SAR Ship Detection Method Based On Deep Learning

Synthetic Aperture Radar（SAR）is capable of all-day and all-weather operation,and it can monitor ships at sea even under the changeable climate conditions.Traditional SAR ship detection methods rely on artificial design features,which require modeling for different sea clutter distributions,and the detection accuracy is limited.In recent years,SAR ship detection methods based on deep learning can automatically learn target features from massive data,with high detection accuracy and strong generalization,attracting the attention of many researchers.However,existing SAR ship detection methods based on deep learning still have some problems,such as:（1）Due to the lack of prior information,the detection accuracy of the existing anchor-free methods is limited（2）Existing methods often use convolutional networks for feature extraction,which makes it difficult to model long range dependencies and extract global information,leading to limited detection accuracy;（3）Due to the majority of small-scale ships and complex background interference,existing SAR ship instance segmentation methods have limited instance segmentation accuracy.In this thesis,the above three problems are studied respectively,and the specific contents are as follows:（1）Aiming at the problem that the detection accuracy of existing anchor-free methods is limited due to the lack of prior information,an anchor-free method of SAR ship detection based on spatial information enhancement is proposed.In this method,a two-path enhancement pyramid is designed,and the spatial information in the hierarchical feature graph is enhanced by adding the bottom-up path.The anchor-free detection head based on spatial attention is designed to guide the boundary box prediction by extracting important spatial information and suppressing interference information.Experimental results show that the proposed method has similar detection accuracy as Faster R-CNN.（2）Aiming at solving the problems of limited detection accuracy due to the limited processing of local neighborhood,limited sensing field and difficulty in modeling long range dependence when using convolutional network to extract features,a high precision SAR ship detection method based on vision Transformer is proposed.In this method,vision Transformer is introduced into the SAR ship detection field to replace convolutional network for feature extraction and model long range dependence.Based on Transformer architecture and nearest neighbor interpolation,a high-resolution HRSWinT backbone network is designed to obtain high-resolution and globally well-informed hierarchical feature maps,and a semantic enhanced pyramid network is designed to further enhance the semantic information in the hierarchical feature maps through multiple top-down pathways.Experimental results show that the proposed method has higher detection accuracy than other methods using convolutional neural network to extract features.（3）Aiming at the problem that existing SAR ship instance segmentation methods are limited due to the majority of small-scale ships and complex background interference,a high-quality SAR ship instance segmentation method based on cascade network is proposed.In this method,an enhanced feature pyramid is designed to obtain high semantic and high-resolution hierarchical feature images,and an enhanced multi-task hybrid cascade network is designed to enhance the information interaction of mask branches and improve the accuracy of instance segmentation.Experimental results show that the proposed method has higher instance segmentation accuracy than other methods.