Research On Ship Target Detection In Optical Remote Sensing Images Based On Deep Learning

Ships are important means of water transportation and military targets.Because of the increasingly busy maritime transportation,efficient and accurate management of ships is of great significance to ensure the normal operation of the economy and the safety of the territorial sea.With the development of computer hardware and the in-depth study of deep learning,target detection based on deep learning is far superior to traditional algorithms in both accuracy and speed.Applying deep learning to ship target detection will bring great convenience to the management of maritime traffic.However,the method based on deep learning has some problems,such as the poor detection effect of small targets and the interference of complex background to the detection effect.This thesis has studied these problems and improved some of them.The main work is as follows.1)This thesis introduces the research background and significance of ship target detection.The algorithm of target detection based on deep learning is studied,and the advantages and disadvantages of each algorithm are analyzed.Using the DIOR data set,the horizontal frame data set and the rotation frame data set for ship target detection are made.2)In the thesis,YOLO v5 algorithm is applied to ship target detection.Three boundary box loss functions are compared and the best one is selected to improve the YOLO v5 algorithm.The Mobile Net v3 network is introduced to lightweight improve the YOLO v5 algorithm.Focal Loss is also introduced to balance the number of positive and negative samples.Experiments have shown that using the improved method in this thesis can greatly improve the detection speed of the model with little accuracy loss.3)The YOLO v5 algorithm is used to detect ship targets using horizontal and rotating frames respectively.In the ship target detection based on horizontal frame,a comparative experiment is carried out to verify the superiority of YOLO v5 in YOLO series of algorithms.In response to the problems encountered in using YOLO v5 for ship target detection,this article has made the following improvements.Firstly,in order to enhance the detection ability of small targets,a dynamic feature fusion module is introduced in the feature fusion stage,which gives different weights to the low-level features and high-level features.Secondly,two attention mechanisms are added to the backbone extraction network and their detection effects are compared.YOLO v5 is improved by selecting the better attention mechanism.Thirdly,this thesis introduces a multi-feature extraction module,which improves the detection speed of YOLO v5 while ensuring the detection accuracy.Through experiments,the method proposed in this paper improves the AP value by 0.15 percentage points and the detection speed by 3.79 FPS.This thesis also uses a rotating frame to detect the ship target and adds a ring smoothing label.The improvement of horizontal frame target detection is applied to rotating frame target detection.Experiments show that the rotating frame can better detect ship targets.The improvement measures proposed in this paper are also effective for the target detection method of rotating frame ship.