基于自注意力对抗网络的VLA稀疏阵列效应消除技术

VLA Sparse Array Effect Elimination Technology Based on Self-Attention Adversarial Network

摘要: 稀疏阵列具有许多的优点，例如其拥有较低的成本效益，较低的相互干扰性，以及很强的灵活性等。但是天线的稀疏化也同时带来一系列问题：分辨率限制、视场失真、较大的旁瓣等，这些问题都能使得观测的同时带来一系列效应。为了最大程度的消除这些效应对观测带来的影响，本文选择深度学习的方法对带有效应的脏图进行重建。本文在CycleGAN网络的生成器当中加入自注意力机制来对效应进行消除。改进后的SACGAN由带有自注意力模块的ResNet与判别器PatchGAN组成，前者用于提取原始图像中的细节特征来生成以假乱真的图片，后者的主要任务是分辨图片是不是生成的图片。实验使用250份天空模型图像，使用3份用VLA.A配置修改的、仅有6根天线的阵列配置来模拟观测。总共750张图像，使用600张进行训练，150张进行测试。测试结果显示，重建图像的PSNR与SSIM分别达到了48.091与0.881。结果表明，本文所使用的深度学习网络能够很好地还原了使用随机模型观测的脏图，在对脏图当中效应的消除起到了良好的作用，能够更好地还原原始天体图像。

Abstract: Sparse arrays have many advantages, such as lower cost-effectiveness, lower mutual interference, and strong flexibility. However, the sparseness of antennas also brings about a series of problems: resolution limitations, field of view distortion, large side lobes, etc. These problems can cause a se-ries of effects during observation. In order to eliminate the impact of these effects on observations to the greatest extent, this paper chooses the deep learning method to reconstruct dirty images with effects. This article adds a self-attention mechanism to the generator of the CycleGAN network to eliminate the effect. The improved SACGAN consists of ResNet with a self-attention module and the discriminator PatchGAN. The former is used to extract detailed features in the original image to generate fake pictures. The main task of the latter is to distinguish whether the picture is a gener-ated picture or not. The experiment used 250 sky model images, with 3 images using VLA.A modi-fied array configuration and with only 6 antennas being used to simulate observation. There are 750 images in total. 600 are used for training and 150 for testing. The test results show that the PSNR and SSIM of the reconstructed image reached 36.6636 and 0.472 respectively. The results show that the deep learning network used in this article can restore well dirty images observed us-ing random models, plays a good role in eliminating the effects of dirty images, and can better re-store original celestial body images.