中国科学院云南天文台机构知识库

新型真空太阳望远镜（NVST）作为研究太阳物理最主要的光学太阳望远镜之一,在采集图像数据过程中会遭受各种噪声干扰,影响对观测数据的研究。一些基于标准监督深度学习的方法在图像去噪领域取得显著成果,但对于干净数据难以获取的天文图像领域该类方法变得不适用。针对此问题,本文将基于自监督深度学习的图像去噪方法应用于NVST图像去噪。为定量评价网络模型性能,首先对经过重建的数据添加仿真噪声;其次将含噪数据通过噪声水平估计网络对噪声水平进行估计;接着利用自监督卷积盲点网络对图像特征进行学习,同时使用贝叶斯推理对图像进行恢复;最后通过峰值信噪比和结构相似度评价指标、相关性分析和功率谱分析对实验结果进行定量分析。实验结果表明,不论对于仿真噪声数据还是实际观测数据,相较于实验中其他图像去噪方法,本文方法能有效降低噪声对NVST图像的干扰,提高图像信噪比。同时也为干净图像数据难以获取的其他工程领域提供解决思路。

As one of the most important optical solar telescopes for studying solar physics, the new vacuum solar telescope(NVST)suffers from various noise interferences in the process of collecting image data, which affects the study of observation data. Some methods based on standard supervised deep learning have achieved remarkable results in the field of image denoising. However, this type of method becomes inapplicable to the field of astronomical images where clean data is difficult to obtain. To solve this problem, this paper applies the image denoising method based on self-supervised deep learning to NVST image denoising. In order to quantitatively evaluate the performance of the network model, first, add simulation noise to the reconstructed data; second, the noisy data is estimated through the noise level estimation network to estimate the noise level; then, self-supervised convolution blind spot network is used to learn image features, and the image is restored by Bayesian inference; finally, the experimental results are quantitatively analyzed through the peak signal-to-noise ratio (PSNR)and structural similarity evaluation indicators, correlation analysis, and power spectrum analysis. Experimental results show that, whether for simulated noise data or actual observation data, compared with other image denoising methods in the experiment, the proposed method can effectively reduce the interference of noise on NVST images and improve the image PSNR. At the same time, it also provides solutions for other engineering fields where clean image data is difficult to obtain.