多类类间最大方差算法在图像分割中的应用

Multi-classes Maximum Variance Method Applied in Image Segmentation

介绍了多类类间最大方差法的基本原理,从实际道路图像着手分析该算法存在的两个问题即分类数难以确定以及当分类数增加时其时间复杂度成指数上升.针对确定分类数问题,本文设计了一种直方图均值确定分类数法,其将确定分类数转换成了以图像直方图的概率密度函数的均值为限定条件来求其波峰个数;当分类数大于3类时使用无导数方法对多类类间最大方差法得到局部最优解.建立总量为7209幅图像的道路图像库,对道路图像做实验分析得到确定分类数算法的时间复杂度为O(L)且其平均耗时为0.717毫秒,对图像库抽样计算得到整个算法的平均耗时小于45毫秒.

In this paper, the base principle of multi-classes maximum variance method is introduced, and two key problems are thrown from analyzing the method applied in reality road images. The one is how to confirm the number of classes, and the other is that the time complexity of method is exponential increased with growth of number of classes. To resolve the problem of numbers of classes, a mean value of histogram method is designed. The method converts the problem of confirming number of classes to counting number of peak of image histogram with mean value of probability density function as limiting condition. When the number of classes is more than 3, derivative-free method is used to get locally optimal results. An image databases with 7029 images is built. In this database, the average time cost of confirming number of classes is 0. 717 milliseconds, and the time complexity is O（ L）. By sampling from this database, the average time cost of the whole algorithm is less than 45 milliseconds.