A deep learning approach for automated segmentation of magnetic bright points in the solar photosphere^★

¹ Yunnan Normal University, School of Information, Kunming, Yunnan 650500, PR China
e-mail: yyang_ynu@163.com
² Yunnan Normal University, School of Physics and Electronic Information, Kunming, Yunnan 650500, PR China
³ Key Laboratory on Adaptive Optics, Chinese Academy of Sciences, Chengdu, Sichuan 610209, PR China
⁴ Institute of Optics and Electronics, Chinese Academy of Sciences, Chengdu, Sichuan 610209, PR China
e-mail: chrao@ioe.ac.cn
⁵ University of Chinese Academy of Sciences, Beijing 100049, PR China
⁶ School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, PR China

Received: 16 May 2023
Accepted: 4 July 2023

Abstract

Context. Magnetic bright points (MBPs) are small, bright, and conspicuous magnetic structures observed in the solar photosphere and are widely recognized as tracers of magnetic flux tubes. Previous studies have underscored the significance of MBPs in elucidating the mechanisms of coronal heating. The continuous advancement of solar telescopes and observation techniques has significantly enhanced the resolution of solar images, enabling a more detailed examination of MBP structures. In light of the growing availability of MBP observation images, the implementation of large-scale automated and precise MBP segmentation methods holds tremendous potential to facilitate significant progress in solar physics research.

Aims. The objective of this study is to propose a deep learning network called MBP-TransCNN that enables the automatic and precise pixel-level segmentation of MBPs in large quantities, even with limited annotated data. This network is designed to effectively handle MBPs of various shapes and backgrounds, including those with complex features.

Methods. First, we normalized our sample of MBP images. We then followed this with elastic deformation and rotation translation to enhance the images and expand the dataset. Next, a dual-branch encoder was used to extract the features of the MBPs, and a Transformer-based global attention mechanism was used to extract global contextual information, while a convolutional neural network (CNN) was used to extract detailed local information. Afterwards, an edge aware module was proposed to extract detailed edge features of MBPs, which were used to optimize the segmentation results. Focal loss was used during the training process to address the problem of the small number of MBP samples.

Results. The average values of precision, recall, F1, pixel accuracy, and intersection over union of the MBP-TransCNN are 0.976, 0.827, 0.893, 0.999, and 0.808, respectively. Experimental results show that the proposed MBP-TransCNN deep learning network can quickly and accurately segment the fine structure of MBPs.