DCGAN模型改进与SAR图像生成研究

Study on DCGAN Model Improvement and SAR Images Generation

针对SAR图像识别软件,通过改进DCGAN模型单生成器与单判别器对抗的结构,采用多生成器与单判别器进行对抗,设计了控制各生成器生成图像平均质量的算法,提出了一种基于改进的DCGAN生成SAR图像的方法。为测试和验证多个同类图像识别软件,并进行择优,需要自行设计不同于训练用的图像来对测软件进行测试。此方法可以为择优测试提供一个公平的基准测试集。实验分别使用原DCGAN模型和改进的DCGAN模型生成目标图像和场景图像,并使用公开判别器分别对两种模型生成的新图像进行质量验证。实验结果表明,改进的DCGAN模型比原DCGAN模型生成的图像效果更好,经其训练生成的新SAR图像与原SAR图像相比,质量相当且多样性更好,可以满足软件择优测试的需要。

This paper proposes a method of generating SAR images based on the improved DCGAN.This method improves DCGAN,adopts the model structure of multi-generator versus single discriminator,and uses the algorithm to control the average image quality generated by each generator.In order to test and verify multiple similar image recognition software and select the best one,testers need to design the images that are different from those used in training to test the testing software.This method can provide a fair set of benchmarks for selective testing.Respectively in the experiments,based on the original DCGAN model and the improved DCGAN model,target images and the images are generated,and the public discriminator is used to verify the quality of the new images generated by the two models.The experimental results show that the improved DCGAN model generates better images than the original DCGAN model,and the new SAR images have the same quality and better diversity as the original SAR images,and they can meet the needs of software selective testing.