基于DCGAN的高分辨率天文图像生成研究  

High Resolution Star-Galaxy Image Generation Using Deep Convolutional Generative Adversarial Neural Networks

作　　者：张光华 王福豹[1] 段渭军[1] ZHANG Guang-hua;WANG Fu-bao;DUAN Wei-jun(School of Electronics and Information,Northwestern Polytechnical University,Xi’an Shanxi 710072,China)

机构地区：[1]西北工业大学电子信息学院

出　　处：《计算机仿真》2019年第12期200-204,262,共6页Computer Simulation

基　　金：美国国家科学基金基础研究项目(ACI-1053575)

摘　　要：为生成分辨率大于等于256×256的恒星和星系图像,帮助预测未知的恒星和星系,采用了深度卷积对抗生成网络(DCGAN)以及改进的神经元抛弃方法。通过研究对抗生成网络(GAN)的结构,优化了图像生成模型和训练过程,得到了稳定的能够生成高分辨率的恒星和星系图像生成模型。为了验证上述方法的有效性,在实验中采用了斯隆数字巡天数据库(SDSS)中的图像作为训练图像,生成了256×256以及512×512两种高分辨率的恒星和星系图像,并将结果与采用GAN生成的结果进行了对比。实验结果证明上述算法有效可行,能够在保证模型稳定性的前提下,生成清晰度更高的高分辨率恒星和星系图像。To generate images of stars and galaxies with a resolution greater than or equal to 256×256 and to help predict unknown stars and galaxies,DCGAN and the improved dropout method of neuron connection are presented.By studying the structure of GAN,the image generation model and training process were optimized,and a stable image generation model of stars and galaxies with high resolution was obtained.In order to verify the effectiveness of this method,images in the Sloan digital sky survey(SDSS)were used as training images in the experiment,and two kinds of high-resolution images of stars and galaxies,256×256 and 512×512,were generated,and the results were compared with those generated by GAN.The results show that the proposed algorithm is effective and feasible which can generate high-resolution images of stars and galaxies with higher definition under the premise of keeping the stability of the model.

关 键 词：高分辨率 深度卷积对抗生成网络 斯隆数字巡天数据库 恒星和星系图像 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]