基于Unet++GAN的天文图像混合退化复原

DEFOCUS DEGRADATION RESTORATION OF ASTRONOMICAL IMAGES BASED ON UNET+KG-*3+ GENERATIVE ADVERSARIAL NETWORK

  • 摘要: 天文观测常常会受到很多干扰,造成采集到的图像产生各种形式退化,其中较为常见且复杂的为离焦模糊及光电子噪声的混合退化,传统复原手段难以恢复出高质量图像。因而创新地提出利用Unet++改进生成对抗网络的方法,采用更精细的网络结构对图像细节进行准确提取,对比实验说明此方法恢复图像质量较高,并通过恢复真实拍摄的离焦图像,证明了方法具有一定的通用性。改进方法适合处理大数据量的天文图像,不仅如此,模型的泛化能力以及训练稳定性有明显提升。

     

    Abstract: Astronomical observations are often interfered, resulting in various types of degradation of the collected images, among which defocus blur and photoelectronic noise are common and complex. It is difficult to recover high quality images by traditional restoration methods. Therefore, the innovative method of using Unet++ to improve the generative adversarial network is proposed. The finer network structure was used to accurately extract the details of the image. Comparative experiments show that this method has higher quality of image restoration, and by restoring real defocus images, it proves that the method has a certain versatility. The improved method is suitable for processing astronomical images with large amount of data. Moreover, the generalization ability and training stability of the model are obviously improved.

     

/

返回文章
返回