基于改进卷积神经网络的极光图像分类算法研究

Aurora image classification algorithm based on improved convolution neural network

极光包含丰富的磁层和日地电磁活动以及能量耦合等空间物理信息,是一种自然放光现象。对极光图像的正确分类有助于探索太阳与地球及地球自身磁场的奥秘。文中针对极光图像分类问题提出一种基于神经网络改进的算法,首先采用迁移学习将在大规模数据集上训练过的VGG16网络用于极光数据库,然后结合VGG16和密集连接的思想提出一种改进的Dense-VGG网络,用该网络提取极光图像的特征,并实现极光图像的自动分类。在中国北极黄河站拍摄的两个极光数据库上进行了实验,其中8 001幅准确率达到96.54%,38 044幅准确率达到98.99%,证明该算法能有效提高极光图像分类准确率。

Aurora contains abundant spatial physical information,such as magnetosphere,solar-geomagnetic activities and energy coupling,is a natural luminescence phenomenon.The correct classification of aurora images is helpful for people to explore the mysteries of the magnetic fields of the sun and the Earth,as well as the Earth itself.An improved algorithm based on the neural network is proposed for aurora image classification.Firstly,the VGG16 network trained on large-scale datasets is used for the aurora database by migration learning.Then,an improved Dense-VGG network is put forward by combining VGG16 and dense connection.The network is used to extract aurora image features and realize automatic aurora image classification.Experiments have been carried out in two aurora databases taken at the Yellow River Station in the Arctic of China.Experimental results show that the accuracy of 8 001 aurora images is 96.54%,and that of 38 044 aurora images is 98.99%.It is proved that the algorithm can effectively improve the accuracy of aurora image classification.