对于图象中不同类别（以其统计分布函数一表征）的可识别性，分离度的统计定义可做出定量的描述。根据分离度在一对应映射变换下的不变性，应用Karhunen-Loeve变换对两类分布撮识别特征，发现分离度只依赖于特征值最大和最小的两个特征向量。而且，分离度的大小依赖于特征值怀某个定定值的偏差。由此我们提出一个识别模型，使每次分类后的样本集的特征值都趋近于两上定值，从而得到最佳的识别效果。由该模型设计的

The separability of two pattern classes can be measured by the divergence for two Gaussian distributions. Because the divergence is invariant under the linear transformation we can extract "good" features for separating two patterns via Karhunen-Loeve transformation. It is shown that the divergence is only dependent on two of n eigenvalues. One property of a "good" dichotomy is that each feature should be effective for classification. Thus, a criterion function is proposed. This algorithm, which is called as sample exchange algorithm, is convergent and it is a reasonable unsupervised clustering method for classification.