Astronomy has entered the era of big data, driven by the rapid advancements in both ground-based and space-based telescopes. These technological leaps have led to the collection of vast quantities of digital images, making the process of denoising these images an increasingly critical task for researchers. Taking solar physics as an example, the Solar Dynamics Observatory (SDO) is a space-borne telescope that plays a significant role in collecting data, with its higher-level derivatives exceeding 7 petabytes. In the realm of galaxies and cosmology, the Hubble Space Telescope (HST) remains one of the most iconic instruments in space exploration. Since its launch in 1990, the HST has completed over 1.5 million observations, offering high-resolution images of distant galaxies, nebulae, and other cosmic phenomena. These observations have provided unprecedented views of the universe, reshaping our understanding of space and time.

This study explores a novel approach to denoising astronomical observation images using rapidly advancing neural network deep learning methods. The research begins by training and validating the technique on solar observation data (daytime astronomy), and then extends it to nighttime astronomical data, resulting in a universal denoising method applicable to all types of astronomical images. This approach achieves the best universal denoising performance currently available worldwide.

The method combines deep learning with restoration processes and breaks the denoising process into three steps:

1. Training the model on a randomly selected image (T-step),
2. Using the trained neural network to denoise similar images (D-step),
3. Adjusting the regression threshold based on the change in the denoised image to ensure the noise level does not exceed the original noise level of the image (R-step).

This denoising method is named DenoisingTDR based on these three steps. Poster image demonstrates the effectiveness of DenoisingTDR by applying it to data from the Solar Dynamics Observatory (SDO) Helioseismic and Magnetic Imager (HMI) and the Hubble Space Telescope (HST). The noise level in the images was reduced from the original 8 Gaussian to around 2 Gaussian. The denoising also enhanced the Hubble 10-second observation data, making it more comparable to 100-second exposure results. For more details on the network structure, implementation process, and specific denoising performance, please refer to the journal article.

DenoisingTDR combining deep learning with restoration processes has introduced a protective mechanism in the field of scientific image denoising. Scientific images, such as those obtained from astronomical observations, require a high level of precision, and the denoising process must ensure that the original image's accuracy is preserved. Traditional deep learning denoising methods are often based on visual judgment, which may not achieve pixel-level precision.

In our approach, we apply restoration processes to each pixel in the image, ensuring that the change in any pixel does not exceed the noise level. This guarantees pixel-level accuracy while achieving effective denoising, minimizing the impact on the original signal. This method is especially crucial in scientific data processing, as even subtle changes, undetectable to the human eye, could exceed the required precision for the data. In today's fast-evolving era of artificial intelligence, as various industries rapidly adopt deep learning, it is important to remain cautious and ensure that AI is used safely and responsibly. The restoration processes used in this study provide us with valuable insights. Since DenoisingTDR is a universal denoising technique, it has broad applications across various fields that involve image and scientific data denoising, offering significant potential for practical use.