Detection And Classification Of Ship Targets In Remote Sensing Images Based On Deep Learning

Ship is an important military and civilian target.It is of great theoretical significance and practical reference value to study ship detection and recognition based on remote sensing image.Target recognition can be divided into three levels for military reconnaissance and target attack: the first level distinguishes between warships and civilian ships,that is,whether they are warships or civilian ships;the second level distinguishes between warships and civilian ships in specific categories,that is,whether warships are aircraft carriers,destroyers,frigates,cargo ships,cruise ships or oil tankers,etc.and the third level distinguishes between warships and civilian ships.The specific types of warships are analyzed.Convolutional neural network(CNN)is widely used in the field of natural image target recognition.In recent years,it has also been used in the field of remote sensing images.Target recognition based on depth learning is a frontier technology to solve ship target detection and classification.For the task of remote sensing ship target recognition,this paper focuses on the detection and classification of warships and civilian ships in the first level of sea background.The main work is as follows:(1)The resolution of high-resolution satellite images in China has reached 1 m,which can be used to make samples for warships and civilian ships.The target types selected in this paper are as follows: warships(including aircraft carriers,frigates,destroyers,ocean-going ships),civilian ships(including cruise ships,fishing vessels,cargo ships,oil tankers).To solve the problem of scarcity of labeled samples,a data set containing 6180 images is constructed by using data augmentation method.(2)In order to construct a new ship target classification and recognition model,two typical target recognition algorithms Faster RCNN and SSD are compared by using the data set.The performance of Faster RCNN is acceptable,but the speed is slow,and the detection speed of SSD algorithm is faster,but the accuracy is poor.On the basis of evaluating various network architectures,this paper takes RetinaNet as the basic framework,adds clustering algorithm to select a priori box,and improves the accuracy of the algorithm without affecting the real-time performance.Experiments show that the target detection and recognition accuracy of the improved RetinaNet algorithm(clustering algorithm design priori box)is similar to that of Faster RCNN algorithm,and its speed is close to that of SSD algorithm.At the same time,the improved RetinaNet algorithm is an end-to-end ship target classification and recognition algorithm.This paper studies ship target detection and classification in remote sensing image based on depth learning,and obtains some meaningful results.Finer ship classification and recognition will be the next research focus.