基于全局特征的极光图像分类

Aurora Image Classification Based on Global Features

摘要: 极光分类对于太阳活动对地球的影响方式的研究具有非常重要的意义，现有的极光分类方法主要是基于极光图像的局部特征，这些局部特征对噪声干扰及极光图像的位置、方向等变化较为敏感，很难满足实际应用的要求。本文提出了一种新的基于全局特征的极光图像分类方法，在该方法中，极光图像通过Radon变换投影到Radon域，然后计算投影矩阵中每列的方差作为特征，为了实现方向变化不变性，对该方差序列进行循环移位使得该序列方差最大的值居于首位，进行旋转归一化处理，然后，应用基于欧氏距离的最近邻分类器实现极光图像分类。实验表明，该方法分类精度明显高于现有的基于局部特征的分类方法，对噪声干扰及极光图像位置、方向的鲁棒性显著高于现有方法，而且计算效率也高于基于局部特征的分类方法。

Abstract: Aurora classification is of great significance to the research on the way and degree of the influence of solar activity on the earth. The existing methods for aurora image classification are mainly based on the local features extracted from the aurora images. These local features are sensitive to noise and variances of the position and orientation, so their classification accuracies and robustness are insufficient for complicated applications. This paper proposed a novel aurora image classification method based on the global feature descriptors. In the proposed method, the aurora image is projected to Radon domain via Radon transform, and then, the variances of columns are determined, the rotation invariant features are obtained via circular shift operation on the variance sequence to let the maximum value in the first place, which completes the rotation normalization of the variance sequence. A nearest neighbor classifier based on Euclidean distance is used for classification. Experimental results show that the proposed approach yields a better performance in terms of the correct classification percentages compared with the aurora image classification method based on representative local feature. It is also shown that the proposed approach yields observably low computational cost and relatively high robustness to noise, the variations of orientation and position of aurora images.