Photometry in Crowded Fields: Parametric Models vs. a Data-Driven Approach

The problem of disentangling flux from objects in a crowded image is central to many problems in observational astronomy. Rather than fitting a parametric generative model (e.g., a Sersic model for galaxies) to the data, we consider component separation to be the primary task. Once separated from both neighbors and background, an object is easier to characterize. Given a Gaussian prior in the N_pixel-dimensional space for each component we can perform a joint inference on all components, conditional on them summing to the data. Despite replacing operations on N_pixel-dimensional vectors with operations on N_pixel by N_pixel covariance matrices, this approach is computationally tractable and gives satisfactory results for crowded stellar fields on complex backgrounds. Performance is relatively robust to PSF uncertainty, mis-centering, and masked pixels. Such an approach generalizes to more complex objects in images (galaxies), and to other vector spaces (spectra).