opengl 实现瘦脸等脸部微调

文章目录

美颜瘦脸先看效果先上源码实现识别人脸并定位特征点人脸分割瘦脸变形 像素偏移后结果去掉特征点后结果

美颜瘦脸

在这个靠脸吃饭的时代，有张漂亮的脸蛋无疑会令你加分不少；那万一天生的颜值不够怎么办呢？。。。还有美颜来拯救。

现在很多相机以及大多数修图软件都提供了瘦脸的功能。现在主流的瘦脸功能都是基于opengl来完成的。

先看效果

原图：

瘦脸：

很明显下面的照片要比上面的照片的脸要瘦了很多。 下面一步步去实现这些功能。

先上源码

VS2019的项目，环境完整，可直接运行

实现

识别人脸并定位特征点

用dlib或者face++或者自己训练个神经网络去识别出特征点。 这里我用dlib 识别出特征点

初始化dlib：

dlib::frontal_face_detector detector; dlib::shape_predictor pose_model; std::string facemodel = "./model/shape_predictor_68_face_landmarks.dat"; detector = dlib::get_frontal_face_detector(); dlib::deserialize(facemodel) >> pose_model;

识别人脸并保存特征点

dlib::cv_image<dlib::bgr_pixel> cimg(src); std::vector<dlib::rectangle> faces = detector(cimg); // Detect faces dlib::full_object_detection shape; // Find the pose of one face facelandmarks.clear(); if (faces.size() > 0) { shape = pose_model(cimg, faces[0]); facerect = cv::Rect(cv::Point2i(faces[0].left(), faces[0].top()), cv::Point2i(faces[0].right() + 1, faces[0].bottom() + 1)); for (int i = 0; i < 68; i++) { facelandmarks.push_back(cv::Point2f(shape.part(i).x(), shape.part(i).y())); } } else { return -1; }

识别的特征点如下 共68个：

人脸分割

这里为了简单只将图片分为26个三角形。若想效果更平滑，可以将图片分的更细。

//人脸分割 cv::Point2f LALL0 = cv::Point2f(-1.0f, -1.0f); cv::Point2f LALL1 = cv::Point2f(-1.0f, 1.0f); cv::Point2f LALL2 = cv::Point2f(1.0f, 1.0f); cv::Point2f LALL3 = cv::Point2f(1.0f, -1.0f); cv::Point2f LFace30 = cv::Point2f(facelandmarks[30].x * 2.0 / width - 1.0, (height - facelandmarks[30].y) * 2.0 / height - 1.0); cv::Point2f LFace3 = cv::Point2f(facelandmarks[3].x * 2.0 / width - 1.0, (height - facelandmarks[3].y) * 2.0 / height - 1.0); cv::Point2f LFace4 = cv::Point2f(facelandmarks[4].x * 2.0 / width - 1.0, (height - facelandmarks[4].y) * 2.0 / height - 1.0); cv::Point2f LFace5 = cv::Point2f(facelandmarks[5].x * 2.0 / width - 1.0, (height - facelandmarks[5].y) * 2.0 / height - 1.0); cv::Point2f LFace6 = cv::Point2f(facelandmarks[6].x * 2.0 / width - 1.0, (height - facelandmarks[6].y) * 2.0 / height - 1.0); cv::Point2f LFace7 = cv::Point2f(facelandmarks[7].x * 2.0 / width - 1.0, (height - facelandmarks[7].y) * 2.0 / height - 1.0); cv::Point2f LFace8 = cv::Point2f(facelandmarks[8].x * 2.0 / width - 1.0, (height - facelandmarks[8].y) * 2.0 / height - 1.0); cv::Point2f LFace9 = cv::Point2f(facelandmarks[9].x * 2.0 / width - 1.0, (height - facelandmarks[9].y) * 2.0 / height - 1.0); cv::Point2f LFace10 = cv::Point2f(facelandmarks[10].x * 2.0 / width - 1.0, (height - facelandmarks[10].y) * 2.0 / height - 1.0); cv::Point2f LFace11 = cv::Point2f(facelandmarks[11].x * 2.0 / width - 1.0, (height - facelandmarks[11].y) * 2.0 / height - 1.0); cv::Point2f LFace12 = cv::Point2f(facelandmarks[12].x * 2.0 / width - 1.0, (height - facelandmarks[12].y) * 2.0 / height - 1.0); cv::Point2f LFace13 = cv::Point2f(facelandmarks[13].x * 2.0 / width - 1.0, (height - facelandmarks[13].y) * 2.0 / height - 1.0); cv::Point2f VALL0 = cv::Point2f(0.0f, 1.0f); cv::Point2f VALL1 = cv::Point2f(0.0f, 0.0f); cv::Point2f VALL2 = cv::Point2f(1.0f, 0.0f); cv::Point2f VALL3 = cv::Point2f(1.0f, 1.0f); cv::Point2f VFace30 = cv::Point2f(facelandmarks[30].x / width, facelandmarks[30].y / height); cv::Point2f VFace3 = cv::Point2f(facelandmarks[3].x / width, facelandmarks[3].y / height); cv::Point2f VFace4 = cv::Point2f(facelandmarks[4].x / width, facelandmarks[4].y / height); cv::Point2f VFace5 = cv::Point2f(facelandmarks[5].x / width, facelandmarks[5].y / height); cv::Point2f VFace6 = cv::Point2f(facelandmarks[6].x / width, facelandmarks[6].y / height); cv::Point2f VFace7 = cv::Point2f(facelandmarks[7].x / width, facelandmarks[7].y / height); cv::Point2f VFace8 = cv::Point2f(facelandmarks[8].x / width, facelandmarks[8].y / height); cv::Point2f VFace9 = cv::Point2f(facelandmarks[9].x / width, facelandmarks[9].y / height); cv::Point2f VFace10 = cv::Point2f(facelandmarks[10].x / width, facelandmarks[10].y / height); cv::Point2f VFace11 = cv::Point2f(facelandmarks[11].x / width, facelandmarks[11].y / height); cv::Point2f VFace12 = cv::Point2f(facelandmarks[12].x / width, facelandmarks[12].y / height); cv::Point2f VFace13 = cv::Point2f(facelandmarks[13].x / width, facelandmarks[13].y / height); float vertices[] = { //----位置---- ---纹理--- LALL0.x, LALL0.y, 0.0f, VALL0.x, VALL0.y, LALL1.x, LALL1.y, 0.0f, VALL1.x, VALL1.y, LALL2.x, LALL2.y, 0.0f, VALL2.x, VALL2.y, LALL3.x, LALL3.y, 0.0f, VALL3.x, VALL3.y, LFace30.x, LFace30.y, 0.0f, VFace30.x, VFace30.y, LFace3.x, LFace3.y, 0.0f, VFace3.x, VFace3.y, LFace4.x, LFace4.y, 0.0f, VFace4.x, VFace4.y, LFace5.x, LFace5.y, 0.0f, VFace5.x, VFace5.y, LFace6.x, LFace6.y, 0.0f, VFace6.x, VFace6.y, LFace7.x, LFace7.y, 0.0f, VFace7.x, VFace7.y, LFace8.x, LFace8.y, 0.0f, VFace8.x, VFace8.y, LFace9.x, LFace9.y, 0.0f, VFace9.x, VFace9.y, LFace10.x, LFace10.y, 0.0f, VFace10.x, VFace10.y, LFace11.x, LFace11.y, 0.0f, VFace11.x, VFace11.y, LFace12.x, LFace12.y, 0.0f, VFace12.x, VFace12.y, LFace13.x, LFace13.y, 0.0f, VFace13.x, VFace13.y }; unsigned int indices[] = { 0,1,5, 5,1,4, 4,1,2, 2,4,15, 15,2,3, 0,5,6, 0,6,7, 0,7,8, 0,8,9, 0,9,10, 0,10,3, 3,10,11, 3,11,12, 3,12,13, 3,13,14, 3,14,15, 4,5,6, 4,6,7, 4,7,8, 4,8,9, 4,9,10, 4,10,11, 4,11,12, 4,12,13, 4,13,14, 4,14,15 };

瘦脸变形

只需要将图中标记的点按照箭头方向拖动，即可实现瘦脸，为了不显得太过于尖锐，我们将范围包括到附近的几个点。

关于像素偏移调节，可以参考这篇文章

shader:

#version 330 core precision mediump float; in vec2 TexCoord; out vec4 outColor; uniform sampler2D inputTexture; uniform float face5x; uniform float face5y; vec2 stretchFun(vec2 textureCoord, vec2 originPosition, vec2 targetPosition, float radius,float curve) { vec2 direction = targetPosition - originPosition; float infect = distance(textureCoord, originPosition)/radius; infect =1.0 - pow(infect,curve); infect = clamp(infect,0.0,1.0); vec2 offset = direction * infect; vec2 result = textureCoord - offset; return result; } void main(){ vec2 A1 = vec2(face5x,face5y); vec2 A2 = vec2(face5x+0.02f,face5y + 0.01f); vec2 TexCoord2 = stretchFun(TexCoord,A1,A2,0.19f,2.0f); vec3 tmpColor = texture(inputTexture, TexCoord2).rgb; outColor = vec4(tmpColor,1.0f); }

像素偏移后结果

去掉特征点后结果

完成