Large-Scale Remote Sensing Ship Image Database Organization And Experimental Analysis

Today,target detection is an important research topic in the field of pattern recognition.The remote sensing image has the characteristics of high efficiency in innovation,large observation area,and abundant collection channels.As a result,more and more fields such as military,meteorology,fishery,and navigation have all involved the application of remote sensing images.The drastic increase in the remote sensing image data has made great contradictions between the methods of remote sensing image processing and application that are still relatively backward today.It makes the practical application of remote sensing images unable to make a breakthrough.In recent years,although there has been a lot of attention on target detection of remote sensing ship images at home and abroad,considerable progress has been made in both theory and technology,however,the fact that some remote sensing images have low resolution and there are no publicly available remote sensing image datasets for certain ship types,the detection results of the algorithms are not ideal,which is a great obstacle to the improvement and development of detection technology.Therefore,based on the existing four deep learning algorithms,this paper aims to establish a large-scale remote sensing ship image database and design a complete ship detection system based on deep learning,and using four kinds of algorithms to compare,so as to start the research of this topic.The main research contents and achievements of this article are described as follows:(1)A method for establishing a large-scale remote sensing ship image database is proposed.The high-resolution remote sensing image collection,sorting,labeling,specification and category classification are designed in detail.Through the three enhancement processing of the data,the sample set was expanded and the remote sensing image database YNU-SRSID was established.The remote sensing image of the database has a resolution of up to 0.23 m,and the target texture is clearer,providing more detailed information.(2)In terms of ship target detection,four deep learning algorithms CNN,RCNN,Fast RCNN,and Faster RCNN were used to compare the test results.The detection accuracy of these four algorithms are: 89.66%,90.39%,91.86%,and 93.44%.Then,statistics were collected on the detection of specific types of ships,including hull adhesion,port conditions under complex conditions,sea conditions without complex environments,small vessels and large vessels,the corresponding detection rates were 89.68%,86.45%,93.99%,89.45% and 95.42%.The detection results show that the remote sensing image ship target detection system basically meets the requirements.The self-built remote sensing image database sample can complete the inspection task,and the effect has basically reached the expectation.