OpenCV基于Landmark实现人脸交换

OpenCV学堂

发布于 2018-07-26 10:37:31

1.2K0

发布于 2018-07-26 10:37:31

OpenCV基于Landmark实现人脸交换

川剧是中国最知名的戏曲剧种之一，变脸是川剧表演的特技之一，在对象传承上有着严格的师门派别。有点扯远啦，回来！其实主要是我们今天要用OpenCV干的事情跟这个有点关系，OpenCV基于Landmark实现人脸关键点提取，对结果善加利用可以实现人脸交换，对特定对象施加变脸术。OpenCV开发者不学川剧也一样可以给各种人变脸，当然前提是会写代码，会做OpenCV。首先简单说一下原理与流程。

一：原理与流程

基本原理是利用OpenCV的级联检测器实现人脸检测然后基于Landmak的LBF模型实现人脸68个关键点提取，基于关键点数据实现三角剖分与维诺图计算，经过几何变换之后得到mask区域，再利用OpenCV无缝克隆算法相关API实现换脸。整个工作流程如下：

二：代码实现

1.创建Landmark关键点检测器

// 人脸检测与Landmark68个关键点检测
CascadeClassifier face_cascade;
face_cascade.load(cascade_name);
FacemarkLBF::Params params;
params.n_landmarks = 68; // 68个标注点
params.initShape_n = 10;
params.stages_n = 5; // 算法的5个强化步骤
params.tree_n = 6; // 模型中每个标注点结构树 数目
params.tree_depth = 5; // 决策树深度

// 创建LBF landmark 检测器
Ptr<FacemarkLBF> facemark = FacemarkLBF::create(params);
facemark->setFaceDetector((FN_FaceDetector)myDetector, &face_cascade);
facemark->loadModel(modelfile_name);
cout << "Loaded model" << endl;

2.Landmark关键点检测

//vector to store the faces detected in the image
vector<Rect> faces1, faces2;
vector< vector<Point2f> > shape1, shape2;

//Detect faces in the current image
float ratio1 = (float)img1.cols / (float)img1.rows;
float ratio2 = (float)img2.cols / (float)img2.rows;
resize(img1, img1, Size((int)(640 * ratio1), (int)(640 * ratio1)), 0, 0, INTER_LINEAR_EXACT);
resize(img2, img2, Size((int)(640 * ratio2), (int)(640 * ratio2)), 0, 0, INTER_LINEAR_EXACT);

Mat img1Warped = img2.clone();
facemark->getFaces(img1, faces1);
facemark->getFaces(img2, faces2);

//Initialise the shape of the faces
facemark->fit(img1, faces1, shape1);
facemark->fit(img2, faces2, shape2);

3.三角剖分变换

vector<Point2f> points1 = shape1[z];
vector<Point2f> points2 = shape2[z];
img1.convertTo(img1, CV_32F);
img1Warped.convertTo(img1Warped, CV_32F);
// Find convex hull
vector<Point2f> boundary_image1;
vector<Point2f> boundary_image2;
vector<int> index;
convexHull(Mat(points2), index, false, false);
for (size_t i = 0; i < index.size(); i++)
{
    boundary_image1.push_back(points1[index[i]]);
    boundary_image2.push_back(points2[index[i]]);
}
// Triangulation for points on the convex hull
vector< vector<int> > triangles;
Rect rect(0, 0, img1Warped.cols, img1Warped.rows);
divideIntoTriangles(rect, boundary_image2, triangles);
// Apply affine transformation to Delaunay triangles
for (size_t i = 0; i < triangles.size(); i++)
{
    vector<Point2f> triangle1, triangle2;
    // Get points for img1, img2 corresponding to the triangles
    for (int j = 0; j < 3; j++)
    {
        triangle1.push_back(boundary_image1[triangles[i][j]]);
        triangle2.push_back(boundary_image2[triangles[i][j]]);
    }
    warpTriangle(img1, img1Warped, triangle1, triangle2);
}

4.计算与模板生成

// 计算与生成模板
vector<Point> hull;
for (size_t i = 0; i < boundary_image2.size(); i++)
{
    Point pt((int)boundary_image2[i].x, (int)boundary_image2[i].y);
    hull.push_back(pt);
}
Mat mask = Mat::zeros(img2.rows, img2.cols, img2.depth());
fillConvexPoly(mask, &hull[0], (int)hull.size(), Scalar(255, 255, 255));

5.无缝克隆

//    无缝克隆 
Rect r = boundingRect(boundary_image2);
Point center = (r.tl() + r.br()) / 2;
Mat output;
img1Warped.convertTo(img1Warped, CV_8UC3);
seamlessClone(img1Warped, img2, mask, center, output, NORMAL_CLONE);
imshow("Face_Swapped", output);
imwrite("D:/face_swap_demo.png", output);