基于生成式对抗网络的合成孔径雷达舰船数据增广在改进单次多盒检测器中的应用

针对合成孔径雷达（SAR）图像舰船目标检测领域舰船数据获取成本较高、数据集稀少的问题，提出一种基于像素对像素（pix2pix）生成式对抗网络(GAN)的数据增广技术。制作一个用于pix2pix GAN的数据集，通过对GAN网络的训练和测试得到800张新的SAR舰船样本，并对生成的典型样本进行了客观评价；针对传统SAR舰船目标检测算法鲁棒性差、易受斑点噪声影响的缺点，提出一种基于改进单次多盒检测器(SSD)的SAR舰船检测算法，通过在SSD加入Inception模块增强其对多尺寸目标适应性,提高检测器性能；将pix2pix GAN生成的SAR舰船数据进行标注后加入改进的SSD中，在SAR舰船检测数据集上进行大量对比实验。实验结果表明：当将生成的样本加入原SSD后，检测精度比原SSD检测算法提高了4.3%；当将生成的样本加入改进的SSD后，检测精度相比改进的SSD提高了1.9%；检测器中没有加入生成样本的情况下，改进SSD算法相比原SSD检测算法，检测精度提升了4.7%.

Application of SAR Ship Data Augmentation Based on Generative Adversarial Network in Improved SSD

For the high cost of data acquisition and few of datasets for ship detection in synthetic aperture radar (SAR) image, a data augmentation technology based on pix2pix is proposed. A dataset is set for pix2pix, and 800 SAR ship samples are obtained by training and testing the generative adversarial network (GAN). The objective evaluation is given for the generated typical samples. And for the problems that the accuracy of traditional ship detection in SAR images is susceptible to speckle noise and its generalization is poor, a ship detection algorithm based on single shot multibox detector (SSD) is proposed. An Inception module is added into the SSD detecting algorithm for enhancing its adaptability to multi-size target and improving the performance of detector. Finally, the SAR ship data generated by pix2pix GAN is marked and added to the improved SSD. A large number of comparison experiments were performed on the SSDD dataset. The experimental results show that the detection accuracy is improved by 4.3% when the generated samples are added to SSD. The detection accuracy is improved by 1.9% after the samples are added to the improved SSD; and the detection accuracy of the improved SSD is improved by 4.7% without the addition of the generated sample in the detector.Key