An Improved U-Net Model for Astronomical Images Denoising

To explore and understand the universe, it is inevitable to analyze astronomical images. But additional and incidental noise information will be added during the acquisition process. These noise will interfere with the sub-sequent image analysis. Hence it is necessary to denoise the astronomical images before analyzing. Astronomical image resolution is large and the star signal is small. Meanwhile,the existing denoising methods are difficult to recover the feature information of stars. To enhance the ability of detail feature extraction while denoising, we propose an improved U-Net model. First, we add the residual connections inside the network to break the symmetry of the network. Residual connections are useful to information transmission, and avoid gradient explosion and overfitting. Meanwhile, residual connections can improve the network representation ability. Secondly, we use the encoding and decoding parts of Receptive Field Block(RFB) connection model. RFB increases the receptive field to strengthen the capability of multi-scale fea-ture extraction. RFB also recover more feature information of stars. Finally, we used an improved upsampling method to avoid generating checkerboard pattern of artifacts. We use the dataset from the Hubble Space Telescope(HST) Archives to verify the experiment. The experimental results verify that the proposed model performs well in the task of astronomical image denoising. Our method can better remove the existing noise of astronomical images and recover the feature in-formation of stars.