基于多通道和卷积神经网络的极光分类-《计算机技术与发展》

文章信息/Info

Title:: Aurora Images Classification Based on Multi－channel Fusion and Convolutional Neural Network

Author(s):: CHEN Chang－hong; LIU Bin; ZHANG Hao; Key Laboratory on Image Processing ＆ Image Communications of Jiangsu Province，School of Telecommunication and Information Engineering，Nanjing University of Posts and Telecommunications，Nanjing 210003，China

Keywords:: multi－channel fusion; convolutional neural network; deep learning; pre－training; classification of aurora images

摘要:: 目前,极光图像分类领域多采用传统特征来进行分类.但传统特征提取图像的某种特定特征,比如纹理特征、局部特征、全局特征等,导致或多或少丢失极光图像的某些重要分类信息,造成分类效果不够好.对此,提出一种基于多通道融合和卷积神经网络的极光图像分类方法.采用多通道融合技术将原图信息和指定有效传统特征信息加以融合形成融合图像,利用预训练卷积神经网络自动提取融合图像的有效特征信息,实现多通道特征与深度学习相结合,得到高效表征极光图像的特征.在2003年北极黄河站越冬观测的4种日侧极光图像数据库上进行实验.在8 001幅典型极光图像数据库上,与人工标记对比分类准确率高达95.2％,高于其他同类方法.实验结果表明该方法能有效用于极光图像分类.

Abstract:: At present，the traditional feature is used for classification in the field of aurora image classification． However，the certain features of image extracted by traditional feature，such as texture feature，local and global feature，leads to more or less loss of some important classification information of aurora images，resulting in a poor classification effect． For this，we propose an aurora image classification method based on multi－channel fusion and convolutional neural network． The multi－channel fusion technology is used to fuse the original image information with the designated effective traditional feature information to form the fusion image． The convolutional neural network is used to automatically extract the effective feature information of the fusion image，to realize the combination of multi－channel features and deep learning，and to obtain the features of efficiently representing the aurora image． The experiment is carried out on the four kinds of daily aurora image databases in the North Pole of the Yellow River station in 2003，which shows that the classification accuracy in comparison with manual markers in the classic database 8，001 aurora images is as high as 95． 2%，higher than other similar methods． The results show that this method can be used in the classification of aurora images．