| Synthetic aperture radar(SAR)is an all-time and all-weather active microwave imaging technology with certain penetrating capability,providing good data support for ship detection in the field of marine monitoring.The key to ship detection is to identify and locate the ship.In recent years,data-driven deep learning ship detection algorithms can get higher accuracy,but they need a large amount of labeled data for training.However,the SAR imaging process is complicated,and the scattering characteristics of ground objects in SAR images increase the difficulty of annotation.In addition,SAR images are sensitive to imaging parameters,the same ship will show differences under different azimuth and pitch angles,which make it hard to obtain the properly labeled SAR ship detection data quickly and in large quantities,thus limiting the performance of deep learning ship detection algorithms.Therefore,effective data augmentation utilizing existing SAR ship data is essential for ship detection.Traditional SAR image generation methods are realized by simulators,but there are still large gaps in fidelity.The generative adversarial networks(GAN)can not only generate high-quality SAR images,but also fine control them,making it possible to obtain a large number of SAR images with sufficient diversity and appropriate labeling.In order to increase the amount of SAR images,this thesis uses GAN to generate SAR ship images with object detection labels,so as to improve the accuracy of ship detection.The main research work is summarized as follows:A position-based conditional generative adversarial networks(PCGAN)is constructed.To solve the problem of original GAN can not generate the ship detection labels,this thesis takes the position of ship as constraint of CGAN to generate ship in specific position of SAR images.Furthermore,the proposed PCGAN can generate ship images with clear direction and contour by introducing mismatched input into the discriminator and using direction information of the ship.In particular,to prevent model collapse,Wasserstein distance is introduced to stabilize training process.Experimental results show that the proposed PCGAN can generate clear and robust ship data,and effectively improve the accuracy of SAR ship detection.A dual attention-driven contrastive unpaired translation neural network(ACUTNN)is built.Contrastive unpaired translation can translate source domain images to target domain,while retaining their detection label,but it can not solve the problem of huge differences in SAR ship sizes between different domains.Therefore,this thesis uses clipping to reduce the possibility of ship deformation.Then,the features extracted by ACUTNN are recalibrated through the spatial and channel attention to strengthen the important features,and the separation of foreground and background is introduced to improve the significance of ship area.Experimental results show that ACUTNN can translate SAR images between different domains and improve the performance of SAR ship detection. |
