您的位置:首页 > 其它

图像相似度(测试)--基于直方图特征的图像搜索

2013-04-09 14:19 477 查看
转自:http://blog.csdn.net/jia20003/article/details/7771651#comments

图像处理之相似图片识别(直方图应用篇)



算法概述:

首先对源图像与要筛选的图像进行直方图数据采集,对采集的各自图像直方图进行归一化再

使用巴氏系数算法对直方图数据进行计算,最终得出图像相似度值,其值范围在[0, 1]之间

0表示极其不同,1表示极其相似(相同)。

算法步骤详解:

大致可以分为两步,根据源图像与候选图像的像素数据,生成各自直方图数据。第二步:使

用第一步输出的直方图结果,运用巴氏系数(Bhattacharyya coefficient)算法,计算出相似程

度值。

第一步:直方图计算

直方图分为灰度直方图与RGB直方图,对于灰度图像直方图计算十分简单,只要初始化一

个大小为256的直方图数组H,然后根据像素值完成频率分布统计,假设像素值为124,则

H[124] += 1, 而对于彩色RGB像素来说直方图表达有两种方式,一种是单一直方图,另外一

种是三维直方图,三维直方图比较简单明了,分别对应RGB三种颜色,定义三个直方图HR,

HG, HB, 假设某一个像素点P的RGB值为(4, 231,129), 则对于的直方图计算为HR[4] += 1,

HG[231] += 1, HB[129] += 1, 如此对每个像素点完成统计以后,RGB彩色直方图数据就生成了。

而RGB像素的单一直方图SH表示稍微复杂点,每个颜色的值范围为0 ~ 255之间的,假设

可以分为一定范围等份,当8等份时,每个等份的值范围为32, 16等份时,每个等份值范

围为16,当4等份时候,每个等份值的范围为64,假设RGB值为(14, 68, 221), 16等份之

后,它对应直方图索引值(index)分别为: (0, 4, 13), 根据计算索引值公式:index = R + G*16 + B*16*16

对应的直方图index = 0 + 4*16 + 13 * 16 * 16, SH[3392] += 1

如此遍历所有RGB像素值,完成直方图数据计算。

第二步:巴氏系数计算,计算公式如下:






其中P, P’分别代表源与候选的图像直方图数据,对每个相同i的数据点乘积开平方以后相加

得出的结果即为图像相似度值(巴氏系数因子值),范围为0到1之间。

程序效果:



相似度超过99%以上,极其相似






相似度为:72%, 一般相似

程序直方图计算源代码如下:

[java]
view plaincopyprint?

public void setGreenBinCount(int greenBinCount) { 
    this.greenBins = greenBinCount; 

 
public void setBlueBinCount(int blueBinCount) { 
    this.blueBins = blueBinCount; 


 
public float[] filter(BufferedImage src, BufferedImage dest) { 
    int width = src.getWidth(); 

       int height = src.getHeight(); 
        
       int[] inPixels =
new int[width*height]; 
       float[] histogramData =
new float[redBins * greenBins * blueBins]; 
       getRGB( src, 0,
0, width, height, inPixels ); 
       int index = 0; 
       int redIdx =
0, greenIdx = 0, blueIdx =0; 

       int singleIndex = 0; 
       float total =
0; 
       for(int row=0; row<height; row++) { 
        int ta =
0, tr = 0, tg =0, tb =
0; 
        for(int col=0; col<width; col++) { 
            index = row * width + col; 
            ta = (inPixels[index] >> 24) &0xff; 

               tr = (inPixels[index] >>
16) & 0xff; 
               tg = (inPixels[index] >> 8) &0xff; 

               tb = inPixels[index] &
0xff; 
               redIdx = (int)getBinIndex(redBins, tr,255); 

               greenIdx = (int)getBinIndex(greenBins, tg,255); 

               blueIdx = (int)getBinIndex(blueBins, tb,255); 

               singleIndex = redIdx + greenIdx * redBins + blueIdx * redBins * greenBins; 
               histogramData[singleIndex] += 1; 
               total += 1; 
        } 
       } 
        
       // start to normalize the histogram data 
       for (int i =0; i < histogramData.length; i++) 

       { 
        histogramData[i] = histogramData[i] / total; 
       } 
        
       return histogramData; 


public void setGreenBinCount(int greenBinCount) {
this.greenBins = greenBinCount;
}

public void setBlueBinCount(int blueBinCount) {
this.blueBins = blueBinCount;
}

public float[] filter(BufferedImage src, BufferedImage dest) {
int width = src.getWidth();
int height = src.getHeight();

int[] inPixels = new int[width*height];
float[] histogramData = new float[redBins * greenBins * blueBins];
getRGB( src, 0, 0, width, height, inPixels );
int index = 0;
int redIdx = 0, greenIdx = 0, blueIdx = 0;
int singleIndex = 0;
float total = 0;
for(int row=0; row<height; row++) {
int ta = 0, tr = 0, tg = 0, tb = 0;
for(int col=0; col<width; col++) {
index = row * width + col;
ta = (inPixels[index] >> 24) & 0xff;
tr = (inPixels[index] >> 16) & 0xff;
tg = (inPixels[index] >> 8) & 0xff;
tb = inPixels[index] & 0xff;
redIdx = (int)getBinIndex(redBins, tr, 255);
greenIdx = (int)getBinIndex(greenBins, tg, 255);
blueIdx = (int)getBinIndex(blueBins, tb, 255);
singleIndex = redIdx + greenIdx * redBins + blueIdx * redBins * greenBins;
histogramData[singleIndex] += 1;
total += 1;
}
}

// start to normalize the histogram data
for (int i = 0; i < histogramData.length; i++)
{
histogramData[i] = histogramData[i] / total;
}

return histogramData;
}
计算巴氏系数的代码如下:

[java]
view plaincopyprint?

/**
* Bhattacharyya Coefficient
* http://www.cse.yorku.ca/~kosta/CompVis_Notes/bhattacharyya.pdf *
* @return
*/ 
public double modelMatch() { 
    HistogramFilter hfilter = new HistogramFilter(); 
    float[] sourceData = hfilter.filter(sourceImage,null); 

    float[] candidateData = hfilter.filter(candidateImage,null); 

    double[] mixedData =new
double[sourceData.length]; 
    for(int i=0; i<sourceData.length; i++ ) { 
        mixedData[i] = Math.sqrt(sourceData[i] * candidateData[i]); 
    } 
     
    // The values of Bhattacharyya Coefficient ranges from 0 to 1, 
    double similarity =
0; 
    for(int i=0; i<mixedData.length; i++ ) { 
        similarity += mixedData[i]; 
    } 
     
    // The degree of similarity 

    return similarity; 



/**
* Bhattacharyya Coefficient
* http://www.cse.yorku.ca/~kosta/CompVis_Notes/bhattacharyya.pdf *
* @return
*/
public double modelMatch() {
HistogramFilter hfilter = new HistogramFilter();
float[] sourceData = hfilter.filter(sourceImage, null);
float[] candidateData = hfilter.filter(candidateImage, null);
double[] mixedData = new double[sourceData.length];
for(int i=0; i<sourceData.length; i++ ) {
mixedData[i] = Math.sqrt(sourceData[i] * candidateData[i]);
}

// The values of Bhattacharyya Coefficient ranges from 0 to 1,
double similarity = 0;
for(int i=0; i<mixedData.length; i++ ) {
similarity += mixedData[i];
}

// The degree of similarity
return similarity;
}
转载文章请务必注明出自本博客
内容来自用户分享和网络整理,不保证内容的准确性,如有侵权内容,可联系管理员处理 点击这里给我发消息
标签: