ResidualDetectron

Due to the increased depth at which the Large Synoptic Survey Telescope (LSST) will be observing, a large fraction of the observed galaxies will overlap with neighboring objects. This is referred to as blending, since light from multiple objects appear blended together. The extent of blending will be comparable to the overlapping object in the HSC Ultra-Deep Field shown below.

HSC COSMOS Ultra-Deep Field

A significant fraction of these overlapping galaxies will not be recognized as blended object by the LSST Science Pipeline. This could result in several issues:

We explore in this project an iterative method to detect objects using Neural Networks. We run a region based Convolutional Neural Network on the model residuals of the current LSST “deblending” algorithm. Let’s discuss this in detail.

Background

Unrecognized Blends

If galaxies are too close to each other, then they might not be detected as separate objects. We call such cases as unrecognized blends.

Current LSST Deblender: Scarlet

The current state of the art multi-band deblender is Scarlet, developed for potential use for ground- and space-based multi-band images. The algorithm models the observed scene as a mixture of components with compact spatial support and uniform spectra over their support while incorporating (optional) constraints such as symmetry and monotonicity. It is important to remember that Scarlet does not perform source detection; it just models the observed blend as a sum sources centered at the input centers and satisfying the input Neural. Thus it requires source centroids + templates as input.

Could unrecognized blends be detected from the residual images?

Neural Networks are well poised to perform detection with non parametric models while utilizing multi-band information. We want the network to learn from the multi-color dipolar patterns in the residual image to identify centers of the undetected source, without shredding it.

Network Architecture

Aim: Run Mask R-CNN detection network, with residual images + the Scarlet model as input, to predict undetected source locations.

Dataset

Results

Performance metrics