Ship Target Detection In Satellite Remote Sensing Images

Ship target detection of satellite remote sensing image is to locate the ship in the image.In typical deep learning algorithms,horizontal-vertical bounding box is usually used to locate the target to be detected.However,the ship target distribution in satellite remote sensing image is arbitrary direction,for the narrow and long ship target with arbitrary direction,using horizontal-vertical bounding box to locate is not accurate enough.Therefore,the use of rotating bounding box in detection is more conducive to the location of ship targets.This paper studies the above problems,and the main research contents are as follows:(1)Aiming at the problem of how to use the rotating bounding box to locate the ship target,this paper proposes a ship detection algorithm in arbitrary direction based on dense sub-region cutting.Firstly,the whole ship is intensively segmented along its long axis into several local sub-regions included in the square annotation boxes,to ensure the high effective-area-ratio of the sub-region in each annotation box and thereafter the high generalization ability of the core detection network YOLOv3.Secondly,the core detection network is trained to detect these sub-regions and some overlapping sub-regions are reorganized during the training process.Lastly,the detected sub-regions are combined based on graph clustering method to estimate the direction angle and other key object parameters.Experimental results show that the algorithm can locate ship targets by rotating bounding box,and compares it with the latest improved YOLOv3,RRCNN,RRPN,R-DFPN-3 and R-DFPN-4 algorithms on the HRSC2016 data.Compared with the R-DFPN-4 control algorithm with the highest detection accuracy,the mAP(IOU=0.5)value increases by 0.9 %,and the average timeconsuming decreases by 57.9 %.(2)There are deficiencies in the ship detection algorithm in arbitrary direction of dense sub-region cutting: a)For two ship targets with a head and tail,the algorithm cannot separate the head and tail when merging;b)There are some ships due to the subregion results are not dense enough,the algorithm in the merger of multiple continuous detection boxes.In view of the above deficiencies,this paper proposed a ship detection algorithm in arbitrary direction based on dense sub-region type subdivision.Firstly,according to the characteristics of the ship’s bow and stern,the ship’s dense sub-region is divided into bow and stern class and hull class;Secondly,the two sub-regions of bow and stern and hull are used as the detection targets,and the core network YOLOv5 is used for detection.Lastly,the parameters of the ship target are estimated according to the dense sub-region of the hull detected by the core network,and then the pre-estimated parameters of the ship target are corrected by detecting the dense sub-region of the bow and stern to obtain the complete overall parameters of the ship target.Experimental results show that the algorithm solves the problem that the ship detection algorithm in arbitrary direction based on dense sub-region cutting can’t distinguish the connected parts of the ship and the continuous detection boxes appear in some ship detection.In the same experimental environment,the performance of the algorithm is improved compared with that of the ship detection algorithm in arbitrary direction based on dense sub-region cutting,and its mAP is increased by 2.9 %,and the average time-consuming is reduced by 23.8 %.(3)According to the specific experimental process of the two algorithms,a series of auxiliary software related to the algorithm experiment is designed,including data enhancement software,dense sub-region cutting software,image ship target detection software and other auxiliary software,which makes the experiment more convenient.