基于作者主题模型的遥感图像自动类别标注方法

摘要

以概率图模型为基础，提出一种基于作者主题模型ATM（Author Topic Model）的多光谱遥感图像类别标注方法。该方法采用了一种新的基于颜色和形状特征的描述符，并结合ATM对遥感图像进行类别标注。首先采用一组定义了语义鸿沟的图像作为训练图像，然后采用基于颜色和形状特征的视觉单词描绘训练图像，最后结合ATM对遥感图像进行类别标注。通过对实际的遥感图像进行类别标注验证，可以看出，所提出的基于ATM的遥感图像标注方法在区域类别较少的情况下具有较高的分类准确率。%In this paper,we propose a new method of category tagging for multispectral remote sensing images taking the probability plot as the basis,the method is based on author-topic model (ATM).It adopts a new colour and shape feature-based descriptor,and integrates ATMto perform category tagging on remote sensing images.First,the method uses a set of images with defined semantic gap as the training images. Then,it uses colour and shape feature-based visual words to describe the training images.Finally,it conducts category tagging on remote sensing images in combination with ATM.Through the verification of category tagging on actual remote sensing images,it is demonstrated that the ATM-based remote sensing image tagging method has higher classification accuracy under the condition of less regional categories.