Development of a high-speed identification model for infrared-ring structures using deep learning

Machine learning-based analysis has become essential to efficiently handle the increasing massive data from modern astronomical instruments in recent years. Churchwell et al. (2006, 2007) identified infrared ring structures, which are believed to relate to the formation of massive stars, with the human eye. Recently, Ueda et al. (2020) showed that Convolutional Neural Networks (CNN) can detect objects with indistinct boundaries such as infrared rings with comparable accuracy as the human eye. However, such a classification-based object detector requires a long processing time, making it impractical to apply to existing all-sky 12 μm and 22 μm data captured by WISE. We introduced the Single Shot MultiBox Detector (SSD, Liu W. et al. 2016), which directly outputs the locations and confidences of targets, to significantly reduce the time for identification. We applied an SSD model to the rings toward the 6 deg2 region in the Galactic plane which is the same region used in Ueda et al. (2020), and confirmed that the time for identification was reduced by about 1/80 with maintaining almost the same accuracy. Since detecting small rings is still difficult by even this model, an input image should be cropped into small images, which increases the number of applications of the model. There is still room for reducing the processing time. In the future, we will try to solve this problem and detect the rings faster.