基于生成对抗网络的空中目标图像生成算法研究

Research on Air Target Image Generation Algorithm Based on Generative Adversarial Networks

复杂实战背景下的空中机动军事目标视频图像数据较难获得,导致缺少训练数据,从而限制了基于深度学习的目标检测和识别算法的性能。为了解决这一问题,提出一种基于深度卷积生成对抗网络(deep convolutional generative adversarial networks,DCGAN)改进算法,采用随机向量作为输入,生成空中机动目标数据。另外,判别器使用改进的Wasserstein距离衡量生成样本和真实样本的数据分布,优化损失函数,提高DCGAN模型训练过程的稳定性和生成图像的质量。实验结果表明,基于改进的DCGAN图像生成算法能够很好地生成各种实际场景下的空中机动军事目标图像,改进后模型训练过程稳定,损失函数波动明显下降,生成图像更真实。32×32分辨率图像中FID(Frechet inception distance)和IS(inception score)得分分别提高了9.4%和7.6%;64×64分辨率图像中FID和IS得分分别提高了5.9%和4.8%。

It is difficult to obtain the video image data of air maneuvering military targets under the complex actual combat background.The lack of training data limit the performance of target detection and recognition algorithms based on deep learning.Thus,this paper proposes an improved algorithm based on the deep convolutional generative adversarial networks(DCGAN).The enhanced algorithm could generate air maneuvering target data by injecting random vectors.Besides,the discriminator in this paper uses an improved Wasserstein distance measurement to generate the data distribution of the sample and the real sample,thereby optimizing the loss function and improving the stability of the DCGAN model training process and the quality of the generated image.Experimental results show that the improved DCGAN image generation algorithm can generate air maneuvering military targets images in various actual scenarios.After the improvement,the model training process is stable,the loss function fluctuations are significantly reduced,and the generated images are more realistic.The FID and IS scores in 32×32 resolution images have increased by 9.4%and 7.6%respectively.The FID and IS scores in 64×64 resolution images have increased by 5.9%and 4.8%respectively.