Radio Galaxy Morphology Classification with Mask R-CNN

Abstract

Due to the sheer scale of data generated from newer radio galaxy surveys, automating the classification of radio galaxy morphologies has become an active area of study in recent years. One promising solution is CLARAN: A proof-of-concept radio source morphology classifier based upon the Faster Region-based Convolutional Neural Network (Fast R-CNN) method. Using the same dataset based on FIRST (Faint Images of the Radio Sky at Twenty-centimeters) defined in CLARAN we adopt the next generation Mask R-CNN to the problem of radio galaxy morphology classification and compare their performance. The advantage of applying Mask R-CNN to this problem is its additional capability to predict a mask that works in tandem with the region proposal network to classify radio galaxy morphologies. Our results show that Mask R-CNN average precision compares to CLARAN only for the single component single source subjects while CLARAN clearly outperforms Mask R-CNN in other classes. We provide explanations on where the performance gaps are coming from and propose future improvements to close the performance gap.