基于Inception区域候选定位网络的太阳黑子群描述研究

Sunspot groups caption based on Inception-RPN localization network

准确地检测和描述全日面图像中的黑子群可以为监控和预测太阳活动提供依据。目前图像描述技术已有大量研究工作,但关于太阳黑子群描述方面的研究仍未涉及。针对苏黎世分类法中的9类太阳黑子群,制作了一个黑子群图像和描述文本的数据集,设计了一种Inception区域候选定位网络IRLN模型,首次将Inception区域候选网络Inception-RPN应用到图像描述中,通过使用Inception-RPN获得候选区域,并根据黑子群的特征改进了Inception模块的结构,提高网络对多尺度黑子群的检测能力。实验结果表明,本文模型在Visual Genome(VG)数据集上mAP为6.09%,比全卷积定位网络FCLN模型提高了0.7%;Meteor为31.9%,比FCLN模型提高了4.6%。在太阳黑子群数据集上mAP为74.47%,比FCLN模型提高了16%;Meteor为47.6%,比FCLN模型提高了14.2%。

Accurately detecting and describing sunspot groups in full-disk solar images can provide a basis for monitoring and predicting solar activities.At present,there are a lot of research work in the field of image caption technology.However,the research on the caption of sunspot groups is still not covered.For the nine sunspot groups in Zurich classification,a dataset containing the sunspot group images and their caption texts is produced,and an Inception-RPN Localization Network(IRLN)model is designed.For the first time,the Inception-RPN is applied to the image caption.IRLN adopts the Inception-RPN to produce the candidate regions,and improves the structure of Inception module according to the characteristics of sunspot groups,thereby improving the detection ability of multi-scale sunspot groups.Experimental results show that the mAP of the proposed method on Visual Genome(VG)dataset is 6.09%,which is 0.7%higher than the Fully Convolutional Localization Network(FCLN.Its Meteor is 31.9%,which is 4.6%higher than the FCLN method.The mAP of the proposed method on the sunspot groups dataset is 74.47%,which is 16%higher than the FCLN method.Its Meteor is 47.6%,which is 14.2%higher than the FCLN method.