SAR Image Generation Based On Conditional Adversarial Variational Autoencoder

Image generation aims to enrich the diversity of data,achieve data augmentation,sample balance,and cross-modal learning tasks.At present,the proposed traditional generative models have limitations such as the poor ability to operate high-dimensional images,lack of details in the generated images,and slow calculation speeds.In order to improve these problems,two classic image generation models based on deep learning,namely,variational autoencoders and generative adversarial networks have been applied to many fields.Moreover,the labeled data of synthetic aperture radar(SAR)is insufficient.In order to the make full use of generative models to generate SAR images,and then achieve data argument to improve the problem of insufficient SAR image samples.As such,considering the characteristics of SAR images and in order to ease shortcomings of the current generative models,two novel generative models are developed to yield the high-quality SAR images.The contributions of this paper are listed as follows:Based on the variational autoencoder(VAE)model,a new generative model(i.e.,adversarial VAE,AVAE)with a novel adversarial training strategy is constructed,which consists of two modules,such as a VAE model and a discriminator.Specifically,different from the traditional generative adversarial network,an improved objective function is constructed,which is built by incorporating the discriminative loss among the generated images,the reconstructed images,and the real images into the objective function,thus solving the problem of insufficient training in the generator module.After that,a composite objective function is designed to jointly integrate the properties of these two modules.Furthermore,the variational lower bound function is trained,and by the idea of adversarial training,the gradient of the discriminator is added to the above objective function,improving the clarity of the generated images by the game process of the discriminator.Finally,a single Gaussian distribution and a Gaussian mixture distribution are selected as the prior probability distribution,and the high-quality and clear SAR images that conforms to the real data distribution are generated with the help of reparameterization trick.Images usually contain the rich semantic information,such as length,width,shape,and other attribute information.In order to make full use of these semantic information,modify the one-hot vector form to make the corresponding semantic vector.Then an effective conditional AVAE(CAVAE)model is proposed.Based on experimental data,an appropriate scheme is further designed to represent the semantic information,thus generating images which are more in line with the attributes of the original data.To verify the effectiveness of two proposed generative models,the value of FID is applied to evaluate the quality of the generating images,which is used as a reference to generate the needed SAR images.Then,with the help of the YOLOv3 framework,the expanded SAR images are applied to the SAR ship detection task,whose experimental results illustrate the superiority of the proposed generation algorithms for higher detection performance.