Ship Segmentation Method Of SAR Image In Large Complex Scene

As a sensing device for microwave imaging,Synthetic Aperture Radar has all-weather imaging processing capabilities compared with traditional optical remote sensing devices.It has been widely used in civilian remote sensing,surveying and mapping,and military reconnaissance.With the successful launch of China’s Gaofen satellites,it provides strong data support for the detection of ships.The traditional method of ship detection in SAR images mainly combines statistical learning with the constant false alarm rate(CFAR)method.However,this type of method mostly relies on the modeling of ocean clutter,requires the manual design of features,has limited operating speed,and is only effective for ship segmentation in some simple scenarios.Because the SAR image imaging area is generally tens of square kilometers or even larger.In solving the problem of ship detection in large complex scenes,due to the existence of oceans,land,ports,nearshore waters,and other targets with similar backscattering mechanisms,the above methods produce higher false alarm rates.Based on the above problems,this thesis solves the problem of ship target detection in complex large scenes from the perspective of semantic segmentation and creates a ship segmentation dataset in large complex scenes.At the same time,due to the scarcity of SAR image data,this thesis uses transfer learning and few-shot learning to solve the problem of ship detection in large complex scenes.The specific work is as follows:(1)The single-polarization SAR image carries less information,so we combined it with the frequency domain information.A 3D atrous encoder-decoder neural network with global attention modules(GAM-EDNet)is proposed to achieve ship segmentation in SAR images.In order to increase the structural information of the single-polarization SAR images,a 3D image cube is designed as the input of the GAM-EDNet.Aiming at the multi-scale problem of ship targets,GAM-EDNet adopts an encoder-decoder structure and stacks 8 wavelet decomposition coefficient images with the original image to form a 3D image block,and 3D convolution is used to solve the problem of image resolution loss due to down-sampling.At the same time,a global attention mechanism is introduced to guide high-level features at different scales to low-level features to get more accurate results.Experimental results show that the proposed method effectively enhances the segmentation accuracy of SAR ship targets at different scales and gets sharper edges.(2)Due to the problem of small data scale in large complex scenarios,the model has insufficient generalization ability,which often results in superior performance in the source domain data,but performance degradation in the target domain.To solve these problems,a ship segmentation method based on transfer learning is proposed.Based on the proposed GAM-EDNet,the region with the most abundant ship target SAR images is used as the source domain for pre-training,and then a small amount of target domain data is used for fine-tuning,which provides a good prior knowledge for the segmentation of ships.We also combine warmup and snapshot ensembles strategy to avoid the negative transfer effect brought by the small amount of data.Our model-based transfer learning method and training strategy can provide a good knowledge prior to the learning of SAR images in the target domain,reduce the difficulty of learning SAR images that containing fewer ship targets,and effectively improve the segmentation performance of ship targets.(3)To solve the problem of ship segmentation in heterogeneous SAR images,a multi-scale similarity guidance network is proposed.The SAR image ship segmentation problem in different regions is modeled into different tasks,a ship segmentation dataset is constructed,called SARShip-4i.In order to enhance the adaptability of the algorithm to multiple targets of different scales,a multi-scale similarity guidance module is used to improve the segmentation effect of different ship scales.The method in this chapter effectively reduces the amount of annotation data required for the target domain,and provides a solution to the problem of SAR image segmentation under heterologous images.It has achieved relatively good segmentation results with limited target samples.