Galaxy Model Subtraction with a Convolutional Denoising Autoencoder

TL;DR

The paper tackles the challenge of subtracting a galaxy's smooth light to reveal faint nearby sources, particularly in complex morphologies. It introduces a convolutional denoising autoencoder (DAE) trained on GALFIT-generated models injected into real NGVS backgrounds, enabling fully automatic, fast galaxy subtraction and high-quality residuals. Across NGVS data and injection-recovery tests, the DAE matches ellipse fitting for smooth ellipticals and outperforms it for edge-on, spiral, and star-forming galaxies, with negligible photometric bias and residual artifacts. The approach, capable of processing cutouts in under 0.1 s, supports scalable, survey-wide pipelines for upcoming facilities, and the code is openly available for community use.

Abstract

Galaxy model subtraction removes the smooth light of nearby galaxies so that fainter sources (e.g., stars, star clusters, background galaxies) can be identified and measured. Traditional approaches (isophotal or parametric fitting) are semi-automated and can be challenging for large data sets. We build a convolutional denoising autoencoder (DAE) for galaxy model subtraction: images are compressed to a latent representation and reconstructed to yield the smooth galaxy, suppressing other objects. The DAE is trained on GALFIT-generated model galaxies injected into real sky backgrounds and tested on real images from the Next Generation Virgo Cluster Survey (NGVS). To quantify performance, we conduct an injection-recovery experiment on residual images by adding mock globular clusters (GCs) with known fluxes and positions. Our tests confirm a higher recovery rate of mock GCs near galaxy centers for complex morphologies, while matching ellipse fitting for smooth ellipticals. Overall, the DAE achieves subtraction equivalent to isophotal ellipse fitting for regular ellipticals and superior results for galaxies with high ellipticities or spiral features. Photometry of small-scale sources on DAE residuals is consistent with that on ellipse-subtracted residuals. Once trained, the DAE processes an image cutout in $\lesssim 0.1$ s, enabling fast, fully automatic analysis of large data sets. We make our code available for download and use.

Figures (9)

• Figure 1: Galaxy image and model subtraction results for VCC 0407 using traditional methods. The left image shows the original $g'$-band image of VCC 0407, an early-type dwarf galaxy in the Virgo cluster with $M_{g'}=-16.82$ mag. The size of this image is $512\times 512$ pixels, corresponding to $95\hbox{$.\!\!^{\prime\prime}$}7\times 95\hbox{$.\!\!^{\prime\prime}$}7$ in observed units. The center image shows the residual image produced by a ring median filter with an inner radius of 11 pixels (2$.\!\!^{\prime\prime}$057) and an outer radius of 16 pixels (2$.\!\!^{\prime\prime}$992). The right image shows the residual image produced by ellipse subtraction. The ring median filter is unable to subtract the bright galaxy center, while the ellipse fitting struggles to fit the spiral structures in this galaxy.
• Figure 2: An example pair of images in the training set. We use the injected galaxy image (left) as input and the clean model galaxy image (right) as targeted output. Each image is of size $512\times 512$ pixels ($95.7\times 95.7$ arcsecs in observed units). The left image is the injected galaxy image (or "noisy" image in training) produced by injecting GALFIT galaxy models into "blank sky" background images taken from NGVS dataset. The sky background images are chosen so they do not contain bright stars or large galaxies. The right image is the clean model galaxy image (or "clean" image in training) generated using GALFIT package. We bin the training data by apparent magnitudes. This model galaxy ($g'=14.41$ mag) was used in the training set for galaxies with $14<g'<15$ mag.
• Figure 3: Graphical illustration of our DAE model architecture. The encoder (left) compresses the input data into a lower-dimensional latent space vector with key features. The decoder (right) reconstructs the input data from this latent space into a target output which shares some key features with input data but is not identical to the input data. In our denoising autoencoder, the input data are a group of observed galaxy images (with the galaxy, foreground stars, and background objects), and the output data are the corresponding clean galaxy-only images.
• Figure 4: Training loop of the denoising autoencoder. During training, the "noisy" images (galaxies injected into sky backgrounds) are put through the denoising autoencoder, producing output denoised images. The output images are then compared with the clean model galaxy image to calculate the loss by Mean Absolute Error (MAE). Weights in the denoising autoencoder are updated by back propagation to minimize the loss.
• Figure 5: Test result comparisons on Virgo Cluster galaxies in apparent magnitude range $14{-}15$. (a) VCC 0033, a smooth dwarf elliptical galaxy ($M_{g'} = -16.23$). (b) VCC 1304, an edge-on galaxy ($M_{g'} = -16.10$). (c) VCC 0407, a spiral galaxy with $M_{g'} = -16.83$. (d) VCC 1725, a star-forming galaxy ($M_{g'} = -16.80$) with significant artifacts with ellipse model subtraction. Ellipse subtraction and the DAE perform similarly well on smooth galaxies, but the DAE yields cleaner residuals for edge-on, spiral, and star-forming galaxies.