圖像分割之(四)OpenCV的GrabCut函數使用和源碼解讀
上一文對GrabCut做了一個瞭解。OpenCV中的GrabCut算法是依據《"GrabCut" - Interactive Foreground Extraction using Iterated Graph Cuts》這篇文章來實現的。現在我對源碼做了些註釋,以便我們更深入的瞭解該算法。一直覺得論文和代碼是有比較大的差別的,個人覺得脫離代碼看論文,最多能看懂70%,剩下20%或者更多就需要通過閱讀代碼來獲得了,那還有10%就和每個人的基礎和知識儲備相掛鉤了。
接觸時間有限,若有錯誤,還望各位前輩指正,謝謝。原論文的一些淺解見上一博文:
http://blog.csdn.net/zouxy09/article/details/8534954
一、GrabCut函數使用
在OpenCV的源碼目錄的samples的文件夾下,有grabCut的使用例程,請參考:
opencv\samples\cpp\grabcut.cpp。
而grabCut函數的API說明如下:
void cv::grabCut( InputArray _img, InputOutputArray _mask, Rect rect,
InputOutputArray _bgdModel, InputOutputArray _fgdModel,
int iterCount, int mode )
/*
****參數說明:
img——待分割的源圖像,必須是8位3通道(CV_8UC3)圖像,在處理的過程中不會被修改;
mask——掩碼圖像,如果使用掩碼進行初始化,那麼mask保存初始化掩碼信息;在執行分割的時候,也可以將用戶交互所設定的前景與背景保存到mask中,然後再傳入grabCut函數;在處理結束之後,mask中會保存結果。mask只能取以下四種值:
GCD_BGD(=0),背景;
GCD_FGD(=1),前景;
GCD_PR_BGD(=2),可能的背景;
GCD_PR_FGD(=3),可能的前景。
如果沒有手工標記GCD_BGD或者GCD_FGD,那麼結果只會有GCD_PR_BGD或GCD_PR_FGD;
rect——用於限定需要進行分割的圖像範圍,只有該矩形窗口內的圖像部分才被處理;
bgdModel——背景模型,如果爲null,函數內部會自動創建一個bgdModel;bgdModel必須是單通道浮點型(CV_32FC1)圖像,且行數只能爲1,列數只能爲13x5;
fgdModel——前景模型,如果爲null,函數內部會自動創建一個fgdModel;fgdModel必須是單通道浮點型(CV_32FC1)圖像,且行數只能爲1,列數只能爲13x5;
iterCount——迭代次數,必須大於0;
mode——用於指示grabCut函數進行什麼操作,可選的值有:
GC_INIT_WITH_RECT(=0),用矩形窗初始化GrabCut;
GC_INIT_WITH_MASK(=1),用掩碼圖像初始化GrabCut;
GC_EVAL(=2),執行分割。
*/
二、GrabCut源碼解讀
其中源碼包含了gcgraph.hpp這個構建圖和max flow/min cut算法的實現文件,這個文件暫時沒有解讀,後面再更新了。
- /*M///////////////////////////////////////////////////////////////////////////////////////
- //
- // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
- //
- // By downloading, copying, installing or using the software you agree to this license.
- // If you do not agree to this license, do not download, install,
- // copy or use the software.
- //
- //
- // Intel License Agreement
- // For Open Source Computer Vision Library
- //
- // Copyright (C) 2000, Intel Corporation, all rights reserved.
- // Third party copyrights are property of their respective owners.
- //
- // Redistribution and use in source and binary forms, with or without modification,
- // are permitted provided that the following conditions are met:
- //
- // * Redistribution's of source code must retain the above copyright notice,
- // this list of conditions and the following disclaimer.
- //
- // * Redistribution's in binary form must reproduce the above copyright notice,
- // this list of conditions and the following disclaimer in the documentation
- // and/or other materials provided with the distribution.
- //
- // * The name of Intel Corporation may not be used to endorse or promote products
- // derived from this software without specific prior written permission.
- //
- // This software is provided by the copyright holders and contributors "as is" and
- // any express or implied warranties, including, but not limited to, the implied
- // warranties of merchantability and fitness for a particular purpose are disclaimed.
- // In no event shall the Intel Corporation or contributors be liable for any direct,
- // indirect, incidental, special, exemplary, or consequential damages
- // (including, but not limited to, procurement of substitute goods or services;
- // loss of use, data, or profits; or business interruption) however caused
- // and on any theory of liability, whether in contract, strict liability,
- // or tort (including negligence or otherwise) arising in any way out of
- // the use of this software, even if advised of the possibility of such damage.
- //
- //M*/
- #include "precomp.hpp"
- #include "gcgraph.hpp"
- #include <limits>
- using namespace cv;
- /*
- This is implementation of image segmentation algorithm GrabCut described in
- "GrabCut — Interactive Foreground Extraction using Iterated Graph Cuts".
- Carsten Rother, Vladimir Kolmogorov, Andrew Blake.
- */
- /*
- GMM - Gaussian Mixture Model
- */
- class GMM
- {
- public:
- static const int componentsCount = 5;
- GMM( Mat& _model );
- double operator()( const Vec3d color ) const;
- double operator()( int ci, const Vec3d color ) const;
- int whichComponent( const Vec3d color ) const;
- void initLearning();
- void addSample( int ci, const Vec3d color );
- void endLearning();
- private:
- void calcInverseCovAndDeterm( int ci );
- Mat model;
- double* coefs;
- double* mean;
- double* cov;
- double inverseCovs[componentsCount][3][3]; //協方差的逆矩陣
- double covDeterms[componentsCount]; //協方差的行列式
- double sums[componentsCount][3];
- double prods[componentsCount][3][3];
- int sampleCounts[componentsCount];
- int totalSampleCount;
- };
- //背景和前景各有一個對應的GMM(混合高斯模型)
- GMM::GMM( Mat& _model )
- {
- //一個像素的(唯一對應)高斯模型的參數個數或者說一個高斯模型的參數個數
- //一個像素RGB三個通道值,故3個均值,3*3個協方差,共用一個權值
- const int modelSize = 3/*mean*/ + 9/*covariance*/ + 1/*component weight*/;
- if( _model.empty() )
- {
- //一個GMM共有componentsCount個高斯模型,一個高斯模型有modelSize個模型參數
- _model.create( 1, modelSize*componentsCount, CV_64FC1 );
- _model.setTo(Scalar(0));
- }
- else if( (_model.type() != CV_64FC1) || (_model.rows != 1) || (_model.cols != modelSize*componentsCount) )
- CV_Error( CV_StsBadArg, "_model must have CV_64FC1 type, rows == 1 and cols == 13*componentsCount" );
- model = _model;
- //注意這些模型參數的存儲方式:先排完componentsCount個coefs,再3*componentsCount個mean。
- //再3*3*componentsCount個cov。
- coefs = model.ptr<double>(0); //GMM的每個像素的高斯模型的權值變量起始存儲指針
- mean = coefs + componentsCount; //均值變量起始存儲指針
- cov = mean + 3*componentsCount; //協方差變量起始存儲指針
- for( int ci = 0; ci < componentsCount; ci++ )
- if( coefs[ci] > 0 )
- //計算GMM中第ci個高斯模型的協方差的逆Inverse和行列式Determinant
- //爲了後面計算每個像素屬於該高斯模型的概率(也就是數據能量項)
- calcInverseCovAndDeterm( ci );
- }
- //計算一個像素(由color=(B,G,R)三維double型向量來表示)屬於這個GMM混合高斯模型的概率。
- //也就是把這個像素像素屬於componentsCount個高斯模型的概率與對應的權值相乘再相加,
- //具體見論文的公式(10)。結果從res返回。
- //這個相當於計算Gibbs能量的第一個能量項(取負後)。
- double GMM::operator()( const Vec3d color ) const
- {
- double res = 0;
- for( int ci = 0; ci < componentsCount; ci++ )
- res += coefs[ci] * (*this)(ci, color );
- return res;
- }
- //計算一個像素(由color=(B,G,R)三維double型向量來表示)屬於第ci個高斯模型的概率。
- //具體過程,即高階的高斯密度模型計算式,具體見論文的公式(10)。結果從res返回
- double GMM::operator()( int ci, const Vec3d color ) const
- {
- double res = 0;
- if( coefs[ci] > 0 )
- {
- CV_Assert( covDeterms[ci] > std::numeric_limits<double>::epsilon() );
- Vec3d diff = color;
- double* m = mean + 3*ci;
- diff[0] -= m[0]; diff[1] -= m[1]; diff[2] -= m[2];
- double mult = diff[0]*(diff[0]*inverseCovs[ci][0][0] + diff[1]*inverseCovs[ci][1][0] + diff[2]*inverseCovs[ci][2][0])
- + diff[1]*(diff[0]*inverseCovs[ci][0][1] + diff[1]*inverseCovs[ci][1][1] + diff[2]*inverseCovs[ci][2][1])
- + diff[2]*(diff[0]*inverseCovs[ci][0][2] + diff[1]*inverseCovs[ci][1][2] + diff[2]*inverseCovs[ci][2][2]);
- res = 1.0f/sqrt(covDeterms[ci]) * exp(-0.5f*mult);
- }
- return res;
- }
- //返回這個像素最有可能屬於GMM中的哪個高斯模型(概率最大的那個)
- int GMM::whichComponent( const Vec3d color ) const
- {
- int k = 0;
- double max = 0;
- for( int ci = 0; ci < componentsCount; ci++ )
- {
- double p = (*this)( ci, color );
- if( p > max )
- {
- k = ci; //找到概率最大的那個,或者說計算結果最大的那個
- max = p;
- }
- }
- return k;
- }
- //GMM參數學習前的初始化,主要是對要求和的變量置零
- void GMM::initLearning()
- {
- for( int ci = 0; ci < componentsCount; ci++)
- {
- sums[ci][0] = sums[ci][1] = sums[ci][2] = 0;
- prods[ci][0][0] = prods[ci][0][1] = prods[ci][0][2] = 0;
- prods[ci][1][0] = prods[ci][1][1] = prods[ci][1][2] = 0;
- prods[ci][2][0] = prods[ci][2][1] = prods[ci][2][2] = 0;
- sampleCounts[ci] = 0;
- }
- totalSampleCount = 0;
- }
- //增加樣本,即爲前景或者背景GMM的第ci個高斯模型的像素集(這個像素集是來用估
- //計計算這個高斯模型的參數的)增加樣本像素。計算加入color這個像素後,像素集
- //中所有像素的RGB三個通道的和sums(用來計算均值),還有它的prods(用來計算協方差),
- //並且記錄這個像素集的像素個數和總的像素個數(用來計算這個高斯模型的權值)。
- void GMM::addSample( int ci, const Vec3d color )
- {
- sums[ci][0] += color[0]; sums[ci][1] += color[1]; sums[ci][2] += color[2];
- prods[ci][0][0] += color[0]*color[0]; prods[ci][0][1] += color[0]*color[1]; prods[ci][0][2] += color[0]*color[2];
- prods[ci][1][0] += color[1]*color[0]; prods[ci][1][1] += color[1]*color[1]; prods[ci][1][2] += color[1]*color[2];
- prods[ci][2][0] += color[2]*color[0]; prods[ci][2][1] += color[2]*color[1]; prods[ci][2][2] += color[2]*color[2];
- sampleCounts[ci]++;
- totalSampleCount++;
- }
- //從圖像數據中學習GMM的參數:每一個高斯分量的權值、均值和協方差矩陣;
- //這裏相當於論文中“Iterative minimisation”的step 2
- void GMM::endLearning()
- {
- const double variance = 0.01;
- for( int ci = 0; ci < componentsCount; ci++ )
- {
- int n = sampleCounts[ci]; //第ci個高斯模型的樣本像素個數
- if( n == 0 )
- coefs[ci] = 0;
- else
- {
- //計算第ci個高斯模型的權值係數
- coefs[ci] = (double)n/totalSampleCount;
- //計算第ci個高斯模型的均值
- double* m = mean + 3*ci;
- m[0] = sums[ci][0]/n; m[1] = sums[ci][1]/n; m[2] = sums[ci][2]/n;
- //計算第ci個高斯模型的協方差
- double* c = cov + 9*ci;
- c[0] = prods[ci][0][0]/n - m[0]*m[0]; c[1] = prods[ci][0][1]/n - m[0]*m[1]; c[2] = prods[ci][0][2]/n - m[0]*m[2];
- c[3] = prods[ci][1][0]/n - m[1]*m[0]; c[4] = prods[ci][1][1]/n - m[1]*m[1]; c[5] = prods[ci][1][2]/n - m[1]*m[2];
- c[6] = prods[ci][2][0]/n - m[2]*m[0]; c[7] = prods[ci][2][1]/n - m[2]*m[1]; c[8] = prods[ci][2][2]/n - m[2]*m[2];
- //計算第ci個高斯模型的協方差的行列式
- double dtrm = c[0]*(c[4]*c[8]-c[5]*c[7]) - c[1]*(c[3]*c[8]-c[5]*c[6]) + c[2]*(c[3]*c[7]-c[4]*c[6]);
- if( dtrm <= std::numeric_limits<double>::epsilon() )
- {
- //相當於如果行列式小於等於0,(對角線元素)增加白噪聲,避免其變
- //爲退化(降秩)協方差矩陣(不存在逆矩陣,但後面的計算需要計算逆矩陣)。
- // Adds the white noise to avoid singular covariance matrix.
- c[0] += variance;
- c[4] += variance;
- c[8] += variance;
- }
- //計算第ci個高斯模型的協方差的逆Inverse和行列式Determinant
- calcInverseCovAndDeterm(ci);
- }
- }
- }
- //計算協方差的逆Inverse和行列式Determinant
- void GMM::calcInverseCovAndDeterm( int ci )
- {
- if( coefs[ci] > 0 )
- {
- //取第ci個高斯模型的協方差的起始指針
- double *c = cov + 9*ci;
- double dtrm =
- covDeterms[ci] = c[0]*(c[4]*c[8]-c[5]*c[7]) - c[1]*(c[3]*c[8]-c[5]*c[6])
- + c[2]*(c[3]*c[7]-c[4]*c[6]);
- //在C++中,每一種內置的數據類型都擁有不同的屬性, 使用<limits>庫可以獲
- //得這些基本數據類型的數值屬性。因爲浮點算法的截斷,所以使得,當a=2,
- //b=3時 10*a/b == 20/b不成立。那怎麼辦呢?
- //這個小正數(epsilon)常量就來了,小正數通常爲可用給定數據類型的
- //大於1的最小值與1之差來表示。若dtrm結果不大於小正數,那麼它幾乎爲零。
- //所以下式保證dtrm>0,即行列式的計算正確(協方差對稱正定,故行列式大於0)。
- CV_Assert( dtrm > std::numeric_limits<double>::epsilon() );
- //三階方陣的求逆
- inverseCovs[ci][0][0] = (c[4]*c[8] - c[5]*c[7]) / dtrm;
- inverseCovs[ci][1][0] = -(c[3]*c[8] - c[5]*c[6]) / dtrm;
- inverseCovs[ci][2][0] = (c[3]*c[7] - c[4]*c[6]) / dtrm;
- inverseCovs[ci][0][1] = -(c[1]*c[8] - c[2]*c[7]) / dtrm;
- inverseCovs[ci][1][1] = (c[0]*c[8] - c[2]*c[6]) / dtrm;
- inverseCovs[ci][2][1] = -(c[0]*c[7] - c[1]*c[6]) / dtrm;
- inverseCovs[ci][0][2] = (c[1]*c[5] - c[2]*c[4]) / dtrm;
- inverseCovs[ci][1][2] = -(c[0]*c[5] - c[2]*c[3]) / dtrm;
- inverseCovs[ci][2][2] = (c[0]*c[4] - c[1]*c[3]) / dtrm;
- }
- }
- //計算beta,也就是Gibbs能量項中的第二項(平滑項)中的指數項的beta,用來調整
- //高或者低對比度時,兩個鄰域像素的差別的影響的,例如在低對比度時,兩個鄰域
- //像素的差別可能就會比較小,這時候需要乘以一個較大的beta來放大這個差別,
- //在高對比度時,則需要縮小本身就比較大的差別。
- //所以我們需要分析整幅圖像的對比度來確定參數beta,具體的見論文公式(5)。
- /*
- Calculate beta - parameter of GrabCut algorithm.
- beta = 1/(2*avg(sqr(||color[i] - color[j]||)))
- */
- static double calcBeta( const Mat& img )
- {
- double beta = 0;
- for( int y = 0; y < img.rows; y++ )
- {
- for( int x = 0; x < img.cols; x++ )
- {
- //計算四個方向鄰域兩像素的差別,也就是歐式距離或者說二階範數
- //(當所有像素都算完後,就相當於計算八鄰域的像素差了)
- Vec3d color = img.at<Vec3b>(y,x);
- if( x>0 ) // left >0的判斷是爲了避免在圖像邊界的時候還計算,導致越界
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y,x-1);
- beta += diff.dot(diff); //矩陣的點乘,也就是各個元素平方的和
- }
- if( y>0 && x>0 ) // upleft
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y-1,x-1);
- beta += diff.dot(diff);
- }
- if( y>0 ) // up
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y-1,x);
- beta += diff.dot(diff);
- }
- if( y>0 && x<img.cols-1) // upright
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y-1,x+1);
- beta += diff.dot(diff);
- }
- }
- }
- if( beta <= std::numeric_limits<double>::epsilon() )
- beta = 0;
- else
- beta = 1.f / (2 * beta/(4*img.cols*img.rows - 3*img.cols - 3*img.rows + 2) ); //論文公式(5)
- return beta;
- }
- //計算圖每個非端點頂點(也就是每個像素作爲圖的一個頂點,不包括源點s和匯點t)與鄰域頂點
- //的邊的權值。由於是無向圖,我們計算的是八鄰域,那麼對於一個頂點,我們計算四個方向就行,
- //在其他的頂點計算的時候,會把剩餘那四個方向的權值計算出來。這樣整個圖算完後,每個頂點
- //與八鄰域的頂點的邊的權值就都計算出來了。
- //這個相當於計算Gibbs能量的第二個能量項(平滑項),具體見論文中公式(4)
- /*
- Calculate weights of noterminal vertices of graph.
- beta and gamma - parameters of GrabCut algorithm.
- */
- static void calcNWeights( const Mat& img, Mat& leftW, Mat& upleftW, Mat& upW,
- Mat& uprightW, double beta, double gamma )
- {
- //gammaDivSqrt2相當於公式(4)中的gamma * dis(i,j)^(-1),那麼可以知道,
- //當i和j是垂直或者水平關係時,dis(i,j)=1,當是對角關係時,dis(i,j)=sqrt(2.0f)。
- //具體計算時,看下面就明白了
- const double gammaDivSqrt2 = gamma / std::sqrt(2.0f);
- //每個方向的邊的權值通過一個和圖大小相等的Mat來保存
- leftW.create( img.rows, img.cols, CV_64FC1 );
- upleftW.create( img.rows, img.cols, CV_64FC1 );
- upW.create( img.rows, img.cols, CV_64FC1 );
- uprightW.create( img.rows, img.cols, CV_64FC1 );
- for( int y = 0; y < img.rows; y++ )
- {
- for( int x = 0; x < img.cols; x++ )
- {
- Vec3d color = img.at<Vec3b>(y,x);
- if( x-1>=0 ) // left //避免圖的邊界
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y,x-1);
- leftW.at<double>(y,x) = gamma * exp(-beta*diff.dot(diff));
- }
- else
- leftW.at<double>(y,x) = 0;
- if( x-1>=0 && y-1>=0 ) // upleft
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y-1,x-1);
- upleftW.at<double>(y,x) = gammaDivSqrt2 * exp(-beta*diff.dot(diff));
- }
- else
- upleftW.at<double>(y,x) = 0;
- if( y-1>=0 ) // up
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y-1,x);
- upW.at<double>(y,x) = gamma * exp(-beta*diff.dot(diff));
- }
- else
- upW.at<double>(y,x) = 0;
- if( x+1<img.cols && y-1>=0 ) // upright
- {
- Vec3d diff = color - (Vec3d)img.at<Vec3b>(y-1,x+1);
- uprightW.at<double>(y,x) = gammaDivSqrt2 * exp(-beta*diff.dot(diff));
- }
- else
- uprightW.at<double>(y,x) = 0;
- }
- }
- }
- //檢查mask的正確性。mask爲通過用戶交互或者程序設定的,它是和圖像大小一樣的單通道灰度圖,
- //每個像素只能取GC_BGD or GC_FGD or GC_PR_BGD or GC_PR_FGD 四種枚舉值,分別表示該像素
- //(用戶或者程序指定)屬於背景、前景、可能爲背景或者可能爲前景像素。具體的參考:
- //ICCV2001“Interactive Graph Cuts for Optimal Boundary & Region Segmentation of Objects in N-D Images”
- //Yuri Y. Boykov Marie-Pierre Jolly
- /*
- Check size, type and element values of mask matrix.
- */
- static void checkMask( const Mat& img, const Mat& mask )
- {
- if( mask.empty() )
- CV_Error( CV_StsBadArg, "mask is empty" );
- if( mask.type() != CV_8UC1 )
- CV_Error( CV_StsBadArg, "mask must have CV_8UC1 type" );
- if( mask.cols != img.cols || mask.rows != img.rows )
- CV_Error( CV_StsBadArg, "mask must have as many rows and cols as img" );
- for( int y = 0; y < mask.rows; y++ )
- {
- for( int x = 0; x < mask.cols; x++ )
- {
- uchar val = mask.at<uchar>(y,x);
- if( val!=GC_BGD && val!=GC_FGD && val!=GC_PR_BGD && val!=GC_PR_FGD )
- CV_Error( CV_StsBadArg, "mask element value must be equel"
- "GC_BGD or GC_FGD or GC_PR_BGD or GC_PR_FGD" );
- }
- }
- }
- //通過用戶框選目標rect來創建mask,rect外的全部作爲背景,設置爲GC_BGD,
- //rect內的設置爲 GC_PR_FGD(可能爲前景)
- /*
- Initialize mask using rectangular.
- */
- static void initMaskWithRect( Mat& mask, Size imgSize, Rect rect )
- {
- mask.create( imgSize, CV_8UC1 );
- mask.setTo( GC_BGD );
- rect.x = max(0, rect.x);
- rect.y = max(0, rect.y);
- rect.width = min(rect.width, imgSize.width-rect.x);
- rect.height = min(rect.height, imgSize.height-rect.y);
- (mask(rect)).setTo( Scalar(GC_PR_FGD) );
- }
- //通過k-means算法來初始化背景GMM和前景GMM模型
- /*
- Initialize GMM background and foreground models using kmeans algorithm.
- */
- static void initGMMs( const Mat& img, const Mat& mask, GMM& bgdGMM, GMM& fgdGMM )
- {
- const int kMeansItCount = 10; //迭代次數
- const int kMeansType = KMEANS_PP_CENTERS; //Use kmeans++ center initialization by Arthur and Vassilvitskii
- Mat bgdLabels, fgdLabels; //記錄背景和前景的像素樣本集中每個像素對應GMM的哪個高斯模型,論文中的kn
- vector<Vec3f> bgdSamples, fgdSamples; //背景和前景的像素樣本集
- Point p;
- for( p.y = 0; p.y < img.rows; p.y++ )
- {
- for( p.x = 0; p.x < img.cols; p.x++ )
- {
- //mask中標記爲GC_BGD和GC_PR_BGD的像素都作爲背景的樣本像素
- if( mask.at<uchar>(p) == GC_BGD || mask.at<uchar>(p) == GC_PR_BGD )
- bgdSamples.push_back( (Vec3f)img.at<Vec3b>(p) );
- else // GC_FGD | GC_PR_FGD
- fgdSamples.push_back( (Vec3f)img.at<Vec3b>(p) );
- }
- }
- CV_Assert( !bgdSamples.empty() && !fgdSamples.empty() );
- //kmeans中參數_bgdSamples爲:每行一個樣本
- //kmeans的輸出爲bgdLabels,裏面保存的是輸入樣本集中每一個樣本對應的類標籤(樣本聚爲componentsCount類後)
- Mat _bgdSamples( (int)bgdSamples.size(), 3, CV_32FC1, &bgdSamples[0][0] );
- kmeans( _bgdSamples, GMM::componentsCount, bgdLabels,
- TermCriteria( CV_TERMCRIT_ITER, kMeansItCount, 0.0), 0, kMeansType );
- Mat _fgdSamples( (int)fgdSamples.size(), 3, CV_32FC1, &fgdSamples[0][0] );
- kmeans( _fgdSamples, GMM::componentsCount, fgdLabels,
- TermCriteria( CV_TERMCRIT_ITER, kMeansItCount, 0.0), 0, kMeansType );
- //經過上面的步驟後,每個像素所屬的高斯模型就確定的了,那麼就可以估計GMM中每個高斯模型的參數了。
- bgdGMM.initLearning();
- for( int i = 0; i < (int)bgdSamples.size(); i++ )
- bgdGMM.addSample( bgdLabels.at<int>(i,0), bgdSamples[i] );
- bgdGMM.endLearning();
- fgdGMM.initLearning();
- for( int i = 0; i < (int)fgdSamples.size(); i++ )
- fgdGMM.addSample( fgdLabels.at<int>(i,0), fgdSamples[i] );
- fgdGMM.endLearning();
- }
- //論文中:迭代最小化算法step 1:爲每個像素分配GMM中所屬的高斯模型,kn保存在Mat compIdxs中
- /*
- Assign GMMs components for each pixel.
- */
- static void assignGMMsComponents( const Mat& img, const Mat& mask, const GMM& bgdGMM,
- const GMM& fgdGMM, Mat& compIdxs )
- {
- Point p;
- for( p.y = 0; p.y < img.rows; p.y++ )
- {
- for( p.x = 0; p.x < img.cols; p.x++ )
- {
- Vec3d color = img.at<Vec3b>(p);
- //通過mask來判斷該像素屬於背景像素還是前景像素,再判斷它屬於前景或者背景GMM中的哪個高斯分量
- compIdxs.at<int>(p) = mask.at<uchar>(p) == GC_BGD || mask.at<uchar>(p) == GC_PR_BGD ?
- bgdGMM.whichComponent(color) : fgdGMM.whichComponent(color);
- }
- }
- }
- //論文中:迭代最小化算法step 2:從每個高斯模型的像素樣本集中學習每個高斯模型的參數
- /*
- Learn GMMs parameters.
- */
- static void learnGMMs( const Mat& img, const Mat& mask, const Mat& compIdxs, GMM& bgdGMM, GMM& fgdGMM )
- {
- bgdGMM.initLearning();
- fgdGMM.initLearning();
- Point p;
- for( int ci = 0; ci < GMM::componentsCount; ci++ )
- {
- for( p.y = 0; p.y < img.rows; p.y++ )
- {
- for( p.x = 0; p.x < img.cols; p.x++ )
- {
- if( compIdxs.at<int>(p) == ci )
- {
- if( mask.at<uchar>(p) == GC_BGD || mask.at<uchar>(p) == GC_PR_BGD )
- bgdGMM.addSample( ci, img.at<Vec3b>(p) );
- else
- fgdGMM.addSample( ci, img.at<Vec3b>(p) );
- }
- }
- }
- }
- bgdGMM.endLearning();
- fgdGMM.endLearning();
- }
- //通過計算得到的能量項構建圖,圖的頂點爲像素點,圖的邊由兩部分構成,
- //一類邊是:每個頂點與Sink匯點t(代表背景)和源點Source(代表前景)連接的邊,
- //這類邊的權值通過Gibbs能量項的第一項能量項來表示。
- //另一類邊是:每個頂點與其鄰域頂點連接的邊,這類邊的權值通過Gibbs能量項的第二項能量項來表示。
- /*
- Construct GCGraph
- */
- static void constructGCGraph( const Mat& img, const Mat& mask, const GMM& bgdGMM, const GMM& fgdGMM, double lambda,
- const Mat& leftW, const Mat& upleftW, const Mat& upW, const Mat& uprightW,
- GCGraph<double>& graph )
- {
- int vtxCount = img.cols*img.rows; //頂點數,每一個像素是一個頂點
- int edgeCount = 2*(4*vtxCount - 3*(img.cols + img.rows) + 2); //邊數,需要考慮圖邊界的邊的缺失
- //通過頂點數和邊數創建圖。這些類型聲明和函數定義請參考gcgraph.hpp
- graph.create(vtxCount, edgeCount);
- Point p;
- for( p.y = 0; p.y < img.rows; p.y++ )
- {
- for( p.x = 0; p.x < img.cols; p.x++)
- {
- // add node
- int vtxIdx = graph.addVtx(); //返回這個頂點在圖中的索引
- Vec3b color = img.at<Vec3b>(p);
- // set t-weights
- //計算每個頂點與Sink匯點t(代表背景)和源點Source(代表前景)連接的權值。
- //也即計算Gibbs能量(每一個像素點作爲背景像素或者前景像素)的第一個能量項
- double fromSource, toSink;
- if( mask.at<uchar>(p) == GC_PR_BGD || mask.at<uchar>(p) == GC_PR_FGD )
- {
- //對每一個像素計算其作爲背景像素或者前景像素的第一個能量項,作爲分別與t和s點的連接權值
- fromSource = -log( bgdGMM(color) );
- toSink = -log( fgdGMM(color) );
- }
- else if( mask.at<uchar>(p) == GC_BGD )
- {
- //對於確定爲背景的像素點,它與Source點(前景)的連接爲0,與Sink點的連接爲lambda
- fromSource = 0;
- toSink = lambda;
- }
- else // GC_FGD
- {
- fromSource = lambda;
- toSink = 0;
- }
- //設置該頂點vtxIdx分別與Source點和Sink點的連接權值
- graph.addTermWeights( vtxIdx, fromSource, toSink );
- // set n-weights n-links
- //計算兩個鄰域頂點之間連接的權值。
- //也即計算Gibbs能量的第二個能量項(平滑項)
- if( p.x>0 )
- {
- double w = leftW.at<double>(p);
- graph.addEdges( vtxIdx, vtxIdx-1, w, w );
- }
- if( p.x>0 && p.y>0 )
- {
- double w = upleftW.at<double>(p);
- graph.addEdges( vtxIdx, vtxIdx-img.cols-1, w, w );
- }
- if( p.y>0 )
- {
- double w = upW.at<double>(p);
- graph.addEdges( vtxIdx, vtxIdx-img.cols, w, w );
- }
- if( p.x<img.cols-1 && p.y>0 )
- {
- double w = uprightW.at<double>(p);
- graph.addEdges( vtxIdx, vtxIdx-img.cols+1, w, w );
- }
- }
- }
- }
- //論文中:迭代最小化算法step 3:分割估計:最小割或者最大流算法
- /*
- Estimate segmentation using MaxFlow algorithm
- */
- static void estimateSegmentation( GCGraph<double>& graph, Mat& mask )
- {
- //通過最大流算法確定圖的最小割,也即完成圖像的分割
- graph.maxFlow();
- Point p;
- for( p.y = 0; p.y < mask.rows; p.y++ )
- {
- for( p.x = 0; p.x < mask.cols; p.x++ )
- {
- //通過圖分割的結果來更新mask,即最後的圖像分割結果。注意的是,永遠都
- //不會更新用戶指定爲背景或者前景的像素
- if( mask.at<uchar>(p) == GC_PR_BGD || mask.at<uchar>(p) == GC_PR_FGD )
- {
- if( graph.inSourceSegment( p.y*mask.cols+p.x /*vertex index*/ ) )
- mask.at<uchar>(p) = GC_PR_FGD;
- else
- mask.at<uchar>(p) = GC_PR_BGD;
- }
- }
- }
- }
- //最後的成果:提供給外界使用的偉大的API:grabCut
- /*
- ****參數說明:
- img——待分割的源圖像,必須是8位3通道(CV_8UC3)圖像,在處理的過程中不會被修改;
- mask——掩碼圖像,如果使用掩碼進行初始化,那麼mask保存初始化掩碼信息;在執行分割
- 的時候,也可以將用戶交互所設定的前景與背景保存到mask中,然後再傳入grabCut函
- 數;在處理結束之後,mask中會保存結果。mask只能取以下四種值:
- GCD_BGD(=0),背景;
- GCD_FGD(=1),前景;
- GCD_PR_BGD(=2),可能的背景;
- GCD_PR_FGD(=3),可能的前景。
- 如果沒有手工標記GCD_BGD或者GCD_FGD,那麼結果只會有GCD_PR_BGD或GCD_PR_FGD;
- rect——用於限定需要進行分割的圖像範圍,只有該矩形窗口內的圖像部分才被處理;
- bgdModel——背景模型,如果爲null,函數內部會自動創建一個bgdModel;bgdModel必須是
- 單通道浮點型(CV_32FC1)圖像,且行數只能爲1,列數只能爲13x5;
- fgdModel——前景模型,如果爲null,函數內部會自動創建一個fgdModel;fgdModel必須是
- 單通道浮點型(CV_32FC1)圖像,且行數只能爲1,列數只能爲13x5;
- iterCount——迭代次數,必須大於0;
- mode——用於指示grabCut函數進行什麼操作,可選的值有:
- GC_INIT_WITH_RECT(=0),用矩形窗初始化GrabCut;
- GC_INIT_WITH_MASK(=1),用掩碼圖像初始化GrabCut;
- GC_EVAL(=2),執行分割。
- */
- void cv::grabCut( InputArray _img, InputOutputArray _mask, Rect rect,
- InputOutputArray _bgdModel, InputOutputArray _fgdModel,
- int iterCount, int mode )
- {
- Mat img = _img.getMat();
- Mat& mask = _mask.getMatRef();
- Mat& bgdModel = _bgdModel.getMatRef();
- Mat& fgdModel = _fgdModel.getMatRef();
- if( img.empty() )
- CV_Error( CV_StsBadArg, "image is empty" );
- if( img.type() != CV_8UC3 )
- CV_Error( CV_StsBadArg, "image mush have CV_8UC3 type" );
- GMM bgdGMM( bgdModel ), fgdGMM( fgdModel );
- Mat compIdxs( img.size(), CV_32SC1 );
- if( mode == GC_INIT_WITH_RECT || mode == GC_INIT_WITH_MASK )
- {
- if( mode == GC_INIT_WITH_RECT )
- initMaskWithRect( mask, img.size(), rect );
- else // flag == GC_INIT_WITH_MASK
- checkMask( img, mask );
- initGMMs( img, mask, bgdGMM, fgdGMM );
- }
- if( iterCount <= 0)
- return;
- if( mode == GC_EVAL )
- checkMask( img, mask );
- const double gamma = 50;
- const double lambda = 9*gamma;
- const double beta = calcBeta( img );
- Mat leftW, upleftW, upW, uprightW;
- calcNWeights( img, leftW, upleftW, upW, uprightW, beta, gamma );
- for( int i = 0; i < iterCount; i++ )
- {
- GCGraph<double> graph;
- assignGMMsComponents( img, mask, bgdGMM, fgdGMM, compIdxs );
- learnGMMs( img, mask, compIdxs, bgdGMM, fgdGMM );
- constructGCGraph(img, mask, bgdGMM, fgdGMM, lambda, leftW, upleftW, upW, uprightW, graph );
- estimateSegmentation( graph, mask );
- }
- }