Classification and Segmentation of Galactic Structuresin Large Multi-spectral Images

Abstract

Extensive and exhaustive cataloguing of astronomical objects is imperative for studies seeking to understand mechanisms which drive the universe. Such cataloguing tasks can be tedious, time consuming and demand a high level of domain specific knowledge. Past astronomical imaging surveys have been catalogued through mostly manual effort. Immi-nent imaging surveys, however, will produce a magnitude of data that cannot be feasibly processed through manual cataloguing. Furthermore, these surveys will capture objects fainter than the night sky, termed low surface brightness objects, and at unprecedented spatial resolution owing to advancements in astronomical imaging. In this thesis, we in-vestigate the use of deep learning to automate cataloguing processes, such as detection, classification and segmentation of objects. A common theme throughout this work is the adaptation of machine learning methods to challenges specific to the domain of low surface brightness imaging.We begin with creating an annotated dataset of structures in low surface brightness images. To facilitate supervised learning in neural networks, a dataset comprised of input and corresponding ground truth target labels is required. An online tool is presented, allowing astronomers to classify and draw over objects in large multi-spectral images. A dataset produced using the tool is then detailed, containing 227 low surface brightness images from the MATLAS survey and labels made by four annotators. We then present a method for synthesising images of galactic cirrus which appear similar to MATLAS images, allowing pretraining of neural networks.A method for integrating sensitivity to orientation in convolutional neural networks is then presented. Objects in astronomical images can present in any given orientation, and thus the ability for neural networks to handle rotations is desirable. We modify con-volutional filters with sets of Gabor filters with different orientations. These orientations are learned alongside network parameters during backpropagation, allowing exact optimal orientations to be captured. The method is validated extensively on multiple datasets and use cases.We propose an attention based neural network architecture to process global contami-nants in large images. Performing analysis of low surface brightness images requires plenty of contextual information and local textual patterns. As a result, a network for processing low surface brightness images should ideally be able to accommodate large high resolu-tion images without compromising on either local or global features. We utilise attention to capture long range dependencies, and propose an efficient attention operator which significantly reduces computational cost, allowing the input of large images. We also use Gabor filters to build an attention mechanism to better capture long range orientational patterns. These techniques are validated on the task of cirrus segmentation in MAT-LAS images, and cloud segmentation on the SWIMSEG database, where state of the art performance is achieved.Following, cirrus segmentation in MATLAS images is further investigated, and a com-prehensive study is performed on the task. We discuss challenges associated with cirrus segmentation and low surface brightness images in general, and present several tech-niques to accommodate them. A novel loss function is proposed to facilitate training of the segmentation model on probabilistic targets. Results are presented on the annotated MATLAS images, with extensive ablation studies and a final benchmark to test the limits of the detailed segmentation pipeline.Finally, we develop a pipeline for multi-class segmentation of galactic structures and surrounding contaminants. Techniques of previous chapters are combined with a popu-lar instance segmentation architecture to create a neural network capable of segmenting localised objects and extended amorphous regions. The process of data preparation for training instance segmentation models is thoroughly detailed. The method is tested on segmentation of five object classes in MATLAS images. We find that unifying the tasks of galactic structure segmentation and contaminant segmentation improves model perfor-mance in comparison to isolating each task.