Research On Detection Algorithm Of Ship Objects In SAR Images Based On Deep Learning

SAR image interpretation has become the frontier of remote sensing applications.Ship detection based on deep learning is an important research topic in the field of SAR.However,SAR image has the characteristics of sparse data set,strong diversity,strong port interference and image noise,which seriously affect the detection performance.This paper studies SAR ship detection based on deep learning,and proposes an overall scheme to improve the effect of target detection.The research work mainly includes three parts: data enhancement based on style transfer,image processing module,and improved target detection algorithm.According to the characteristics of sparse and high diversity of SAR image data sets,the fusion data set was analyzed and prepared,and the key point of SAR ship detection is to improve the generalization ability of the model.Therefore,the radar data enhancement network based on AdaIN is designed to effectively expand the data set and greatly improve the generalization ability of the model.A new style transfer network is designed by combining Encoder,AdaIN module and Decoder in order.According to the characteristics of wide range of target scale in data set,multi-scale module is designed to enhance the network’s ability to extract features of different scales.For the characteristics of coastal ships that are greatly disturbed by land,channel attention aggregation module is designed to enhance the feature extraction of important channels and strengthen the corresponding features.The jump connection composed of deformable convolution is used to restore the low loss of deep network.In order to improve the information extraction ability of the network,the multi sampling module is designed,and the density loss and edge loss are proposed to enhance the detail effect,so that the algorithm is more suitable for data enhancement.Finally,it analyzes the effective reasons of style transfer from three aspects of feature space,target shape and polarization mode,and designs the methods of train enhancement and test enhancement.According to the two characteristics of SAR image,single channel image and noise,two plug and play modules are designed.The density edge feature extraction module is designed to extract the density and edge features of the target,strengthen the feature information of the gray image,and transform the single channel image into threedimensional image.The denoising module is designed to improve the input quality of the network.The combination of the two modules can be used as the pre-processing module of the detection network to realize the end-to-end training.According to the characteristics of SAR image,the YOLOv5 m detection network is improved.The gray level energy correction is proposed to adaptively adjust the loss of different quality detection frames.The global attention module is improved through the adding connection module,the context semantic information is considered,the information interaction of the network is strengthened,the features of different scales are fully integrated and the performance of the network is improved.The overall scheme proposed in this paper can effectively improve the detection accuracy and generalization through the experiments on the open SAR ship target detection data set,and had achieved the second place in the international SAR ship detection competition of IEEE JSTARS.