基于像素的皮肤检测技术
介绍一种基于颜色空间的皮肤检测技术，可以检测亚洲人种与白人的皮肤，皮肤检测
人脸识别的基础，也是很多人像识别技术的基础操作，在实际应用中还是非常有用的。
基于像素的皮肤检测主要是寻找正确的颜色空间几何，图像处理中，常见的颜色空间
有如下几种
1. RGB色彩空间 – R代表单色红，G代表单色绿，B代表单色蓝
2. HSV色彩空间 – H 代表色彩， S代表饱和度，V代表强度值
3. YCbCr色彩空间 – 是数字电视的色彩空间
RGB转换为HSV的Java代码如下：
public static float[] rgbToHSV(int tr, int tg, int tb) {
float min, max, delta;
float hue, satur, value;
min = Math.min(tr, Math.min(tg, tb));
max = Math.max(tr, Math.max(tg, tb));
value = max;
delta = max - min;
if(max != 0) {
satur = delta/max;
} else {
satur = 0;
hue = -1;
}
if(tr == max) {
hue = (tg - tb)/delta;
}
else if(tg == max) {
hue = 2 + (tb-tr)/delta;
} else {
hue = 4 + (tr-tg)/delta;
}
hue = hue * 60.0f;
if(hue <0) {
hue = hue + 360;
}
return new float[]{hue, satur, value};
}
RGB转换为YCbCr的Java代码如下：
public static int[] rgbToYcrCb(int tr, int tg, int tb) {
double sum = tr + tg + tb;
double r = ((double)tr)/sum;
double g = ((double)tg)/sum;
double b = ((double)tb)/sum;
double y = 65.481 * r + 128.553 * g + 24.966 * b + 16.0d;
double Cr = -37.7745 * r - 74.1592 * g + 111.9337 * b + 128.0d;
double Cb = 111.9581 * r -93.7509 * g -18.2072 * b + 128.0d;
return new int[]{(int)y, (int)Cr, (int)Cb};
}
一个简单的基于RGB颜色空间的皮肤算法如下：
(R, G, B) is classified as skin if
R > 95 and G > 40 and B > 20and max{R, G, B} – min{R, G, B} > 15 and |R-G| > 15
and R > G and R > B
实现代码如下：
public boolean isSkin(int tr, int tg, int tb) {
int max = Math.max(tr, Math.max(tg, tb));
int min = Math.min(tr, Math.min(tg, tb));
int rg = Math.abs(tr - tg);
if(tr > 95 && tg > 40 && tb > 20 && rg > 15 &&
(max - min) > 15 && tr > tg && tr > tb) {
return true;
} else {
return false;
}
}
一个简单的基于HSV颜色空间的皮肤算法如下：
(H, S, V) will be classified as skin if
H > 0 and H <50 and S > 0.23 andS <0.68
实现代码如下：
public boolean isSkin(int tr, int tg, int tb) {
float[] HSV = ColorUtil.rgbToHSV(tr, tg, tb);
if((HSV[0] > 0.0f && HSV[0] <50.0f ) && (HSV[1] > 0.23f && HSV[1] <0.68f)){
return true;
} else {
return false;
}
}
一个简单的基于YCbCr颜色空间的皮肤算法如下：
(Y, Cb, Cr) will be classified as skin if:
Y > 80 and 85<Cb <135 and 135 <Cr <180, and (Y,Cb,Cr)= [0,255]
对于的Java代码如下:
public boolean isSkin(int tr, int tg, int tb) {
int y = (int)(tr * 0.299 + tg * 0.587 + tb * 0.114);
int Cr = tr - y;
int Cb = tb - y;
if(y> 80 && y <255 && Cr > 133 && Cr <173 && 77 基于上述三个算法实现的皮肤检测的效果如下：
皮肤检测滤镜的源代码如下：
package com.process.blur.study;
import java.awt.Color;
import java.awt.image.BufferedImage;
import com.gloomyfish.skin.dection.DefaultSkinDetection;
import com.gloomyfish.skin.dection.FastSkinDetection;
import com.gloomyfish.skin.dection.GaussianSkinDetection;
import com.gloomyfish.skin.dection.HSVSkinDetection;
import com.gloomyfish.skin.dection.ISkinDetection;
public class SkinFilter extends AbstractBufferedImageOp {
private ISkinDetection skinDetector;
public SkinFilter(int type) {
if(type == 2) {
skinDetector = new FastSkinDetection();
} else if(type == 4) {
skinDetector = new HSVSkinDetection();
} else if(type == 8) {
skinDetector = new GaussianSkinDetection();
} else {
skinDetector = new DefaultSkinDetection();
}
}
@Override
public BufferedImage filter(BufferedImage src, BufferedImage dst) {
int width = src.getWidth();
int height = src.getHeight();
if ( dst == null )
dst = createCompatibleDestImage( src, null );
int[] inPixels = new int[width*height];
int[] outPixels = new int[width*height];
getRGB( src, 0, 0, width, height, inPixels );
if(skinDetector instanceof GaussianSkinDetection) {
((GaussianSkinDetection)skinDetector).setDispSample(getDispersion(src));
}
int index = 0;
for(int row=0; row> 24) & 0xff;
tr = (inPixels[index] >> 16) & 0xff;
tg = (inPixels[index] >> 8) & 0xff;
tb = inPixels[index] & 0xff;
if(skinDetector.isSkin(tr, tg, tb)) {
outPixels[index] = (ta <<24) | (tr <<16) | (tg <<8) | tb;
} else {
tr = tg = tb = 0;
outPixels[index] = (ta <<24) | (tr <<16) | (tg <<8) | tb;
}
}
}
setRGB( dst, 0, 0, width, height, outPixels );
return dst;
}
public Color getDispersion(BufferedImage image) {
// calculate means of pixel
int index = 0;
int height = image.getHeight();
int width = image.getWidth();
int[] inPixels = new int[width*height];
getRGB(image, 0, 0, width, height, inPixels );
double redSum = 0, greenSum = 0, blueSum = 0;
Color meanColor = getMean(image);
double redmeans = meanColor.getRed();
double greenmeans = meanColor.getGreen();
double bluemeans = meanColor.getBlue();
double total = height * width;
for(int row=0; row> 24) & 0xff;
tr = (inPixels[index] >> 16) & 0xff;
tg = (inPixels[index] >> 8) & 0xff;
tb = inPixels[index] & 0xff;
double rd = (tr - redmeans);
double gd = (tg - greenmeans);
double bd = (tb - bluemeans);
redSum += rd * rd;
greenSum += gd * gd;
blueSum += bd * bd;
}
}
int reddiff = (int)Math.sqrt((redSum / total));
int greendiff = (int)Math.sqrt((greenSum / total));
int bluediff = (int)Math.sqrt(blueSum / total);
System.out.println(" red dispersion value = " + reddiff);
System.out.println(" green dispersion value = " + greendiff);
System.out.println(" blue dispersion value = " + bluediff);
return new Color(reddiff, greendiff, bluediff);
}
public Color getMean(BufferedImage image) {
// calculate means of pixel
int index = 0;
int height = image.getHeight();
int width = image.getWidth();
int[] inPixels = new int[width*height];
getRGB(image, 0, 0, width, height, inPixels );
double redSum = 0, greenSum = 0, blueSum = 0;
double total = height * width;
for(int row=0; row> 24) & 0xff;
tr = (inPixels[index] >> 16) & 0xff;
tg = (inPixels[index] >> 8) & 0xff;
tb = inPixels[index] & 0xff;
redSum += tr;
greenSum += tg;
blueSum +=tb;
}
}
int redmeans = (int)(redSum / total);
int greenmeans = (int)(greenSum / total);
int bluemeans = (int)(blueSum / total);
System.out.println(" red average value = " + redmeans);
System.out.println(" green average value = " + greenmeans);
System.out.println(" blue average value = " + bluemeans);
return new Color(redmeans, greenmeans, bluemeans);
}
}
讨论:
皮肤检测中的后续处理非常重要，可以除去噪声，平滑图像，是皮肤检测的结果
更加的准确，输出的更容易接受。
参考引用：
《A New Fast Skin Color Detection Technique》 - Tarek M. Mahmoud
《Improved Automatic Skin Detection in Color Images》 - Filipe Tomaz
and Tiago Candeias and Hamid Shahbazkia
《Skin Detection using HSV color space》- unknown author
京公网安备 11010802041100号