SAR ship detection through generative knowledge transfer
To address data acquisition and labeling data in the training process of SAR ship-detection network based on deep convolutional neural network,we propose a SAR ship-detection framework via generative knowledge transfer of a knowledge transfer network for SAR image generation and a SAR ship-detection network.The knowledge transfer network consists of three parts:a cycle consistency GAN to synthesize virtual features which have spatial distribution of optical image domain and feature distribution of SAR image domain as well;We further use an identity loss to encourage pseudo-SAR images generated by the knowledge transfer networks to have more of the intrinsic features of SAR images.To alleviate the SAR feature confusion issue,we introduce a feature boundary decision loss to maximize the decision boundary of real SAR features and the pseudo ones.Therefore The knowledge transfer network generates pseudo-SAR images consistent with the spatial distribution of labeled optical remote-sensing images and has a feature distribution similar to those of SAR images.Our proposed method is evaluated from three aspects:(1)The evaluation on the generated pseudo-SAR images.When the object detection network is trained on 70％of SSDD and the pseudo-SAR images,The remaining 30％of SSDD is test set,the AP can reaches 97.50％.As for 0％,10％,20％,30％,and 50％of SSDD,the AP is 64.55％,91.14％,94.69％,96.21％,and 96.84％,respectively.When there is no real SAR images involved in the training process,Ap can still reach 64.55％.(2)Ablation study on loss functions.On the basis of using cycle consistency loss in knowledge transfer network,the best performance comes when applying both the identity loss and the feature boundary decision loss,the AP reaches 64.55％.(3)The evaluation on the ship detection network.The generated pseudo-SAR images in this paper are used in the training process of SSD,Faster R-CNN and YOLOv3 detection networks,which can increase the object detection network to learn more parameters suitable for SAR images,thus improving the detection effect of the network.Experiments in the above three aspects prove the effectiveness of the proposed method.
NETL
NSTL
万方数据
