Research On Ship Detection Method In SAR Image Based On Deep Learning

Since 21 st century,the rapid development of artificial intelligence technology,deep learning on their autonomous learning ability of neural network algorithms in classification,detection and segmentation task to get in-depth study,the current target detection algorithm based on depth study on behavior monitoring,target detection and tracking,automatic driving,and other practical application has proved its excellent performance in the scene.With the progress of remote sensing technology,Radar wave has successfully reached the millimeter level.Synthetic Aperture Radar(SAR),as a kind of Radar wave with the most application and strong penetration,can effectively detect various targets with deep hiding and strong camouflage.High-resolution images similar to optical photographs are generated in active imaging mode.And by virtue of its all-day all-weather working advantage,it has a very considerable application prospect in the national economic life,human natural exploration and military strategic activities.In the 1980 s,with the gradual development of SAR technology,radar detection as the most effective reconnaissance technology in the military field,SAR images have been widely used in maritime target tracking,maritime rescue and other tasks.The generation of a large number of high-resolution SAR images has made researchers pay more attention to the interpretation and processing of SAR images.Therefore,the ship detection task of SAR image has become a key research work in Marine operations.However,the ship detection task of SAR image also faces great challenges in the face of complex sea environment,large target scale and span,dense ship target and variable direction.Traditional target detection algorithms have some limitations,such as poor detection effect,large amount of computation,and vulnerability to sea clutter.Most detection algorithms based on deep learning are vertical frame detection based on anchor point frame,without considering the characteristics of ship targets in SAR images.Therefore,this paper mainly improves the Anchor-free detection algorithm based on deep learning to realize the tilt box detection of ship targets in SAR images.The main work of this thesis is as follows:1)In this thesis,the Anchor-free Center Net network is improved to realize tilt box detection for ship targets in SAR images.This thesis proposes an Anchor-free ship detection algorithm for SAR image,which uses Center Net network to directly return to the target without the end-to-end operation of non-maximum suppression,and can detect the SAR ship target quickly and accurately.In this thesis,an Angle Side Angle(ASA)model structure is proposed based on the improvement of Center Net network to realize the oblique frame detection of ship targets,and a double-branching structure is composed of multi-scale sensing field block and deconvolution layer.Then,the attention block is used to extract and fuse features to provide more effective feature information for detecting head branches.The experimental results prove that the ASA tilt box detection method proposed in this thesis is more suitable for detecting ships with variable directions on the sea.The Anchor-free network used achieves good results in both speed and accuracy of detection.The introduced multi-scale receptive field is beneficial to multi-scale ship detection in complex scenes.2)In practical application scenarios,SAR image detection system is mostly carried in space equipment.Considering the efficiency of detection system and the limitations of hardware equipment,the model studied should be as efficient,accurate and lightweight as possible.In this thesis,a lightweight ship detection model of SAR image is proposed based on the Center Net network.Light-Center Net adopts lightweight Ghost Net network as feature extraction network to achieve the lightweight model;The three-branch structure of the detection head is modified into four-branch structure to avoid the interaction between the two kinds of parameters:scale and Angle.In order to increase the sensitivity of the model to ships of different scales,the feature fusion part uses FPN hierarchical fusion of different scales to obtain the feature information of different receptive fields.The proposed Light-Centernet can also efficiently complete the task of ship detection in SAR images under the experimental result.