Solar Filament Segmentation Based on AA-UNet

Abstract

As a tracer of the solar atmospheric magnetic field, the solar filament is extremely important for studying the solar magnetic field. In order to solve the problems of low segmentation accuracy and noise in the existing filament segmentation methods, this paper proposes to replace the convolutional block with an axial attention block in the Encoder part based on the Unet structure. The AA-UNet network takes into account the contextual information among non-adjacent pixels, which helps to perform accurate segmentation. From the results of the comparison experiments in this paper, the proposed method can still achieve good segmentation results even in the case of uneven image quality. The Jac, MCC, and F1-Score metrics on our solar image data test set reach 0.63005, 0.77058, and 0.76659, respectively.