Study on Star-Galaxy Image Generation Method Based on GAN

• Format: Article
• Language: zho
• Published: The Northwestern Polytechnical University 2019-04-01
• Series: Xibei Gongye Daxue Xuebao
• Subjects: generative adversarial neural network; images of stars and galaxies; stabilized training; loss function
• Online Access: https://www.jnwpu.org/articles/jnwpu/full_html/2019/02/jnwpu2019372p315/jnwpu2019372p315.html

GAN technology has been widely used in image generation field. Generating images of stars and galaxy is of great significance for the prediction of unknown stars and galaxy. GAN has been used to generate star-galaxy images in this paper; the GAN model structure was built; the training strategy for GAN was designed; in order to stabilize the training procedure, we proposed a gird search method for the optimization of several hyper-parameters and an improved neuron discard method, EM-distance was used to modify the loss function in original GAN model. Taking the star-galaxy images in the Sloan digital sky survey (SDSS) as the training dataset, the improved method proposed in this paper and the original GAN were respectively used to generate two kinds of stars and galaxy images with different resolutions, and the comparison has been made to verify the effectiveness of the improved method.