Large scale characterisation of galaxy morphology: a deep learning approach

这个跨学科项目的目的是设计数据科学方法和算法，用于从太空望远镜图像中大规模自动表征星系形态。描述星系的形态特征对天体物理学研究至关重要，因为它们提供了对推动它们进化历史的物理过程的见解。然而，天体物理学界目前面临着大量的图像，这些图像只能通过开发自动和适应的图像分析方法来解决。该项目将通过开发充分利用观测数据所需的工具和方法来解决这一紧迫问题。因此，它将极大地影响天体物理学界开展基于大观测数据的高质量研究的能力。在这个项目中，我们将与天体物理学合作者一起设计最先进的计算机视觉和机器学习技术，以便从大量观测数据中自动精确地表征星系的形态。我们将研究以前从未解决过的挑战，例如理解，建模和解释星系的外观如何因红移而变化，或者同时利用多光谱图像以获得更大的鲁棒性。这些实际上是天体物理学图像分析中的关键挑战。我们还将解决仍然没有解决的问题，即共同检测物体的属性（例如螺旋臂或杆的存在）并估计它们的强度（例如臂的数量和角度以及杆的厚度）。我们的工作将基于深度学习，这是机器学习的一个新领域，可以在许多大数据和图像分析任务中获得最先进的结果，并在计算机视觉社区中引起广泛的热情。我们将开发进一步推进这一令人兴奋的新研究领域的方法，同时特别适用于天体物理图像的分析。这个项目是高度协作和多学科的。它涉及在数据科学，计算机视觉，机器学习和深度学习（斯旺西大学计算机科学系）以及星系形态学和天体物理成像（斯特拉斯堡天文台（法国），布里斯托尔大学）方面具有互补专业知识的机构。新的计算机视觉和深度学习方法的发展将得到天体物理实验室提供的真实和模拟图像的支持。真实的图像数据将从几个观测任务中获得，比如天体物理学合作者正在参与的欧洲航天局欧几里得任务。

The aim of this interdisciplinary project is to design data science methods and algorithms for the large-scale automatic characterisation of galaxy morphology from space telescope images. Characterising the morphology of galaxies is essential for astrophysics research, as they provide insights into the physical processes that drove their evolutionary history. However, the astrophysics community currently faces a deluge of images that can only be resolved through the development of automatic and adapted image analysis methods. This project will address this burning issue by developing the tools and methods necessary to fully exploit observational data. As such, it will greatly impact the ability of the astrophysics community to carry out high quality research based on big observational data.During this project, we will work with astrophysics collaborators to design state-of-the-art computer vision and machine learning techniques for automatically and precisely characterising the morphology of galaxies from a large range of observational data. We will investigate challenges that have never been tackled before, such as understanding, modelling, and accounting for how the appearance of galaxies vary due to redshift, or exploiting multi-spectral images simultaneously for greater robustness. These are in fact key challenges in astrophysics image analysis in general.We will also tackle the still unsolved problem of jointly detecting attributes of objects (e.g. presence of spiral arms or a bar) and estimating their intensity (e.g. number and angle of arms and thickness of a bar). Our work will be based on deep learning, a new area of machine learning that achieves state-of-the-art results for many big data and image analysis tasks and arouses widespread enthusiasm among the computer vision community. We will develop methods that further advance this new and exciting field of research, while being specifically adapted to the analysis of astrophysical images.This project is highly collaborative and multidisciplinary. It involves institutions with complementary expertise in data science, computer vision, machine learning, and deep learning (Computer Science Department at Swansea University), and in galaxy morphology and astrophysics imaging (Strasbourg Observatory (France), University of Bristol). The development of new computer vision and deep learning methods will be supported by real and simulated images provided by the astrophysics laboratories. Real image data will be obtained from several observation missions, such as the European Space Agency's Euclid mission of which the astrophysics collaborators are taking part.