritime ship monitoring system based on computer vision has gradually developed and improved.This kind of monitoring system can obtain the specific position,category,channel and other information of all maritime ships in the monitoring area,and can even predict the intention of ships through this information by analyzing remote sensing data It greatly reduces the labor cost of processing and interpreting the remote sensing data.And provides great convenience for the monitoring work of maritime ships.According to the system,its workflow is: remote sensing data processing,ship detection,fine-grained ship classification.This paper conducts research on the fine-grained ship classification task,and proposes corresponding solutions to the difficult problems in the task.Among the remote sensing data sources,optical remote sensing data and synthetic aperture radar data are suitable for the computer vision based maritime ship monitoring system.The readability of optical remote sensing data is excellent and the data is sufficient,but it is easily affected by light,weather,etc.,and cannot work at night and in the case of clouds and fog.Due to its special imaging mechanism,synthetic aperture radar data can realize all-weather and full-time earth observation,but its noise is complex and its readability is weak.These remote sensing data and the ship target characteristics have brought many challenges to the maritime ship monitoring task,such as:(1)The ship target is a slender geometric target,and the slice containing the ship target must be extracted from the remote sensing data by the ship detection algorithm before classification.The wrong extraction method will greatly affect the classification result;(2)The number of various types of ships in remote sensing data shows a long-tailed distribution.Some types of ships that need to be focused on are often located at the tail of the distribution,and the data is scarce;(3)The ship finegrained classification task aims to predict the specific class of ship samples,such as bulk carrier,oil tanker.Due to the influence of weather,noise,etc.,ship samples in remote sensing data often have large intra-class differences and small inter-class differences,and their discriminative features are difficult to extract.In this paper,the fine-grained ship classification task is fully investigated and researched on the above difficulties,and the corresponding solutions are proposed.The research of this paper is summarized as follows:1.Aiming at the problem of how to correctly extract slices containing ship objects,this paper defines it as the problem of studying effective ship image standardization strategies.A total of six standardization strategies are discussed,and is confirmed which standardization can effectively improve the accuracy of fine-grained ship classification.Corresponding solutions are proposed for optical remote sensing and synthetic aperture radar data.2.Aiming at the problem that the distribution of ship samples in remote sensing data is a long-tail distribution.This paper proposes an improvement in the training strategy of the network.By adopting a special sampling strategy which undersampling of dominant classes and resampling of non-dominant classes,to ensure the training batches have balanced samples in every class.Thereby alleviating the negative impact of long-tail data distribution on classification models.3.Aiming at the challenges of discriminative feature extraction of ship,large intra-class differences and small inter-class differences.In this paper,we propose a unified multiple proxy deep metric learning framework embedding distribution optimization from the perspective of loss function design.This framework successfully improves the accuracy of fine-grained ship classification by combining the advantages of multiple proxy learning,pairwise learning,and constraints on the divergence of proxies and realistic distribution.This paper verifies the effectiveness and advancement of the proposed solutions through sufficient experiments,and at the same time provides guidance for the follow-up research work on the current task needs and lack of work.