ISP模块之色彩增强算法--HSV空间Saturation通道调整

    色彩增强不同于彩色图像增强，图像增强的一般处理方式为直方图均衡化等，目的是为了增强图像局部以及整体对比度。而色彩增强的目的是为了使的原有的不饱和的色彩信息变得饱和、丰富起来。对应于Photoshop里面的“色相/饱和度”调节选项里面对饱和度的操作。色彩增强的过程，并不改变原有彩色图像的颜色以及亮度信息。

    在我的色彩增强算法模块里面，始终只针对色彩饱和度(Saturation)信息做研究，调整。这样的话，那就不得不介绍HSV颜色空间了，H代表Hue（色彩），S代表Saturation（饱和度），V代表Value，也可用B表示（Brightness，明度），HSV空间也可称作HSB空间。

    HSV空间在wikipedia上的介绍，https://en.wikipedia.org/wiki/HSL_and_HSV 

    下面根据自己的理解介绍一下HSV空间，以及其各通道在Matlab和OpenCV中的不同。

    HSV的圆柱模型

    

    HSV的圆锥模型

    

    从上图可以看出，在HSV空间中，Hue通道的取值从0-360°变化时，颜色从红->黄->绿->青->蓝逐步变化。Saturation从0->1变化时，色彩逐渐加深变成纯色（pure）。Value值从0->1变化时，图像整体亮度增加，V值为0时，图像为全黑，V值为1时，图像为全白。

    Matlab RGB色彩空间向HSV转换，采用函数rgb2hsv,转换后的hsv各通道的元素取值范围为[0,1];OpenCV中彩色图像向HSV空间中转换，cvtColor(src,srcHsv,CV_BGR2HSV),转换后H的取值范围为[0,180],S,V的取值范围为[0,255].

   下面介绍自己的算法处理思路，后面会给出完整的Matlab代码： 

   步骤一、给出一张原图src，用PS进行饱和度(Saturation)+40处理后另存为src_40;

   步骤二、将以上两张图像分别转换到hsv空间，提取出饱和度信息，分别为S，S_40;

   步骤三、统计饱和度增加40后，原色彩饱和度与饱和度增量之间的对应关系，即S -- (S_40-S);

   步骤四、对关系S -- (S_40-S)进行二次多项式曲线拟合，得到二次曲线f(x) = p1*x^2 + p2*x + p3；

   为什么是二次？1.对应关系呈现出抛物线形状；2.更高次曲线并没有明显改善拟合性能，且计算消耗会变高。

   步骤五、任意给定输出图像input，根据其色彩饱和度信息，即可进行色彩增强40处理，新的饱和度信息可以表示为S'(x) = S(x) + f(x),得到增强后的色彩信息后返回RGB图像输出；

   步骤六、分别对原图+20，+40，+60后进行饱和度信息统计，并得到相应拟合参数，设置为色彩增强的低、中、高三挡，在实际处理过程中，根据输入图像input自身色彩饱和度信息(均值)自适应选取相应参数进行色彩增强；

   步骤七、按需对某一单独颜色通道进行色彩增强处理，例如绿色范围为105°-135°，在对该范围进行增强的同时，还需对75°-105°,135°-165°进行一半强度的增强，这样才会保证色彩的连续性，不会出现色斑；

   步骤八、按需对色彩(Hue)进行转换；

   代码部分：第一部分用作估计拟合参数，在Curve fitting tool里面对X，Y进行拟合，得到曲线参数。

% Color Enhancement
clc,clear,close all
src1 = imread('src.bmp');
src2 = imread('src_40.bmp');

src1_hsv = rgb2hsv(src1);
src2_hsv = rgb2hsv(src2);

h1 = src1_hsv(:,:,1);
s1 = src1_hsv(:,:,2);
v1 = src1_hsv(:,:,3);

h2 = src2_hsv(:,:,1);
s2 = src2_hsv(:,:,2);
v2 = src2_hsv(:,:,3);
% 
meanS1 = mean(s1(:));
varS1  = std2(s1); 
% 
meanS2 = mean(s2(:));
varS2  = std2(s2); 
% 
deltaS = s2 - s1;
deltaV = v2 - v1;

%% test1 : 观测“原饱和度-饱和度调整增量”的关系 saturation and delta saturation
figure;
oriS = zeros(101,2);
s3 = s1;
j = 1;
for i = 0: 0.01 : 1
    oriS(j,1) = i + 0.01;
    oriS(j,2) =  mean(deltaS(find(s1 > i & s1< i + 0.01)));
    j = j + 1;
end
X  = oriS(:,1);
Y  = oriS(:,2);
XX = oriS(:,1) * 255;
YY = oriS(:,2) * 255;
plot(XX,YY)

   第二部分，对输入图像进行高、中、低三级自适应增强处理

%% Color Enhancement Module -- Authored by HuangDao,08/17/2015
% functions: input a image of type BMP or PNG, the program will decide to
% do the Color Enhancement choice for you.There are four types of Enhanced
% intensity - 20,40,60,80.The larger number stands for stronger
% enhancement.
% And we can also choose the simple color channel(eg.R,G,B) to do the
% enhancement.There are also four different types of enhanced intensity.
%
% parameters table
%  ------------------------------------------------------------------------
% | Enhanced  |     MATLAB params             |      OpenCV params         |
% | intensity |p1        p2        p3         | p1        p2        p3     |
% | 20        |-0.1661   0.2639    -0.003626  |-0.0006512 0.2639    -0.9246|
% | 40        |-0.4025   0.6238    -0.0005937 |0.001578   0.6238    -0.1514|
% | 60        |1.332     1.473     -0.01155   |-0.005222  1.473     -2.946 |
% | 80        |-4.813    3.459     -0.004568  |-0.01887   3.459     -1.165 |
%  ------------------------------------------------------------------------

clc; clear ;close all
% 载入文件夹
pathName = '.\';
fileType = '*.bmp';
files    = dir([pathName fileType]);
len      = length(files);

for pic = 5%1:1:len
    srcName = files(pic).name;
    srcImg  = imread(srcName);
    srcHSV  = rgb2hsv(srcImg);
    srcH    = srcHSV(:,:,1);
    srcS    = srcHSV(:,:,2);
    srcV    = srcHSV(:,:,3);
    meanS   = mean(srcS(:));
    varS    = std2(srcS);
    %图像整体进行色彩增强处理
    if (meanS >= 0.5)
        p1 = 0;p2 = 0;p3 = 0;
    else if (meanS >= 0.35 && meanS < 0.5)
            p1 = -0.1661;p2 = 0.2639;p3 = -0.003626;
        else if (meanS >=0.2 && meanS <0.35)
                p1 = -0.4025;p2 = 0.6238;p3 = -0.0005937;
            else
                p1 = 1.332;p2 = 1.473;p3 = -0.01155;
            end
        end
    end
    dstS = srcS + p1*srcS.*srcS + p2*srcS + p3 ;
    dstHSV = srcHSV;
    dstHSV(:,:,2) = dstS;
    dstImg = hsv2rgb(dstHSV);
    figure;imshow(srcImg);
    figure;imshow(dstImg);
    %指定R,G,B通道进行色彩增强处理,红色范围（[225-255]）,绿色范围（75-[105-135]-165）,蓝色范围（[-15-15]）
    p11 = -0.4025;p21 = 0.6238;p31 = -0.0005937;%周边杂色调整系数，40
    p12 = 1.332;  p22 = 1.473; p32 = -0.01155;  %纯色区域调整系数，60
    compHue = srcH;
    GcompS  = dstS;
    RcompS  = dstS;
    BcompS  = dstS;
    channel = 'B';
    switch channel
        case 'G'
            I1 = find(compHue > 0.2083 & compHue <0.2917);
            GcompS(I1) = dstS(I1) + dstS(I1).*dstS(I1)*p11 + dstS(I1)*p21 + p31;
            I2 = find(compHue >= 0.2917 & compHue <= 0.3750);
            GcompS(I2) = dstS(I2) + dstS(I2).*dstS(I2)*p12 + dstS(I2)*p22 + p32;
            I3 = find(compHue > 0.3750 & compHue <0.4583);
            GcompS(I3) = dstS(I3) + dstS(I3).*dstS(I3)*p11 + dstS(I3)*p21 + p31;
            compHSV = dstHSV;
            compHSV(:,:,2) = GcompS;
            dstImgG = hsv2rgb(compHSV);
            figure;imshow(dstImgG);
        case 'R'
            I1 = find(compHue > 0.875 & compHue <0.9583);
            RcompS(I1) = dstS(I1) + dstS(I1).*dstS(I1)*p11 + dstS(I1)*p21 + p31;
            I2 = find(compHue >= 0.9583 | compHue <= 0.0417);
            RcompS(I2) = dstS(I2) + dstS(I2).*dstS(I2)*p12 + dstS(I2)*p22 + p32;
            I3 = find(compHue > 0.0417 & compHue <0.125);
            RcompS(I3) = dstS(I3) + dstS(I3).*dstS(I3)*p11 + dstS(I3)*p21 + p31;
            compHSV = dstHSV;
            compHSV(:,:,2) = RcompS;
            dstImgR = hsv2rgb(compHSV);
            figure;imshow(dstImgR);
        case 'B'
            I1 = find(compHue > 0.5417 & compHue <0.625);
            BcompS(I1) = dstS(I1) + dstS(I1).*dstS(I1)*p11 + dstS(I1)*p21 + p31;
            I2 = find(compHue >= 0.625 & compHue <= 0.7083);
            BcompS(I2) = dstS(I2) + dstS(I2).*dstS(I2)*p12 + dstS(I2)*p22 + p32;
            I3 = find(compHue > 0.7083 & compHue <0.7917);
            BcompS(I3) = dstS(I3) + dstS(I3).*dstS(I3)*p11 + dstS(I3)*p21 + p31;
            compHSV = dstHSV;
            compHSV(:,:,2) = BcompS;
            dstImgB = hsv2rgb(compHSV);
            figure;imshow(dstImgB);
    end
    %进行R,G,B通道之间的互换
    convH = zeros(size(srcH,1),size(srcH,2)); %convert
    deltaHue = 240;
    switch deltaHue
        case 120
            disp('R -> G')
            convH = srcH + 1/3;
            convH(find(convH >= 1)) = convH(find(convH >= 1)) - 1;
        case 240
            disp('R -> B')
            convH = srcH + 2/3;
            convH(find(convH >= 1)) = convH(find(convH >= 1)) - 1;
    end
    convHSV = dstHSV;
    convHSV(:,:,1) = convH;
    convImg = hsv2rgb(convHSV);
    figure;imshow(convImg)
    pause();
end

   添加OpenCV代码段：

	Mat srcHSV,sat,satAdj,dstMerge,dst;     //sat - saturation饱和度分量
	Mat imageAwb = imread("m_ImageAwb.bmp");
	vector channels,channels1;
	double p1,p2,p3;

	cvtColor(imageAwb,srcHSV,CV_BGR2HSV);
	split(srcHSV,channels);
	split(srcHSV,channels1);
	sat = channels.at(1);
	Scalar m = mean(sat);

	if (m(0) <= 51.5)	                   
	{p1 = -0.002714 , p2 = 0.9498, p3 = -0.5073;  AfxMessageBox("High Color Enhancement!"); }//高
	else if (m(0) > 38.5 && m(0) <= 89.5)  
	{p1 = -0.001578  , p2 = 0.6238, p3 = -0.1514;AfxMessageBox("Middle Color Enhancement!"); }//中
	else if (m(0) > 89.5 && m(0) <=127.5)  
	{p1 = -0.0006512, p2 = 0.2639, p3 = -0.9246;AfxMessageBox("Low Color Enhancement!");}//低
	else                                   
	{p1 = 0,p2 = 0,p3 =0;AfxMessageBox("No Color Enhancement!");}

	satAdj = sat;
	for (int i = 0 ; i < sat.rows;i ++)
	{
		for (int j = 0;j < sat.cols;j ++)
		{
			uchar val = sat.at(i,j);
			satAdj.at(i,j) = (val + p1 * val * val + p2 * val + p3) ;
		}
	}

	channels1.at(1) = satAdj;
	merge(channels1,dstMerge);
	cvtColor(dstMerge,dst,CV_HSV2BGR);
	imwrite("m_ImageCE.bmp",dst);

   最后给出算法效果图：

Group1.原图->增强后

Group2.原图->R通道增强->颜色通道改变R2B

Group3.原图->增强后->颜色通道改变R2B

完！下篇讲Local Tone Mapping。