Research On Deep Learning-Based SAR Ship Detection And Recognition Technology

Ship detection and ship recognition from synthetic aperture radar(SAR)imagery are both playing a considerable role among marine traffic surveillance,fishery management,emergent shipwreck rescue,national defense protection,and many other fields.In recent years,with the rise of artificial intelligence,deep learning is providing a new pathway to achieve high-performance SAR ship detection and recognition.Nevertheless,nowadays,there are still some challenging problems to be urgent to resolved.For SAR ship detection,the multi-scale characteristic of ship targets and the complexity of scenes both reduce the detection accuracy,current many large-scale networks miserably sacrifice the detection speed,small-scene datasets restrict detection models from applying to large-scene marine surveillance tasks,and current many box-level detection models hinder the pixel-level refined representation of ships.For SAR ship recognition,the excessive dependence on networks’ abstract features may reduce the reliability of recognition models,and the lack of radar polarization information sufficient utilization hinders the further improvement of the recognition accuracy.Therefore,in order to address the above problems,this dissertation has carried out relevant research.The main research contents of this dissertation are as follows:(1)Aiming at the low-accuracy problem of multi-scale ship detection,multi-scale SAR ship detection based on the feature pyramid network(FPN)is studied.Specifically,an improved feature pyramid network named Deform-FPN based on the deformable convolution(Deform-Conv)is proposed to perform the adaptive geometric outline modeling of multi-scale ships.In addition,a novel quad feature pyramid network(QuadFPN)is proposed to achieve the asymptotic scale information enhancement interaction of multi-scale ship features.Finally,the experimental results on the open multi-scale bounding box SAR ship detection dataset(BBox-SSDD)show that the proposed DeformFPN and Quad-FPN can improve multi-scale SAR ship detection accuracy.(2)Aiming at the low-accuracy problem of ship detection among complex scenes,SAR ship detection among complex scenes based on the balance learning(BL)is studied.Specifically,a balance scene learning mechanism(BSLM)is proposed,which realizes the automatic augmentation of training samples of inshore complex scenes and also achieves the focused learning on more complex scenes.In addition,a more comprehensive balance learning network(BL-Net)is proposed to achieve detection models’ multi-level learning balance from the data-level,sampling-level,feature-level,and task-level.Eventually,the experimental results on the open high-resolution SAR images dataset(HRSID)which is provided with many complex scenes indicate that the proposed BSLM and BL-Net can improve SAR ship detection accuracy among complex scenes.(3)Aiming at the problem that previous many large-volume networks sacrifice the detection speed,SAR ship fast detection based on light-weight models is studied.Specifically,a deep detection network based on the depthwise separable convolutional neural network(DS-CNN)is proposed,which can reduce network parameters and can also decrease calculation costs.In addition,a hyper-lightweight deep network named Hyper Light-Net on the basis of the structure optimization and feature enhancement is proposed,which achieves the joint optimization of the detection accuracy and detection speed.Eventually,the experimental results on the public early SAR ship detection dataset(SSDD)indicate that the proposed DS-CNN and Hyper Li-Net can improve SAR ship detection speed meanwhile ensuring the detection accuracy.(4)Aiming at the problem that small-scene datasets hinder models’ applications among large scenarios,large-scene SAR ship detection based on Sentinel-1 is studied.Specifically,a large-scene small target SAR ship detection dataset named LS-SSDD-v1.0based on Sentinel-1 is established.This dissertation reveals its significant advantages and huge potential values among large-scene marine surveillance tasks.Additionally,a false alarm suppression method based on the pure background hybrid training(PBHT)is proposed to reduce land’s detection false alarms when detecting ships among large scenes.Eventually,the experimental results indicate that the released LS-SSDD-v1.0 and PBHT are helpful for detection models’ migration applications among large scenarios.(5)Aiming at the problem that current many box-level detection models limit the pixel-level refined representation of ships,SAR ship pixel-level detection based on the instance segmentation is studied.Specifically,a full-level context squeeze-and-excitation region of interest extractor and mask enhancement prediction network(FL-CSE-ROIEMEP-Net)is proposed to achieve more refined ship pixel prediction.Additionally,a mask attention interaction and multi-scale enhancement network(MAI-MSE-Net)are proposed to achieve more accurate pixel prediction of multi-scale ships.Finally,the experimental results on the public pixel-level polygon segmentation SAR ship detection dataset(PSegSSDD)indicate that FL-CSE-ROIE-MEP-Net and MAI-MSE-Net can achieve superior SAR ship pixel-level representation.(6)Aiming at the problem that current many recognition models excessively rely on networks’ abstract features which leads to reducing decision-making credibility,SAR ship recognition based on the traditional hand-crafted feature fusion is studied.Specifically,a novel deep recognition network named HOG-Ship CLS-Net based on the histogram of orientated gradient(HOG)feature fusion is proposed,which can achieve the organic combination between networks’ abstract features and traditional hand-crafted expert features.Moreover,another more general traditional hand-crafted feature fusion approach is proposed to further popularize this technology of combining deep networks’ abstract features with traditional hand-crafted features.Ultimately,the experimental results on the public Sentinel-1 Open SARShip-3 three-classification and Gao Fen-3 FUSAR-Ship-7seven-classification dataset show that the proposed HOG-Ship CLS-Net and another more generalized traditional hand-crafted feature fusion application promotion approach can improve SAR ship recognition accuracy so as to potentially enhance the credibility of the recognition model.(7)Aiming at the problem that the lack of radar polarization information sufficient utilization may hinder the further improvement of the recognition accuracy,SAR ship recognition based on the radar polarization information fusion is studied.Specifically,a multi-polarization channel attention with squeeze-and-excitation and Laplacian pyramid network multi-resolution analysis network with the dual-polarization feature fusion(SELPN-DPFF)is proposed,which can achieve the refining extraction of ship polarization features from coarse to fine.Additionally,a novel multi-level polarization feature fusion and geometric feature adaptive embedding recognition network named PFGFE-Net is proposed,which can achieve multi-level polarization feature fusion at the data-level,feature-level,and decision-level,and can also achieve the adaptive joint decision-making with SAR ship geometric features.Ultimately,the experimental results on the public complex dual-polarization Sentinel-1 Open SARShip-3-Complex three-classification dataset and Open SARShip-6-Complex six-classification dataset show that the proposed SE-LPN-DPFF and PFGFE-Net can further improve SAR ship recognition accuracy.