Astrophotography Image Restoration Using a Convolutional Autoencoder

Abstract

Noise is commonly present in astrophotography images and can obscure faint astronomical structures and reduce their scientific value. Traditional noise-reduction techniques often blur important details and fail to preserve fine astronomical structures. This study investigated whether a deep learning–based approach could more effectively reduce noise while preserving fine structural features in astronomical images. It was hypothesized that a U-Net-style convolutional autoencoder (CAE) trained on synthetically noised astrophotography data can serve as an effective, innovative tool to reconstruct higher quality images. To test this hypothesis, a CAE was trained on publicly available astrophotography images (107 images from the Messier catalog), and synthetic noise was added to these images to create paired training data, allowing the model to learn a direct mapping from noisy inputs to clean outputs. Across 11 test images, the mean peak signal-to-noise ratio (PSNR) increased from 11.9 dB in noisy inputs to 33.0 dB after denoising, while the mean structural similarity index (SSIM) improved from 0.28 to 0.89. The results showed that the CAE reduced the synthetic noise while preserving the most fine astronomical structure. These findings motivate further evaluation of deep learning–based denoising methods for astronomical image analysis.