Research On Multi-frame Cascade Reconstruction Method For Solar Adaptive Optics Images

Most of the energy to support life on earth comes from the sun,and it is important for humans to monitor solar activity in real time.However,when observing the sun with ground-based telescopes,the light is inevitably affected by atmospheric turbulence,resulting in severe wavefront distortion and degradation of the imaging quality.Adaptive optics(AO)systems can improve the image quality of the telescope by correcting some of the wavefront aberrations in real time,but the corrected images still has residual aberrations and requires further image post-processing.Several common AO image post-processing methods include: phase diversity,speckle imaging,blind deconvolution,and deep learning methods.In this paper,we design a multi-frame reconstruction algorithm for AO images based on deep learning.The main research contents are as follows:1.We introduced the theory of AO correction system and solar AO image reconstruction,including three common traditional AO image post-processing techniques and the basic theories of deep learning,and then studied the existing deep learning-based solar AO image reconstruction algorithms and their corresponding shortcomings.2.The spatio-temporal attention network unit is constructed with pixel-level optical flow alignment and attention mechanism for the imaging characteristics of solar AO images: 1)each TSAN unit aligns three consecutive short-exposure frames at the pixel level using the PWC-Net alignment model;2)the aligned three frames are then fed into the TSAN and the temporal sharpness(TSP).TSA calculates the matrix similarity between neighboring frames and the reference frame through the attention mechanism to assign different weights to different neighboring frames,and the pyramid structure with up and down sampling to assign the weights of shallow and deep features.The TSP extracts the clearer pixel areas in the reference frame through the temporal sharpness prior.3)The reconstruction network with encoder-decoder structure takes the output of TSA and TSP as input features and reconstructs the clear latent image.4)the loss function adopts the hard example mining strategy(HEM),which allows the model to focus on the areas more difficult to restore.3.To address the shortcomings of existing deep learning AO image reconstruction algorithms,we propose a deep learning multi-frame cascade for solar AO images.CTSAN is constructed by two-stage cascaded architecure cascading with multiple TSAN cascades,and updates only the parameters of a single TSAN unit during back-propagation,and then reuses the same set of TSAN parameters four times during forward propagation,thus enhancing the stability of the algorithm while reducing the number of trainable parameters and computing costs.The results of CTSAN on five real solar AO image datasets are compared quantitatively and qualitatively with other deep learning-based solar AO image reconstruction algorithms.The experimental results show that the CTSAN deep learning algorithm has better reconstruction accuracy,model generalization and stability in reconstructing solar AO images.