针对现有合成孔径雷达(SAR)图像数据生成方法大多无法同时生成舰船图像及其检测标签的问题,面向SAR舰船图像生成及目标检测任务,构建基于位置信息的条件生成对抗网络(PCGAN).首先,提出将舰船位置信息作为约束条件用于限制生成图像中舰船的位置,并将其作为舰船图像的检测标签;随后,引入Wasserstein距离稳定PCGAN的训练过程;最后,利用生成的SAR舰船图像及对应检测标签完成YOLOv3网络的端到端训练,实现舰船数据增强与目标检测的协同学习,进而获得更耦合目标检测实际应用的多样性数据.在HRSID(high resolution SAR image dataset)数据集上的实验结果表明,PCGAN方法能生成清晰、鲁棒的SAR舰船数据,舰船检测准确度最高提升1.01％,验证了所提出方法的有效性.
SAR Image Generation Method via PCGAN for Ship Detection
Most of existing synthetic aperture radar(SAR)image generation methods fail to generate SAR images and their detection labels simultaneously.A position-based conditional generative adversarial network(PCGAN)is constructed for SAR ship image generation and target detection.Firstly,ship position information is used as a constraint and a detection label to restrict its position in the generated image.Then,the Wasserstein distance is further introduced to stabilize the training process of PCGAN.Finally,the generated SAR images and their corresponding labels are applied for the end-to-end training of YOLOv3,so as to realize the cooperative learning of data enhancement and ship detection,and further obtain the diversified ship data more coupled with the practical application of ship detection.Experimental results conducted on HRSID dataset illustrate that PCGAN can generate clear and robust SAR ship data,and the accuracy of ship detection can be improved by up to 1.01％,thus verifying the proposed method.
synthetic aperture radargenerative adversarial network(GAN)data enhancementship detectionposition information
万方数据
维普
