星系图像超分辨率重建技术研究

Research on Super Resolution Reconstruction Technique of Galaxy Image

SKA1-low接收到的无线电信号转换为图像数据后,由于各种因素的干扰会导致图像的分辨率模糊不清,使得天文学家难以从中获取清晰完整的星系信息。针对这一现象,本文计划在生成对抗网络SRGAN的基础上,对其基础模块进行调整优化,使之更利于重建SKA星系图像。本文采用改进SRGAN的星系图像超分辨率重建算法在星系图像数据集上进行训练及测试,实验结果在原模型上PSNR值提升了3.02 db,SSIM值提升了0.0551 db;此外,本文另从定性与定量两方面与SRResCCGAN、IMDN、LAPAR、及BebyGAN等主流模型作对比,从定量角度分析,PSNR值分别提升了1.96 DB、1.59 DB、4.88 DB、0.81 DB,SSIM值分别提升了0.1103、0.056、0.0381、0.0141,两种主流评价指标数值均有所提升;从定性角度分析,本文模型相较于其他经典模型,重建所得图像的边缘信息更加清晰,更有效地复原了星系图像的细节信息。

After the radio signal received by SKA1-low is converted into image data, the resolution of the image will be blurred due to various factors, making it difficult for astronomers to obtain clear and complete galaxy information. In view of this phenomenon, this paper plans to adjust and optimize the basic module of SRGAN based on the generation of adversarial network, so as to make it more conducive to the reconstruction of SKA galaxy images. In this paper, the galaxy image super-resolution reconstruction algorithm with improved SRGAN is used for training and testing on the galaxy image dataset. The experimental results show that the PSNR value is increased by 3.05 DB and SSIM value is increased by 0.0551 DB on the original model. In addition, this pa-per also made qualitative and quantitative comparisons with SRResCCGAN, IMDN, LAPAR and BebyGAN models. From a quantitative perspective, the PSNR value increased by 1.96 DB, 1.59 DB, 4.88 DB and 0.81 DB respectively. SSIM values increased by 0.1103, 0.056, 0.0381 and 0.0141 respectively, and the values of the two mainstream evaluation indicators were all improved. From the qualitative point of view, compared with other classical models, the edge information of the reconstructed image is clearer and the details of the galaxy image can be recovered more effectively.