Research On Ship Detection In SAR Images Based On Deep Learning

Synthetic Aperture Radar(SAR)can image in complex external environment with the electromagnetic wave emitted by itself.It has great application potential in marine monitoring.SAR image contains more detailed information,which provides a basis for accurate detection of ship targets in SAR image.Traditional ship detection methods are difficult to deal with the complex marine environment.Convolutional Neural Network(CNN)shows excellent performance in computer vision.CNN can automatically extract targets for accurate detection.Therefore,CNN-based ship detection has gradually become a new research trend in SAR ship detection.Based on CNN technology,this paper designs a SAR image ship target detection algorithm combining the characteristics of ship targets in SAR images.The main work is as follows:Firstly,this article introduces the development of SAR image ship target detection algorithms,and divides the existing detection algorithms into two categories,which are based on traditional algorithms and based on deep learning methods.Then we introduce the theoretical basis and analyze the classic object detection methods.Finally,we summarize the main process of SAR image ship target detection.Secondly,in order to solve the detection problem of various sizes ships and densely arranged ships in SAR images,this paper proposed a ship target detection method with spatial pyramid pooling.The spatial pyramid pooling module can fuse local features and global features.A new loss function is used to balance the contribution of different size targets.Double Threshold Soft non-maximum suppression(DTS-NMS)strategy is proposed to alleviates missing detection of ship targets in dense scenes.Finally,in view of the large number of small-size ships which are difficult to detect in SAR,this paper redesigns the feature extraction network structure of the neural network model.Dilated convolution residual unit used to reduce the missing feature of small ships and path argument fusion network(PAFN)is proposed to enrich the spatial and semantic information of different feature maps.Through the experiments,our method has good performance for small ships detection.