一种新的连续特征量化方法

对连续特征进行有效量化是水下目标分类中有待解决的一个重要问题。本文提出一种加权距离量化方法。该量化方法使用类别相对频率构造了两相邻区间的加权距离,将加权距离作为特征量化标准,在量化过程中,将加权距离最小的相邻区间进行合并,直到满足终止条件为止。文中使用递归最小信息熵、Chi2、加权距离等五种量化算法对27维水下目标的识别特征进行了量化处理,比较了各量化方法的性能。结果表明,使用加权距离量化算法对水下目标的识别特征进行量化处理之后,所产生的量化区间数目较少,量化时间较短,量化数据较好的保持了原数据的分类能力,且量化数据的分类时间也大大缩短。

A New Discretization Method for Continuous Features

The development of an effective discretization algorithm for continuous features is an important problem to be solved in underwater targets classification. A weighted-distance based discretization method is proposed in this paper. The weighted distance between two adjacent intervals is defined using relative class frequency in the algorithm. And the two adjacent intervals with minimum weighted distance are merged until some stopping criterion is achieved. 27 features of underwater targets are discretized using recursive entropy minimization, Chi2 and the weighted-distance based discretization algorithm, and properties of each algorithm are compared. The comparison results demonstrate that the weighted-distance based algorithm can obtain fewer intervals in less time. In addition, the classification time is reduced greatly while maintaining the classification capability of the original datasets.