基于多核学习的典型地物目标特征描述

Feature Description of Typical Ground Objects Based on Multiple Kernel Learning

为了寻找针对典型地物目标进行描述的最优特征组合,提出了一种基于多核学习的地物目标的特征描述分析方法.首先,提取图像的纹理特征、颜色特征等多种低层次特征.然后,采用多核学习算法训练出多核分类模型.最后,从特征权重矩阵中抽取对应于特定类型地物目标的特征向量,按权重大小进行降序排序,从而找出适用于特定地物目标描述的最优特征组合.此外,还提出了一种综合排序方法来计算全部特征在所有地物类别上的权重排列顺序.结果表明,对于不同类型的地物目标适合于描述它们的最优特征也不同,利用多核学习算法配置多种特征来描述图像要远远优于使用单独特征来描述图像.

In order to find the optimal feature combination for description of typical ground objects, a fea- ture description analysis method based on multi kernel learning was proposed in this paper. First, some low level features of images, e. g. , texture features and color features, were extracted. Then, a multi kernel classification model was trained by using the multi kernel learning algorithm. Finally, the feature vec- tors of some typical ground objects were extracted from the feature weight matrix, and the feature vectors were ranked in order of descending weights, thereby the optimal features combination suitable to describe typical ground ohiects were found out. In addition, a comprehensive ranking method was also proposed to compute the sort order of all the features on all ground object categories. The experimental results show that the optimal feature combination are different for different types of ground objects, and using the multi kernel learning algorithm to configurate the features to describe images are superior to using one single feature to describe images.