Abstract
Astronomy has entered the era of large photometric surveys, transitioning from a relatively data-scarce field of study to a very data-rich one. The images being collected by these new surveys present challenging features such as large size (may be a order of magnitude larger than the images we are used to working with), high number of channels (may reach a few dozen), low signal-to-noise ratio, signal saturation, missing information, unknown objects, and so on. Besides that, there is an increasingly large amount of images available. In this scenario, deep-learning approaches emerge as a very suitable alternative for extracting information from data.Our main proposal is to explore structural information, both spatial and spectral, from astronomical images by making use of deep learning techniques. We expect to develop deep network models that generate rich and general representations of the images, appropriate for a variety of tasks such as object detection, outlier detection or clustering. This would enable the development of end-to-end models for large scale image processing and analysis, which in turn may lead to more precise results and new discoveries in astronomy. As case studies, we intend to explore the problems of automatic detection of objects such as stars and galaxies as well as their fine-grained classification.This project counts with the collaboration of researchers from the Institute of Astronomy, Geophysics and Atmospheric Sciences of the University of São Paulo (IAG-USP), who lead the Southern Photometric Local Universe Survey (S-PLUS), an ongoing sky survey.
|