A Study Of Sunspot Group Description Based On FCLN

Sunspot groups are the basic signs of solar activities.According to related researches,different types of sunspot groups are closely related to different solar activities.Therefore,scientific detection and caption of sunspot groups can provide a basis for monitoring and predicting solar activities.At present,there are lots of research works in the fields of image caption technology.However,the research on the caption of sunspot groups is still not covered.If the images including sunspot groups can be described in text,astronomers can understand the current status and changes of the sunspot groups intuitively,and understand the impact to the Earth’s climate of sunspot groups.Therefore,it is meaningful to describe the sunspot groups in the full-disk solar image.Aiming at the current situation of zero research on the caption of sunspot groups,and the lack of sufficient contextual relationship between existing image caption methods,the paper based on a Fully Convolutional Localization Network(FCLN)Model,and establishes an end-to-end IRLN(Inception-RPN Localization Network)sunspot groups image caption model.The research contents are as follows:(1)Making the data set of sunspot groupsAiming at the problem that there is no sunspot groups data set,this paper makes a sunspot groups data set.The data set includes two parts: the sunspot groups image data set and the caption text data set.The data source of the former comes from continuous spectrum full-disk solar images provided by the HMI observation device on the SDO satellite.The images are preprocessed,and then the sunspot groups in the full-disk solar image are classified according to the Zurich classification method.The latter is manual tag caption of the classified sunspot groups.Each sunspot group is labeled as a feature caption.After discussing with experts in related fields of the National Observatory and Yunnan Observatory repeatedly,the sunspot groups data set was further modified and improved.(2)Researching on the caption of IRLN sunspot groupsAiming at the characteristics of the sunspot groups image caption task,this paper improves the RPN(Region Proposal Network)network and designs an IRLN sunspot groups image caption model based on the FCLN modle.For the first time,Inception-RPN is applied in image caption.Candidate regions are obtained by using Inception-RPN.The structure of the Inception module is improved according to the feature that the sizes of the sunspot groups are greatly different.The sliding window of the inception module is used to generate candidate regions,which is applied to the caption of sunspot groups.Experimental results show that the m AP of the IRLN model on the VG data set is 6.09%,which is 0.7% higher than the FCLN model;the Meteor is 31.9%,which is 4.6% higher than the FCLN model;and the m AP of the IRLN model on the sunspot groups data set is 74.47%,Which is 16.25% higher than the FCLN model;Meteor is 47.6%,which is 14.2% higher than the FCLN model.In summary,the IRLN model based on the FCLN design in this paper breaks the current situation of zero research on sunspot groups image caption.Meanwhile,the improved Inception-RPN network uses different sizes of sliding windows to slide at each position.This process is constrained by center alignment,which can effectively retain local and contextual information.It effectively improves the lack of sufficient contextual relationship in the existing image caption methods.Therefore,the IRLN sunspot groups image caption model can better describe the sunspot groups.It can enhance the recognition performance of multi-scale sunspot groups.This model further improves the accuracy of describing the sunspot groups.