A non rigid registration model for solar photosphere images based on a hybrid cross attention mechanism

The registration between solar photosphere image sequences is crucial for measured solar physics studies with high spatial and temporal resolutions. In previous studies, we found that the high-resolution photosphere images observed by ground-based telescopes still have some large-scale non-rigid distortions after high-resolution reconstruction, which cannot be eliminated by conventional sequence correlation alignment. These distortions will affect the study of small-scale fine activity in the solar photosphere. To solve this problem, we propose an image registration method based on deep learning in this paper（HCAM-Net）. The method adopts an encoder-decoder framework, and introduces a hybrid attention mechanism to enhance the capture ability of context information, so as to obtain accurate deformation fields and realize non-rigid registration between images. The experimental results on the photosphere images observed by NVST show that the proposed method can achieve accurate non-rigid registration. Compared with the state-of-the-art methods, VoxelMorph and TransMorp, the quantitative indicators and the visualize results suggest the proposed method outperforms the other network in non-rigid registration of solar photosphere images. The SSIM index reaches 0.965 and R² score is 0.976.