介紹基于皮膚檢測之後的，尋找最大連通區域，完成臉譜檢測的算法。大緻的算法步驟如下：
原圖如下：
每步處理以後的效果：
程式運作，加載選擇圖像以後的截屏如下：
截屏中顯示圖檔，是适當放縮以後，代碼如下：
image scaledimage = rawimg.getscaledinstance(200, 200, image.scale_fast); // java image api， rawimage is source image
g2.drawimage(scaledimage, 0, 0, 200, 200, null);
第一步：圖像預處理，預處理的目的是為了減少圖像中幹擾像素，使得皮膚檢測步驟可以得
到更好的效果，最常見的手段是調節對比度與亮度，也可以高斯模糊。
這裡調節對比度的算法很簡單，源代碼如下：
package com.gloomyfish.face.detection;
import java.awt.image.bufferedimage;
public class contrastfilter extends abstractbufferedimageop {
private double ncontrast = 30;
public contrastfilter() {
system.out.println("contrast filter");
}
@override
public bufferedimage filter(bufferedimage src, bufferedimage dest) {
int width = src.getwidth();
int height = src.getheight();
double contrast = (100.0 + ncontrast) / 100.0;
contrast *= contrast;
if ( dest == null )
dest = createcompatibledestimage( src, null );
int[] inpixels = new int[width*height];
int[] outpixels = new int[width*height];
getrgb( src, 0, 0, width, height, inpixels );
int index = 0;
int ta = 0, tr = 0, tg = 0, tb = 0;
for(int row=0; row<height; row++) {
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;
// adjust contrast - red, green, blue
tr = adjustcontrast(tr, contrast);
tg = adjustcontrast(tg, contrast);
tb = adjustcontrast(tb, contrast);
// output rgb pixel
outpixels[index] = (ta << 24) | (tr << 16) | (tg << 8) | tb;
}
}
setrgb( dest, 0, 0, width, height, outpixels );
return dest;
public int adjustcontrast(int color, double contrast) {
double result = 0;
result = color / 255.0;
result -= 0.5;
result *= contrast;
result += 0.5;
result *=255.0;
return clamp((int)result);
public static int clamp(int c) {
if (c < 0)
return 0;
if (c > 255)
return 255;
return c;
}
注意：第一步不是必須的，如果圖像品質已經很好，可以直接跳過。
第二步：皮膚檢測，采用的是基于rgb色彩空間的統計結果來判斷一個像素是否為skin像
素，如果是皮膚像素，則設定像素為黑色，否則為白色。給出基于rgb色彩空間的五種皮
膚檢測統計方法，最喜歡的一種源代碼如下：
/**
* this skin detection is absolutely good skin classification,
* i love this one very much
*
* this one should be always primary skin detection
* from all five filters
* @author zhigang
*
*/
public class skinfilter4 extends abstractbufferedimageop {
int ta = 0, tr = 0, tg = 0, tb = 0;
// detect skin method...
double sum = tr + tg + tb;
if (((double)tb/(double)tg<1.249) &&
((double)sum/(double)(3*tr)>0.696) &&
(0.3333-(double)tb/(double)sum>0.014) &&
((double)tg/(double)(3*sum)<0.108))
{
tr = tg = tb = 0;
} else {
tr = tg = tb = 255;
}
setrgb(dest, 0, 0, width, height, outpixels);
第三步：尋找最大連通區域
使用連通元件标記算法，尋找最大連通區域，關于什麼是連通元件标記算法，可以參見這裡
以這裡我完成了一個更具效率的版本，主要思想是對像素資料進行八鄰域尋找連通，然後合
并标記。源代碼如下：
import java.util.arrays;
import java.util.hashmap;
* fast connected component label algorithm
* @date 2012-05-23
public class fastconnectedcomponentlabelalg {
private int bgcolor;
private int[] labels;
private int[] outdata;
private int dw;
private int dh;
public fastconnectedcomponentlabelalg() {
bgcolor = 255; // black color
public int[] dolabel(int[] inpixels, int width, int height) {
dw = width;
dh = height;
int nextlabel = 1;
int result = 0;
labels = new int[dw * dh/2];
outdata = new int[dw * dh];
for(int i=0; i<labels.length; i++) {
labels[i] = i;
// we need to define these two variable arrays.
int[] fourneighborhoodpixels = new int[8];
int[] fourneighborhoodlabels = new int[8];
int[] knownlabels = new int[4];
int srcrgb = 0, index = 0;
boolean existedlabel = false;
for(int row = 0; row < height; row ++) {
for(int col = 0; col < width; col++) {
srcrgb = inpixels[index] & 0x000000ff;
if(srcrgb == bgcolor) {
result = 0; // which means no labeled for this pixel.
// we just find the eight neighborhood pixels.
fourneighborhoodpixels[0] = getpixel(inpixels, row-1, col); // upper cell
fourneighborhoodpixels[1] = getpixel(inpixels, row, col-1); // left cell
fourneighborhoodpixels[2] = getpixel(inpixels, row+1, col); // bottom cell
fourneighborhoodpixels[3] = getpixel(inpixels, row, col+1); // right cell
// four corners pixels
fourneighborhoodpixels[4] = getpixel(inpixels, row-1, col-1); // upper left corner
fourneighborhoodpixels[5] = getpixel(inpixels, row-1, col+1); // upper right corner
fourneighborhoodpixels[6] = getpixel(inpixels, row+1, col-1); // left bottom corner
fourneighborhoodpixels[7] = getpixel(inpixels, row+1, col+1); // right bottom corner
// get current possible existed labels
fourneighborhoodlabels[0] = getlabel(outdata, row-1, col); // upper cell
fourneighborhoodlabels[1] = getlabel(outdata, row, col-1); // left cell
fourneighborhoodlabels[2] = getlabel(outdata, row+1, col); // bottom cell
fourneighborhoodlabels[3] = getlabel(outdata, row, col+1); // right cell
// four corners labels value
fourneighborhoodlabels[4] = getlabel(outdata, row-1, col-1); // upper left corner
fourneighborhoodlabels[5] = getlabel(outdata, row-1, col+1); // upper right corner
fourneighborhoodlabels[6] = getlabel(outdata, row+1, col-1); // left bottom corner
fourneighborhoodlabels[7] = getlabel(outdata, row+1, col+1); // right bottom corner
knownlabels[0] = fourneighborhoodlabels[0];
knownlabels[1] = fourneighborhoodlabels[1];
knownlabels[2] = fourneighborhoodlabels[4];
knownlabels[3] = fourneighborhoodlabels[5];
existedlabel = false;
for(int k=0; k<fourneighborhoodlabels.length; k++) {
if(fourneighborhoodlabels[k] != 0) {
existedlabel = true;
break;
}
}
if(!existedlabel) {
result = nextlabel;
nextlabel++;
} else {
int found = -1, count = 0;
for(int i=0; i<fourneighborhoodpixels.length; i++) {
if(fourneighborhoodpixels[i] != bgcolor) {
found = i;
count++;
}
if(count == 1) {
result = (fourneighborhoodlabels[found] == 0) ? nextlabel : fourneighborhoodlabels[found];
} else {
for(int j=0; j<knownlabels.length; j++) {
if(knownlabels[j] != 0 && knownlabels[j] != result &&
knownlabels[j] < result) {
result = knownlabels[j];
}
boolean needmerge = false;
for(int mm = 0; mm < knownlabels.length; mm++ ) {
if(knownlabels[0] != knownlabels[mm] && knownlabels[mm] != 0) {
needmerge = true;
// merge the labels now....
if(needmerge) {
int minlabel = knownlabels[0];
for(int m=0; m<knownlabels.length; m++) {
if(minlabel > knownlabels[m] && knownlabels[m] != 0) {
minlabel = knownlabels[m];
}
// find the final label number...
result = (minlabel == 0) ? result : minlabel;
// re-assign the label number now...
if(knownlabels[0] != 0) {
setdata(outdata, row-1, col, result);
if(knownlabels[1] != 0) {
setdata(outdata, row, col-1, result);
if(knownlabels[2] != 0) {
setdata(outdata, row-1, col-1, result);
if(knownlabels[3] != 0) {
setdata(outdata, row-1, col+1, result);
outdata[index] = result; // assign to label
// post merge each labels now
mergelabels(index);
// labels statistic
hashmap<integer, integer> labelmap = new hashmap<integer, integer>();
for(int d=0; d<outdata.length; d++) {
if(outdata[d] != 0) {
if(labelmap.containskey(outdata[d])) {
integer count = labelmap.get(outdata[d]);
count+=1;
labelmap.put(outdata[d], count);
labelmap.put(outdata[d], 1);
// try to find the max connected component
integer[] keys = labelmap.keyset().toarray(new integer[0]);
arrays.sort(keys);
int maxkey = 1;
int max = 0;
for(integer key : keys) {
if(max < labelmap.get(key)){
max = labelmap.get(key);
maxkey = key;
system.out.println( "number of " + key + " = " + labelmap.get(key));
system.out.println("maxkey = " + maxkey);
system.out.println("max connected component number = " + max);
return outdata;
private void mergelabels(int index) {
int row = index / dw;
int col = index % dw;
// get current possible existed labels
int min = getlabel(outdata, row, col);
if(min == 0) return;
if(min > getlabel(outdata, row-1, col) && getlabel(outdata, row-1, col) != 0) {
min = getlabel(outdata, row-1, col);
if(min > getlabel(outdata, row, col-1) && getlabel(outdata, row, col-1) != 0) {
min = getlabel(outdata, row, col-1);
if(min > getlabel(outdata, row+1, col) && getlabel(outdata, row+1, col) != 0) {
min = getlabel(outdata, row+1, col);
if(min > getlabel(outdata, row, col+1) && getlabel(outdata, row, col+1) != 0) {
min = getlabel(outdata, row, col+1);
if(min > getlabel(outdata, row-1, col-1) && getlabel(outdata, row-1, col-1) != 0) {
min = getlabel(outdata, row-1, col-1);
if(min > getlabel(outdata, row-1, col+1) && getlabel(outdata, row-1, col+1) != 0) {
min = getlabel(outdata, row-1, col+1);
if(min > getlabel(outdata, row+1, col-1) && getlabel(outdata, row+1, col-1) != 0) {
min = getlabel(outdata, row+1, col-1);
if(min > getlabel(outdata, row+1, col+1) && getlabel(outdata, row+1, col+1) != 0) {
min = getlabel(outdata, row+1, col+1);
if(getlabel(outdata, row, col) == min)
return;
outdata[index] = min;
// eight neighborhood pixels
if((row -1) >= 0) {
mergelabels((row-1)*dw + col);
if((col-1) >= 0) {
mergelabels(row*dw+col-1);
if((row+1) < dh) {
mergelabels((row + 1)*dw+col);
if((col+1) < dw) {
mergelabels((row)*dw+col+1);
if((row-1)>= 0 && (col-1) >=0) {
mergelabels((row-1)*dw+col-1);
if((row-1)>= 0 && (col+1) < dw) {
mergelabels((row-1)*dw+col+1);
if((row+1) < dh && (col-1) >=0) {
mergelabels((row+1)*dw+col-1);
if((row+1) < dh && (col+1) < dw) {
mergelabels((row+1)*dw+col+1);
private void setdata(int[] data, int row, int col, int value) {
if(row < 0 || row >= dh) {
if(col < 0 || col >= dw) {
int index = row * dw + col;
data[index] = value;
private int getlabel(int[] data, int row, int col) {
// handle the edge pixels
return (data[index] & 0x000000ff);
private int getpixel(int[] data, int row, int col) {
return bgcolor;
/**
* binary image data:
*
* 255, 0, 0, 255, 0, 255, 255, 0, 255, 255, 255,
* 255, 0, 0, 255, 0, 255, 255, 0, 0, 255, 0,
* 255, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0,
* 255, 255, 0, 255, 255, 255, 0, 255, 0, 0, 255
* 255, 255, 0, 0, 0, 0, 255, 0, 0, 0, 0
* height = 5, width = 11
* @param args
*/
public static int[] imagedata = new int[]{
255, 0, 0, 255, 0, 255, 255, 0, 255, 255, 255,
255, 0, 0, 255, 0, 255, 255, 0, 0, 255, 0,
255, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0,
255, 255, 0, 255, 255, 255, 0, 255, 0, 0, 255,
255, 255, 0, 0, 0, 0, 255, 0, 0, 0, 0
};
public static void main(string[] args) {
fastconnectedcomponentlabelalg ccl = new fastconnectedcomponentlabelalg();
int[] outdata = ccl.dolabel(imagedata, 11, 5);
for(int i=0; i<5; i++) {
system.out.println("--------------------");
for(int j = 0; j<11; j++) {
int index = i * 11 + j;
if(j != 0) {
system.out.print(",");
system.out.print(outdata[index]);
system.out.println();
找到最大連通區域以後，對最大連通區域資料進行掃描，找出最小點，即矩形區域左上角坐
标，找出最大點，即矩形區域右下角坐标。知道這四個點坐标以後，在原圖上打上紅色矩形
框，标記出臉譜位置。尋找四個點坐标的實作代碼如下：
private void getfacerectangel() {
int width = resultimage.getwidth();
int height = resultimage.getheight();
int[] inpixels = new int[width*height];
getrgb(resultimage, 0, 0, width, height, inpixels);
int index = 0;
int ta = 0, tr = 0, tg = 0, tb = 0;
for(int row=0; row<height; row++) {
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;
if(tr == tg && tg == tb && tb == 0) { // face skin
if(miny > row) {
miny = row;
if(minx > col) {
minx = col;
if(maxy < row) {
maxy = row;
if(maxx < col) {
maxx = col;
}
}
缺點：
此算法不支援多臉譜檢測，不支援裸體中的臉譜檢測，但是根據人臉的
生物學特征可以進一步細化分析，支援裸體人臉檢測