Inshore Ship Target Detection In SAR Based On FCOS

Synthetic Aperture Radar(SAR)is an active microwave sensor,which is not affected by illumination and weather,and can be used in all weather.It has a wide range of applications in ship detection.The traditional SAR ship detection methods have some limitations,such as poor robustness and generalization,and can not distinguish ship targets from the background effectively.In addition,the traditional methods can not accurately locate the position of the closely arranged ships.In recent years,deep learning technology has made important breakthroughs in object detection,and has gradually been widely used in the field of ship detection in SAR images.Firstly,based on the original FCOS network,an improved feature network model is proposed.The convolutional block attention module(CBAM)is integrated into the feature pyramid network to extract more salient feature information at different scales,which solves the problem that small and medium-sized ship targets are difficult to detect.Secondly,a bidirectional feature pyramid network is proposed based on the idea of PANet.A bottom-up pyramid module is added after the feature pyramid to fully obtain the spatial location information of the shallow feature map and extract the positioning features,and the salient features are combined with the global fuzzy features to solve the problem of missing detection of the densely arranged offshore ships.Thirdly,this thesis improves the feature fusion method and proposes a weighted feature fusion method.The method can better extract the output feature map information under different resolutions,leads the feature information extracted by the feature map of different layers to have different emphases,improves the robustness of an algorithm and leads a network model to have more generalization capability.Finally,in order to verify the effectiveness of the proposed algorithm,ablation experiments are carried out on SSDD datasets,and data enhancement operations are carried out on SSDD datasets to increase the diversity of data and improve the generalization ability of the training model.Experimental results show that each module can improve the detection accuracy to a certain extent,which proves the effectiveness of the proposed module.In addition,compared with other classical target detection algorithms,the accuracy of this algorithm reaches 90.46%,which is 9.79% higher than the original model FCOS,and the speed is also significantly faster than other detection algorithms,which proves the effectiveness of this algorithm.In order to further prove the effectiveness of the proposed algorithm for inshore ships,this thesis makes comparative experiments with other deep learning algorithms on the SAR-Ship-Dataset dataset,and proves that the proposed algorithm has obvious advantages compared with the existing methods and has higher detection accuracy.Finally,this thesis uses the laboratory SAR image data to test the proposed algorithm,and verifies the generalization of the proposed algorithm.