本文共 5188 字,大约阅读时间需要 17 分钟。
高斯混合模型(GMM)在图像分割、对象识别、视频分析等方面均有应用,对于任意给定的数据样本集合,根据其分布概率, 可以计算每个样本数据向量的概率分布,从而根据概率分布对其进行分类,但是这些概率分布是混合在一起的,要从中分离出单个样本的概率分布就实现了样本数据聚类,而概率分布描述我们可以使用高斯函数实现,这个就是高斯混合模型-GMM。
这种方法也称为D-EM即基于距离的期望最大化。
1. 初始化变量定义-指定的聚类数目K与数据维度D
2. 初始化均值、协方差、先验概率分布
3. 迭代E-M步骤
- E步计算期望
- M步更新均值、协方差、先验概率分布
-检测是否达到停止条件(最大迭代次数与最小误差满足),达到则退出迭代,否则继续E-M步骤
4. 打印最终分类结果
package com.gloomyfish.image.gmm; import java.util.ArrayList; import java.util.Arrays; import java.util.List; /** * * @author gloomy fish * */ public class GMMProcessor { public final static double MIN_VAR = 1E-10; public static double[] samples = new double[]{10, 9, 4, 23, 13, 16, 5, 90, 100, 80, 55, 67, 8, 93, 47, 86, 3}; private int dimNum; private int mixNum; private double[] weights; private double[][] m_means; private double[][] m_vars; private double[] m_minVars; /*** * * @param m_dimNum - 每个样本数据的维度, 对于图像每个像素点来说是RGB三个向量 * @param m_mixNum - 需要分割为几个部分,即高斯混合模型中高斯模型的个数 */ public GMMProcessor(int m_dimNum, int m_mixNum) { dimNum = m_dimNum; mixNum = m_mixNum; weights = new double[mixNum]; m_means = new double[mixNum][dimNum]; m_vars = new double[mixNum][dimNum]; m_minVars = new double[dimNum]; } /*** * data - 需要处理的数据 * @param data */ public void process(double[] data) { int m_maxIterNum = 100; double err = 0.001; boolean loop = true; double iterNum = 0; double lastL = 0; double currL = 0; int unchanged = 0; initParameters(data); int size = data.length; double[] x = new double[dimNum]; double[][] next_means = new double[mixNum][dimNum]; double[] next_weights = new double[mixNum]; double[][] next_vars = new double[mixNum][dimNum]; ListcList = new ArrayList (); while(loop) { Arrays.fill(next_weights, 0); cList.clear(); for(int i=0; i 1E-20) ? Math.log10(p) : -20; } currL /= size; // Re-estimation: generate new weight, means and variances. for (int j = 0; j < mixNum; j++) { weights[j] = next_weights[j] / size; if (weights[j] > 0) { for (int d = 0; d < dimNum; d++) { m_means[j][d] = next_means[j][d] / next_weights[j]; m_vars[j][d] = next_vars[j][d] / next_weights[j] - m_means[j][d] * m_means[j][d]; if (m_vars[j][d] < m_minVars[d]) { m_vars[j][d] = m_minVars[d]; } } } } // Terminal conditions iterNum++; if (Math.abs(currL - lastL) < err * Math.abs(lastL)) { unchanged++; } if (iterNum >= m_maxIterNum || unchanged >= 3) { loop = false; } } // print result System.out.println("=================最终结果================="); for(int i=0; i max) { max = v; types[k] = i; } } } double[] counts = new double[mixNum]; for(int i=0; i 0) { for (int d = 0; d < dimNum; d++) { m_vars[i][d] = m_vars[i][d] / counts[i]; // A minimum variance for each dimension is required. if (m_vars[i][d] < m_minVars[d]) { m_vars[i][d] = m_minVars[d]; } } } } System.out.println("=================初始化================="); for(int i=0; i PDF * @param x - 表示采样数据点向量 * @param j - 表示对对应的第J个分类的概率密度分布 * @return - 返回概率密度分布可能性值 */ public double getProbability(double[] x, int j) { double p = 1; for (int d = 0; d < dimNum; d++) { p *= 1 / Math.sqrt(2 * 3.14159 * m_vars[j][d]); p *= Math.exp(-0.5 * (x[d] - m_means[j][d]) * (x[d] - m_means[j][d]) / m_vars[j][d]); } return p; } public static void main(String[] args) { GMMProcessor filter = new GMMProcessor(1, 2); filter.process(samples); } }
结构类DataNode
package com.gloomyfish.image.gmm; public class DataNode { public int cindex; // cluster public int index; public double[] value; public DataNode(double[] v) { this.value = v; cindex = -1; index = -1; } }
这里初始中心均值的方法我是通过随机数来实现,GMM算法运行结果跟初始化有很大关系,常见初始化中心点的方法是通过K-Means来计算出中心点。大家可以尝试修改代码基于K-Means初始化参数,我之所以选择随机参数初始,主要是为了省事!