SAR Image Ship Detection Based On Dual Pooling And Vertex Localization And Hardware Deployment

The sea area of China is vast,and the high precision and rapid maritime ship monitoring technology is of great significance for the protection of maritime rights and interests and maritime security.Synthetic Aperture Radar(SAR)is an important high-resolution microwave imaging remote sensor,which can rapidly collect target images in almost any weather and time,and has become an important technical means for maritime ship target monitoring.The intelligent detection method of SAR images based on the convolutional neural network has been developed rapidly in recent years but still faces the following problems: the sidelobe effect brought by the unique imaging mode of SAR seriously interferes with the extraction of target features;the different sizes of targets also affect the degree of detection frame fitting.Although the effect of sidelobes can be improved by image preprocessing or network model structure optimization,the former requires increased detection time and is prone to loss of target information;the latter cannot take into account the effect of sidelobe interference and target multi-size while optimizing the model structure.To address these problems,this paper proposes a ship detection method based on dual-pooling vertex localization and hardware deployment,with the following main work and innovations:(1)A sidelobe interference ship detection method based on maximum pooling and average pooling is proposed.For the sidelobe interference problem of the SAR ship target,the average pooling and maximum pooling are used to reduce the sidelobe interference in horizontal and vertical directions respectively and enhance the extraction of target subject information.Meanwhile,for the information loss problem caused by average pooling and maximum pooling,a network structure based on bi-directional feature fusion is designed to fuse the feature information of the backbone network.Compared with the conventional detection algorithm,the method improves the average accuracy by 1.40% and 0.60% on the SAR public datasets LS-SSDDv1.0 and SSDD.(2)A sidelobe interference ship detection method based on prediction frame vertex localization is proposed.The penalty term of the conventional regression loss function loses its effect in some specific cases,and the sidelobe interference also affects the judgment of the center of mass of the bounding box.The Euclidean distance between the vertices of the prediction box and the real box is used for regression instead of the center of mass and aspect ratio,which can reduce the sidelobe interference and ship false alarm.Compared with CIOU and DIOU,this method improves the average accuracy by 0.50% on the public dataset LS-SSDDv1.0.(3)The deployment of the SAR ship detection method in this paper on a computationally accelerated hardware platform is implemented.A SAR image ship detection network based on dual pooling and vertex localizations is constructed;the network is deployed on the MLU270-F4 compute accelerated hardware platform and compared with NVIDIA RTX3080 and NVIDIA RTXA5000 in terms of power consumption,memory,computation time and accuracy,and he results show that the power consumption of MLU270-F4 in single-core operation is 182.6w and 192 w lower than NVIDIA RTX3080 and NVIDIA RTXA5000.