Unsupervised Learning Techniques for Detection of Regions of Interest in Solar Images

Identifying regions of interest (ROIs) in images is a very active research problem as it highly depends on the types and characteristics of images. In this paper we present a comparative evaluation of unsupervised learning methods, in particular clustering, to identify ROIs in solar images from the Solar Dynamics Observatory (SDO) mission. With the purpose of finding regions within the solar images that contain potential solar phenomena, this work focuses on describing an automated, non-supervised methodology that will allow us to reduce the image search space when trying to find similar solar phenomenon between multiple sets of images. By experimenting with multiple methods, we identify a successful approach to automatically detecting ROIs for a more refined and robust search in the SDO Content-Based Image-Retrieval (CBIR) system. We then present an extensive experimental evaluation to identify the best performing parameters for our methodology in terms of overlap with expert curated ROIs. Finally we present an exhaustive evaluation of the proposed approach in several image retrieval scenarios to demonstrate that the performance of the identified ROIs is very similar to that of ROIs identified by dedicated science modules of the SDO mission.