| The observation and study of solar activities are of great significance to human beings.Due to the influence of the earth’s atmosphere,the ground-based telescope can only observe the blurred solar speckle image,and the high-resolution image play an important role in the study of the sun.Therefore,high-resolution image reconstruction of solar speckle has always been an important research content in astronomy and solar physics.The traditional reconstruction algorithm makes use of the statistical information of the solar speckle image,the calculation process is complex and the reconstruction time is long.In recent years,deep learning has developed rapidly in the field of computer image processing,especially in the task of image transformation and image reconstruction.Based on the image reconstruction technology of deep learning,this thesis focuses on the application of the generate adversarial network in the reconstruction of single frame solar speckle image.The main research contents are summarized as follows:(1)Combined with the current research background,two kinds of reconstruction algorithms are introduced in detail: traditional solar speckle image reconstruction algorithm and image reconstruction algorithm based on depth learning.We compare the traditional reconstruction algorithm and deep learning network model,and point out the advantage of deep learning.It is proved theoretically that it is possible to apply the generate adversarial network to the reconstruction of sun speckle image.(2)In this thesis,based on the idea of image to image transformation,the algorithm of single frame solar speckle image reconstruction is studied by using the generate adversarial network(GAN),and the mapping between the speckle image and the highresolution reconstruction image is established by using the cycle-consistency generation adversarial network(CycleGAN).Moreover,the structure of encoder,translator and decoder is used in the generator network,and the authenticity and stability of the reconstructed image are improved through the cycle-consistency.(3)Aiming at the problems of too smooth reconstruction results,lack of details and lack of high-frequency information in single frame speckle image.In this thesis,we introduce perceptual loss to improve the CycleGAN model,which is based on the difference of the features of real human visual perception image in image super-resolution reconstruction.And by using VGG network to extract image features,the difference between image features is calculated instead of the difference between image pixels,which improves the problem of insufficient high-frequency information recovery after reconstruction,and further improves the visual effect of the reconstructed image.In this thesis,we use the original data collected from the 1m New Vacuum Solar Telescope(NVST)of Yunnan Observatory of Chinese Academy of Sciences,and construct the data set after some data preprocessing.The experimental results show that the improved algorithm based on the generate adversarial network combined with the cyclic-consistency loss and the perception loss greatly improves the reconstruction quality of single frame solar speckle image.Under the subjective and objective evaluation criteria,the reconstruction accuracy of this method is not less than the traditional algorithm,and the visual effect and algorithm efficiency are better than the traditional algorithm. |
