In astronomical observatory projects, source extraction is the process of extracting celestial objects from observational images and outputting object catalogs. It is the basis for the various analysis tasks that are subsequently performed on the catalog products. With the rapid progress of new, large astronomical projects, observational images will be produced every few seconds. This high speed of image production requires fast source extraction. Unfortunately, current source extraction tools cannot meet the online processing requirement.

In this thesis work, we propose to use the GPU (Graphics Processing Unit) to accelerate source extraction. Specifically, we start from SExtractor, an astronomical source extraction tool widely used in astronomy projects, and study its parall...[ Read more ]

In astronomical observatory projects, source extraction is the process of extracting celestial objects from observational images and outputting object catalogs. It is the basis for the various analysis tasks that are subsequently performed on the catalog products. With the rapid progress of new, large astronomical projects, observational images will be produced every few seconds. This high speed of image production requires fast source extraction. Unfortunately, current source extraction tools cannot meet the online processing requirement.

In this thesis work, we propose to use the GPU (Graphics Processing Unit) to accelerate source extraction. Specifically, we start from SExtractor, an astronomical source extraction tool widely used in astronomy projects, and study its parallelization on the GPU. In our GPU-SExtractor, we re-design and parallelize the three major steps of SExtractor: 1) Background Computation, 2) Multi-Threshold Object Detection and 3) Object Analysis. In particular, we identify the Multi-Threshold Object Detection step as the most complex and time-consuming, and design a parallel detection algorithm based on Connected-Component Labelling. Furthermore, we apply compaction techniques to optimize the detection algorithm to better utilize the massive GPU thread parallelism. In the Object Analysis step, we decompose the analysis into a sequence of GPU-friendly data-parallel primitives to compute the attributes of each extracted object. We have evaluated our GPU-SExtractor in comparison with the original SExtractor on a desktop with an Intel i7 CPU and an NVIDIA GTX670 GPU on a set of real-world and synthetic astronomical images of different sizes. The results show that our GPU-SExtractor outperforms the original SExtractor by a factor of 6, taking a merely 1.9 second to process a typical 4KX4K image containing 167,000 celestial objects.