OpenCV影象分割實戰C++（一）Grabcut摳圖與證件照背景替換
阿新 • • 發佈：2019-01-23
Grabcut摳圖
步驟：
- 輸入原影象
- 矩形輸入
- 開始分類
- GMM描述
- GMM訓練分類
- Graph Cut分類
- 不斷迭代直到收斂分類
API：
void grabCut(
InputArray img, // 待分割影象，8bit，3通道
// 輸入輸出引數，儲存處理後的結果，8bit單通道掩碼(與img同rows cols)，mask元素值只能為 GC_BGD, GC_FGD, GC_PR_BGD, GC_PR_FGD 之一
InputOutputArray mask, // 如果沒有手動標記 GC_BGD或GC_FGD ，那麼結果只會有 GC_PR_BGD或GC_PR_FGD
Rect rect, // 當 mode=GC_INIT_WITH_RECT時使用，rect外部的為GC_BGD，rect內部的為GC_FGD 程式碼：
#include <opencv2/opencv.hpp>
#include <opencv2/xfeatures2d.hpp>
#include<opencv2/face.hpp> 結果：
證件照背景替換
步驟：
- 開始
- 資料組裝
- KMeans分割
- 背景去除
- 遮罩生成
- 遮罩模糊
- 通道混合輸出
程式碼：
#include <opencv2/opencv.hpp>
#include <opencv2/xfeatures2d.hpp>
#include<opencv2/face.hpp>
#include<iostream>
#include<math.h>
#include <string>
#include<fstream>
using namespace cv::face;
using namespace cv;
using namespace std;
using namespace cv::xfeatures2d;
int main() {
Mat src = imread("C:/Users/Administrator/Desktop/pic/z5.jpg");
imshow("src", src);
//組裝資料
int width = src.cols;
int height = src.rows;
int samplecount = width * height;
int dims = src.channels();
//行數為src的畫素點數，列數為通道數，每列資料分別為src的bgr，從上到下 從左到右順序讀資料
Mat points(samplecount, dims, CV_32F, Scalar(10));
int ind = 0;
for (int row = 0; row < height; row++) {
for (int col = 0; col < width; col++) {
ind = row * width + col;//
Vec3b bgr = src.at<Vec3b>(row, col);
points.at<float>(ind, 0) = static_cast<int>(bgr[0]);
points.at<float>(ind, 1) = static_cast<int>(bgr[1]);
points.at<float>(ind, 2) = static_cast<int>(bgr[2]);
}
}
//執行kmeans
int numCluster = 4;
Mat labels;
Mat centers;
TermCriteria criteria = TermCriteria(TermCriteria::EPS + TermCriteria::COUNT, 10, 0.1);
kmeans(points, numCluster, labels, criteria, 3, KMEANS_PP_CENTERS, centers);
//去背景+遮罩生成
Mat mask = Mat::zeros(src.size(), CV_8UC1);
int index = src.rows * 2 + 2;//不取邊緣的左上點，往裡靠2個位置
int cindex = labels.at<int>(index, 0);
int height1 = src.rows;
int width1 = src.cols;
Mat dst;//人的輪廓周圍會有一些雜點，所以需要腐蝕和高斯模糊取干擾
src.copyTo(dst);
for (int row = 0; row < height1; row++) {
for (int col = 0; col < width1; col++) {
index = row * width1 + col;
int label = labels.at<int>(index, 0);
if (label == cindex) {
dst.at<Vec3b>(row, col)[0] = 0;
dst.at<Vec3b>(row, col)[1] = 0;
dst.at<Vec3b>(row, col)[2] = 0;
mask.at<uchar>(row, col) = 0;
}
else {
dst.at<Vec3b>(row, col) = src.at<Vec3b>(row, col);
mask.at<uchar>(row, col) = 255;//人臉部分設為白色，以便於下面的腐蝕與高斯模糊
}
}
}
imshow("dst", dst);
imshow("mask", dst);
//腐蝕+高斯模糊
Mat k = getStructuringElement(MORPH_RECT, Size(3, 3));
erode(mask, mask, k);
GaussianBlur(mask, mask, Size(3, 3), 0, 0);
imshow("gaosimohu", mask);
//通道混合
RNG rng(12345);
Vec3b color;
color[0] = 180;//rng.uniform(0, 255);
color[1] =180;//rng.uniform(0, 255);
color[2] =238;//rng.uniform(0, 255);
Mat result(src.size(), src.type());
double w = 0.0;
int b = 0, g = 0, r = 0;
int b1 = 0, g1 = 0, r1 = 0;
int b2 = 0, g2 = 0, r2 = 0;
double time = getTickCount();
for (int row = 0; row < height1; row++) {
for (int col = 0; col < width; col++) {
int m = mask.at<uchar>(row, col);
if (m == 255) {
result.at<Vec3b>(row, col) = src.at<Vec3b>(row, col);//前景
}
else if (m == 0) {
result.at<Vec3b>(row, col) = color; // 背景
}
else {//因為高斯模糊的關係，所以mask元素的顏色除了黑白色還有黑白邊緣經過模糊後的非黑白值
w = m / 255.0;
b1 = src.at<Vec3b>(row, col)[0];
g1 = src.at<Vec3b>(row, col)[1];
r1 = src.at<Vec3b>(row, col)[2];
b2 = color[0];
g2 = color[0];
r2 = color[0];
b = b1 * w + b2 * (1.0 - w);
g = g1 * w + g2 * (1.0 - w);
r = r1 * w + r2 * (1.0 - w);
result.at<Vec3b>(row, col)[0] = b;//最終邊緣顏色值
result.at<Vec3b>(row, col)[1] = g;
result.at<Vec3b>(row, col)[2] = r;
}
}
}
cout << "time=" << (getTickCount() - time) / getTickFrequency() << endl;
imshow("backgroud repalce", result);
waitKey(0);
}
結果：
