Ship Detection In SAR Images Based On Convolutional Neural Network

Ship detection is to recognize the ships in remote sensing images and regress their bounding boxes to obtain the accurate location of ships.Synthetic Aperture Radar(SAR)is an imaging radar that can work all day and all weather.SAR is one of the main tools used to obtain remote sensing data.Ship detection in SAR images is an important application of SAR technology,which is of great significance to maintain maritime safety and promote the development of marine economy.SAR Ship detection has become a research hotspot in the field of remote sensing.In recent years,deep learning and convolutional neural network(CNN)have been widely concerned with their powerful feature representation ability and generalization performance,and the object detection algorithms based on CNN have made unprecedented achievements.However,due to the special imaging mechanism of SAR images,there are many challenges in SAR ship detection,such as the ships with large scale variation,small objects,and complex background.To meet the above challenges,this paper studies the ship detection algorithms in SAR images based on convolutional neural network.The main work is as follows:(1)An object Detector based on Level-Balancing Pyramid network(LBPDet)is proposed.To detect the objects with large scale variation and small objects in SAR images,InterLevel Mutual Attention Module(ILMAM)and Dual Branch Pyramid Convolution Module(DBPCM)are designed.ILMAM cascades the feature maps of adjacent layers and feeds them respectively into channel attention block and spatial attention block to obtain the mutual attention weights between the feature maps from adjacent layers,to enhance the features conducive to detection and suppress the invalid features.The pyramid convolution operation in the DBPCM can reduce the difference between multi scale feature maps.The top-down and bottom-up pyramid convolution connection paths are used for classification and regression respectively,to provide different features for the two tasks.Finally,experimental results on two SAR ship detection datasets verify this method.The results show that the method greatly improves the detection precision of multi-scale objects and small objects.Compared with baseline,the detection precision of this method on SSDD dataset and SAR-Ship-Dataset has increased by 2.6 and 1.2 respectively,which proves the effectiveness of the method.(2)An object detection method based on Multi Scale Modulated Shared Deformable Convolution(MSMSDC)is proposed.To deal with the difference in spatial receptive fields between feature maps of different layers in existing multi-scale object detection algorithms,MSMSDC is designed to learn the sampling positions and sampling weights of objects with different scale,and a multi-scale parameter sharing mechanism is used to speed up the training and reasoning process.To improve the ability of features to distinguish foreground objects and backgrounds,Attention Dynamic Activation Module(ADAM)is proposed to generate different activation for different tasks,different channels and spatial positions.Finally,the experimental results show that this method can detect positive samples from complex background more accurately.The detection precision is improved by 2.4 and 1.1on SSDD dataset and SAR-Ship-Dataset respectively,which verifies the effectiveness of this method.