K-means算法和矢量量化

语音信号的数字处理课程作业——矢量量化。这里采用了K-means算法,即假设量化种类是已知的,当然也可以采用LBG算法等,不过K-means比较简单。矢量是二维的,可以在平面上清楚的表示出来。

1. 算法描述

本次实验选择了K-means算法对数据进行矢量量化。算法主要包括以下几个步骤

  • 初始化:载入训练数据,确定初始码本中心(4个);
  • 最近邻分类:对训练数据计算距离(此处采用欧式距离),按照距离最小分类;
  • 码本更新:重新生成包腔对应的质心;
  • 重复分类和码本更新步骤,知道达到最大迭代次数或满足一定停止准则;
  • 利用上述步骤得到的码本对测试数据进行矢量量化,并求最小均方误差。

本实验准备使用MATLAB软件完成矢量量化任务,具体步骤实现如下

    1. 将training.dat和to_be_quantized.dat置于当前工作文件夹内,采用load命令载入training.dat 。
    2. 采用合适的规则选取初始的码本中心。如图 1所示。

图 1 码本中心选择

  1. 计算训练数据和每一码本中心之间的距离。
  2. 采用最近邻准则进行分类。
  3. 重新计算质心,计算公式如下所示。
  4. 重复3~5,直到满足最大迭代次数或是两次迭代结果没有发生改变时,此时结果为训练结果。
  5. 利用训练结果对to_be_quantized.dat进行矢量量化。

2. 代码

MATLAB代码如下

 1 %% training
 2 load(‘training.dat‘);
 3 scatter(training(:,1),training(:,2));
 4 %初始中心选取
 5 x_max = max(training(:,1));
 6 x_min = min(training(:,1));
 7 y_max = max(training(:,2));
 8 y_min = min(training(:,2));
 9 z1 = [(3*x_min+x_max)/4 (3*y_min+y_max)/4];
10 z2 = [(3*x_max+x_min)/4 (3*y_min+y_max)/4];
11 z3 = [(3*x_min+x_max)/4 (3*y_max+y_min)/4];
12 z4 = [(3*x_max+x_min)/4 (3*y_max+y_min)/4];
13 z = [z1;z2;z3;z4];
14 hold on;
15 scatter(z(:,1),z(:,2));
16 legend(‘训练数据‘,‘码本‘);grid on;
17 hold off;
18 for k = 1:20
19     %码本分类,欧式距离
20     distancetoz1 = (training - repmat(z1,size(training,1),1)).^2;
21     distancetoz1 = sum(distancetoz1,2);
22     distancetoz2 = (training - repmat(z2,size(training,1),1)).^2;
23     distancetoz2 = sum(distancetoz2,2);
24     distancetoz3 = (training - repmat(z3,size(training,1),1)).^2;
25     distancetoz3 = sum(distancetoz3,2);
26     distancetoz4 = (training - repmat(z4,size(training,1),1)).^2;
27     distancetoz4 = sum(distancetoz4,2);
28     distance = [distancetoz1 distancetoz2 distancetoz3 distancetoz4];
29     % 分类
30     if(classification == (distance == repmat(min(distance,[],2),1,4)))
31         error = mean(min(distance,[],2));
32         break;      %如果两次迭代之间没有变化,结束迭代
33     end;
34     classification = (distance == repmat(min(distance,[],2),1,4));
35     c1 = training(classification(:,1),:);
36     c2 = training(classification(:,2),:);
37     c3 = training(classification(:,3),:);
38     c4 = training(classification(:,4),:);
39     figure;scatter(c1(:,1),c1(:,2));hold on;scatter(c2(:,1),c2(:,2));
40     scatter(c3(:,1),c3(:,2));scatter(c4(:,1),c4(:,2));
41     legend(‘类型1‘,‘类型2‘,‘类型3‘,‘类型4‘);grid on;hold off;
42     % 码本更新
43     z1 = mean(c1);
44     z2 = mean(c2);
45     z3 = mean(c3);
46     z4 = mean(c4);
47     z = [z1;z2;z3;z4];
48 end
49 %% Test
50 load(‘to_be_quantized.dat‘)
51 distancetoz1 = (to_be_quantized - repmat(z1,size(to_be_quantized,1),1)).^2;
52 distancetoz1 = sum(distancetoz1,2);
53 distancetoz2 = (to_be_quantized - repmat(z2,size(to_be_quantized,1),1)).^2;
54 distancetoz2 = sum(distancetoz2,2);
55 distancetoz3 = (to_be_quantized - repmat(z3,size(to_be_quantized,1),1)).^2;
56 distancetoz3 = sum(distancetoz3,2);
57 distancetoz4 = (to_be_quantized - repmat(z4,size(to_be_quantized,1),1)).^2;
58 distancetoz4 = sum(distancetoz4,2);
59 distance = [distancetoz1 distancetoz2 distancetoz3 distancetoz4];
60 testerror = mean(min(distance,[],2));
61
62 classification = (distance == repmat(min(distance,[],2),1,4));
63 c1 = to_be_quantized(classification(:,1),:);
64 c2 = to_be_quantized(classification(:,2),:);
65 c3 = to_be_quantized(classification(:,3),:);
66 c4 = to_be_quantized(classification(:,4),:);
67 figure;scatter(c1(:,1),c1(:,2));hold on;scatter(c2(:,1),c2(:,2));
68 scatter(c3(:,1),c3(:,2));scatter(c4(:,1),c4(:,2));
69 legend(‘类型1‘,‘类型2‘,‘类型3‘,‘类型4‘);grid on;hold off;

3. 实验结果

图 2 训练码本分布

图 3第一次迭代结果                   图 4第四次迭代结果

图 5第八次迭代结果                  图 6第九次迭代结果

图 2展示了训练数据的分布,图 3~6是迭代过程中分类的变化情况,迭代完成后的码本为

  • Z1 = [1.62060631541935 -0.108624145483871]
  • Z2 = [7.96065094375000 -0.999061308437500]
  • Z3 = [1.72161941468750 6.82121444062500]
  • Z4 = [4.43652765757576 2.18874305151515]

4. 实验数据

training.dat

  1   8.4416189e+000 -7.9885975e-001
  2   1.1480908e+000  7.8735044e+000
  3   7.7380144e+000 -1.2165061e+000
  4   8.9727144e-001  7.3962468e+000
  5   7.5343823e+000 -1.1424504e+000
  6  -6.9234039e-001 -1.7096610e+000
  7   7.6418740e+000 -1.3563792e+000
  8   3.1091418e+000  6.3850541e+000
  9   2.3482174e+000  4.7553506e-001
 10  -1.3840364e+000 -2.5480394e+000
 11   8.2008897e+000 -1.1448387e+000
 12  -1.1392497e+000 -2.0809884e+000
 13   3.7970116e+000  1.6906469e+000
 14   3.4484200e+000  1.3980911e+000
 15   2.5701485e+000  5.3755044e+000
 16   8.3899076e+000 -6.6675309e-001
 17   2.0146545e+000  5.6984592e+000
 18   1.8853328e+000  5.2762628e-001
 19   5.6781432e+000  3.2588691e+000
 20   1.0102480e+000  5.8167707e+000
 21   7.7302763e+000 -1.2030348e+000
 22   4.2118845e+000  1.6527181e+000
 23   4.3920049e-001  6.7168970e+000
 24   8.1934984e-001 -5.1917945e-001
 25   4.3708769e+000  2.1613573e+000
 26   1.8569681e+000  4.8380565e+000
 27   3.4732504e+000  1.7953635e+000
 28   7.5822756e+000 -1.1521814e+000
 29   2.6434078e+000  6.3295690e+000
 30   1.9968582e+000  7.3529314e+000
 31   4.0833513e+000  1.4936002e+000
 32   3.6767894e+000  6.7446912e+000
 33   1.3524515e+000  6.8177858e+000
 34   3.9711504e+000  1.5452503e+000
 35   1.5594711e+000  6.3885281e+000
 36   3.4692089e+000  1.7118124e+000
 37   5.2575491e+000  2.5601553e+000
 38   7.8827882e+000 -6.8867840e-001
 39   4.8176593e+000  2.1684005e+000
 40   2.7402486e+000  8.3320174e+000
 41   2.2549011e+000  3.9393641e-001
 42   8.0840542e+000 -7.3155184e-001
 43   8.8753667e-001  6.1607892e+000
 44   1.8067727e+000 -2.1099454e-001
 45   6.8650914e+000  4.4228389e+000
 46   6.4174056e+000  3.7590081e+000
 47   4.0933273e+000  1.3598676e+000
 48   2.2882999e+000  5.1876795e-001
 49   7.9225523e+000 -1.1725456e+000
 50   4.3561335e+000  1.8976163e+000
 51   8.3279098e+000 -1.0232899e+000
 52   6.2551331e+000  3.3449949e+000
 53   3.1276024e+000  7.8463356e-001
 54   6.5241605e+000  3.4561490e+000
 55   4.1588140e-001  6.4974858e+000
 56   2.7379263e+000  6.4746080e+000
 57   7.2185639e+000 -1.3525589e+000
 58   7.5424890e+000 -1.5317814e+000
 59   3.7468423e+000  1.6110753e+000
 60   8.8708536e+000 -5.6439331e-001
 61   7.6960713e+000 -1.1960633e+000
 62   7.5979552e+000 -1.1469059e+000
 63   2.8220978e+000  1.0360184e+000
 64   3.8165165e+000  1.6082223e+000
 65   6.6799248e-002 -1.2910367e+000
 66   2.3054028e+000  2.8450986e-001
 67   4.2788715e+000  5.1995858e+000
 68   3.0006534e+000  9.1250414e-001
 69   7.6051326e+000 -1.1005476e+000
 70   2.5331653e+000  9.7428007e-001
 71   1.0743104e+000  6.0859296e+000
 72   6.7237149e-001  8.6117274e+000
 73   2.4333003e+000  7.1421389e-001
 74   1.7723473e+000  7.1841833e+000
 75   3.5762796e+000  1.5348648e+000
 76   2.7863558e+000  7.3565043e-001
 77   8.0284284e+000 -7.9636983e-001
 78   8.4672682e+000 -8.2062254e-001
 79   2.3519727e+000  8.1632796e-001
 80   7.4240720e+000  4.1800229e+000
 81   1.9724319e+000  4.4328699e-001
 82   7.7622621e+000 -1.3506605e+000
 83   2.3793018e+000 -4.3107386e-001
 84   3.2455220e+000  1.2697488e+000
 85   1.3644859e+000  5.9712644e+000
 86   5.4815655e+000  2.6608754e+000
 87  -1.2002073e+000 -2.1765731e+000
 88  -3.5558595e-001  6.4387512e+000
 89   3.9418185e+000  1.9858047e+000
 90   1.0533626e+000 -7.9068285e-001
 91   1.9560213e+000  6.2001316e+000
 92   7.5555203e+000 -1.2087337e+000
 93   1.7851705e+000  7.0073148e+000
 94   2.2736274e+000  7.9336349e-001
 95   7.6615799e+000 -1.0445564e+000
 96   2.7181608e+000  4.7615418e-001
 97   1.8291149e+000 -6.7261971e-001
 98   7.8640867e+000 -1.4296092e+000
 99   2.6362814e+000  5.8303048e-001
100   3.7771102e+000  1.2928196e+000
101   7.5360359e+000 -9.7942712e-001
102   4.0257498e+000  1.2217666e+000
103   8.4500853e+000 -7.6599648e-001
104   3.0488646e+000  6.2159289e+000
105   2.0954150e+000  2.5848825e-001
106   1.6592148e+000  7.5650162e+000
107   3.5535363e+000  1.3326217e+000
108   4.3388636e+000  2.1235893e+000
109   3.1233524e+000  1.3971470e+000
110   7.6317385e+000 -1.0744610e+000
111   8.5028402e-001 -3.2822876e-001
112   8.6903131e+000 -2.6843242e-001
113   4.4418011e+000  2.5676053e+000
114   2.5119872e+000 -1.0521242e-001
115   1.9613752e+000  7.0072931e+000
116   3.2607143e+000  1.5432286e+000
117   3.2830401e+000  1.0228031e+000
118   8.0201528e+000 -7.0827461e-001
119   3.1597313e+000  7.6750043e+000
120   9.0059933e+000 -9.6130246e-001
121   1.1037820e+000 -1.2980812e-001
122   1.5334911e+000  7.4282719e+000
123   6.0948533e-001  6.3861341e+000
124   4.0065706e-001 -1.1015776e+000
125   2.3451558e+000  8.6384057e+000
126   1.4490876e+000  8.6646066e+000
127   8.0421821e+000 -8.1100509e-001
128   8.0175747e+000 -5.6119093e-001

to_be_quantized

 1   3.7682247e+000  8.3609865e-001
 2   2.6963398e+000  6.5766226e-001
 3   3.3438207e+000  1.2495321e+000
 4   1.3646195e+000 -6.3947640e-001
 5   7.8227583e+000 -8.8616996e-001
 6   1.3532508e+000  7.6607304e+000
 7   2.2741739e+000  6.9387226e+000
 8   3.5361382e+000  5.9729821e+000
 9   8.0409138e+000 -1.1234886e+000
10   7.9630460e+000 -1.3032200e+000
11   2.3478158e+000  6.9759690e+000
12   3.2632942e+000  1.5675470e+000
13   1.5241488e+000  7.1053147e+000
14   5.7320838e+000  3.4042655e+000
15   2.3339411e+000  6.9428434e+000
16   6.5330392e+000  3.4415860e+000
17   3.1068803e+000  8.0080363e+000
18   7.4078126e+000 -1.3416027e+000
19   1.9925474e+000 -2.7782790e-001
20   5.0187915e+000  2.7058427e+000
21   2.6535497e-001 -1.2622069e+000
22   1.4960584e+000  6.3355004e+000
23   3.1933474e-001  7.1467466e+000
24   8.2821020e+000 -9.5178778e-001
25   2.5653586e+000  6.9836115e+000
26   3.6937139e+000  1.1535671e+000
27   8.5390043e+000 -5.0678923e-001
28   7.5436898e-001 -6.7669379e-001
29   2.1638213e+000  7.6142401e+000
30   4.8522826e+000  2.7079076e+000
31   5.4890641e+000  3.3875394e+000
32   4.2525899e+000  1.8861744e+000
33   8.4088615e+000 -1.1920963e+000
34   5.5396960e+000  2.9680110e+000
35   3.3334381e+000  1.4384861e+000
36   3.5212919e+000  1.0327602e+000
37   4.6303492e+000  2.1627805e+000
38   3.9385929e+000  1.0010804e+000
39   8.4553633e+000 -7.2297277e-001
40   1.8111095e+000  7.6132396e+000
41   1.1240984e+000 -2.7029879e-001
42  -3.3840083e-002 -1.5590834e+000
43   7.1674870e+000 -1.5449905e+000
44   8.5103026e+000 -9.8820393e-001
45   7.7529857e+000 -1.4787432e+000
46   1.8704913e+000  6.9370116e+000
47   6.0271939e+000  3.2118915e+000
48   2.8287461e+000  7.3399383e+000
49   4.1568876e+000  1.5631238e+000
50   8.2187067e-001 -5.8546437e-001
51   3.1084965e+000  5.3512449e+000
52   4.1581386e+000  2.1763345e+000
53   3.2267474e+000  1.4105815e+000
54   8.1564752e-001  7.2540175e+000
55   8.0241402e+000 -8.2411742e-001
56   6.2773554e+000  3.1729045e+000
57   8.5460058e+000 -1.0330056e+000
58   8.6215210e+000 -7.4057378e-001
59   7.4872291e+000 -1.0113921e+000
60   3.3155133e+000  9.7636038e-001
61   2.1051593e+000  3.4894654e-001
62   3.6776134e+000  1.5387928e+000
63   2.9009105e+000  5.6931589e+000
64   8.0567164e+000 -1.0000803e+000

时间: 2024-01-09 02:18:01

K-means算法和矢量量化的相关文章

K-means算法

K-means算法很简单,它属于无监督学习算法中的聚类算法中的一种方法吧,利用欧式距离进行聚合啦. 解决的问题如图所示哈:有一堆没有标签的训练样本,并且它们可以潜在地分为K类,我们怎么把它们划分呢?     那我们就用K-means算法进行划分吧. 算法很简单,这么做就可以啦: 第一步:随机初始化每种类别的中心点,u1,u2,u3,--,uk; 第二步:重复以下过程: 然后 ,就没有然后了,就这样子. 太简单, 不解释.

矢量量化(VQ)

作者:桂. 时间:2017-05-31  21:14:56 链接:http://www.cnblogs.com/xingshansi/p/6925955.html 前言 VQ(Vector Quantization)是一个常用的压缩技术,本文主要回顾: 1)VQ原理 2)基于VQ的说话人识别(SR,speaker recognition)技术 〇.分类问题 说话人识别其实也是一个分类问题: 说话人识别技术,主要有这几大类方法: 模板匹配方法 这类方法比较成熟,主要原理:特征提取.模板训练.匹配.

从K近邻算法、距离度量谈到KD树、SIFT+BBF算法

从K近邻算法.距离度量谈到KD树.SIFT+BBF算法 从K近邻算法.距离度量谈到KD树.SIFT+BBF算法 前言 前两日,在微博上说:“到今天为止,我至少亏欠了3篇文章待写:1.KD树:2.神经网络:3.编程艺术第28章.你看到,blog内的文章与你于别处所见的任何都不同.于是,等啊等,等一台电脑,只好等待..”.得益于田,借了我一台电脑(借他电脑的时候,我连表示感谢,他说“能找到工作全靠你的博客,这点儿小忙还说,不地道”,有的时候,稍许感受到受人信任也是一种压力,愿我不辜负大家对我的信任)

K近邻算法

1.1.什么是K近邻算法 何谓K近邻算法,即K-Nearest Neighbor algorithm,简称KNN算法,单从名字来猜想,可以简单粗暴的认为是:K个最近的邻居,当K=1时,算法便成了最近邻算法,即寻找最近的那个邻居.为何要找邻居?打个比方来说,假设你来到一个陌生的村庄,现在你要找到与你有着相似特征的人群融入他们,所谓入伙. 用官方的话来说,所谓K近邻算法,即是给定一个训练数据集,对新的输入实例,在训练数据集中找到与该实例最邻近的K个实例(也就是上面所说的K个邻居),这K个实例的多数属

DM里的K均值算法

1.Preface 因为一直在做的是聚类算法的研究,算是总结了一些心得,这里总结些知识性与思路性的东西,我想在其他地方也是很容易的找到类似的内容的.毕竟,世界就是那么小. 声明:本文比较不适合没有DM基础的人来阅读.我只是胡乱的涂鸦而已 2.聚类算法 在DM里的聚类算法里,有基于划分的算法,基于层次的算法,基于密度的算法,基于网格的算法,基于约束的算法. 其中每一种基于的算法都会衍生出一至几种算法,对应的每一种算法不管在学术界还是工业界都存在着许多的改进的算法 这里想介绍的是基于基于划分的算法里

k均值算法

import matplotlib.pyplot as plt import numpy as np import time from django.template.defaultfilters import center def loadDataSet(fileName): dataMat=[] fr=open(fileName) for line in fr.readlines(): curLine=line.strip().split('\t') fltLine=map(float,cu

『cs231n』作业1问题1选讲_通过代码理解K近邻算法&交叉验证选择超参数参数

通过K近邻算法探究numpy向量运算提速 茴香豆的"茴"字有... ... 使用三种计算图片距离的方式实现K近邻算法: 1.最为基础的双循环 2.利用numpy的broadca机制实现单循环 3.利用broadcast和矩阵的数学性质实现无循环 图片被拉伸为一维数组 X_train:(train_num, 一维数组) X:(test_num, 一维数组) 方法验证 import numpy as np a = np.array([[1,1,1],[2,2,2],[3,3,3]]) b

K 近邻算法

声明: 1,本篇为个人对<2012.李航.统计学习方法.pdf>的学习总结,不得用作商用,欢迎转载,但请注明出处(即:本帖地址). 2,因为本人在学习初始时有非常多数学知识都已忘记,所以为了弄懂当中的内容查阅了非常多资料.所以里面应该会有引用其它帖子的小部分内容,假设原作者看到能够私信我,我会将您的帖子的地址付到以下. 3.假设有内容错误或不准确欢迎大家指正. 4.假设能帮到你.那真是太好了. 描写叙述 给定一个训练数据集,对新的输入实例.在训练数据集中找到与该实例最邻近的K个实例,若这K个实

聚类算法:K-means 算法(k均值算法)

k-means算法:      第一步:选$K$个初始聚类中心,$z_1(1),z_2(1),\cdots,z_k(1)$,其中括号内的序号为寻找聚类中心的迭代运算的次序号. 聚类中心的向量值可任意设定,例如可选开始的$K$个模式样本的向量值作为初始聚类中心.      第二步:逐个将需分类的模式样本$\{x\}$按最小距离准则分配给$K$个聚类中心中的某一个$z_j(1)$.假设$i=j$时, \[D_j (k) = \min \{ \left\| {x - z_i (k)} \right\|