自动瘦脸与眼睛放大可以算作图像局部扭曲算法的一个应用，其参考文献可以追溯至1993年的一篇博士论文：Interactive Image Warping。

图像局部扭曲算法有三个：局部缩放（Local Scaling）算法、局部平移（Local Transition）算法和局部旋转（Local Rotation）算法。其中应用局部缩放算法可实现眼睛放大，局部平移算法则可用于实现瘦脸效果。当然，图像局部缩放算法只是眼睛放大算法流程中的最关键的一步，要实现自动眼睛放大算法还需要额外的步骤。简单来说，给一张美女头像，你首先需要应用自动人脸检测技术定位出图像中的眼睛位置；然后基于此位置坐标应用图像局部缩放算法。自动瘦脸算法流程类似，不同之处在于应用人脸检测技术得到人脸轮廓点，由这些轮廓坐标点应用局部平移算法得到瘦脸效果。

人脸检测可以使用OpenCV自带的人脸检测器，人脸对齐可以使用Face Alignment at3000 FPS via Regressing Local Binary Features中的算法，实现参见https://github.com/yulequan/face-alignment-in-3000fps
图像局部缩放算法 
至于图像局部缩放算法实现，文献中有伪码描述，这里我给出一个简单的OpenGL Shader可以用于实现眼睛放大（其实也可以缩小）。

 precision highp float;varying highp vec2 textureCoordinate; uniform sampler2D inputImageTexture; uniform highp float scaleRatio;// 缩放系数，0无缩放，大于0则放大 uniform highp float radius;// 缩放算法的作用域半径 uniform highp vec2 leftEyeCenterPosition; // 左眼控制点，越远变形越小 uniform highp vec2 rightEyeCenterPosition; // 右眼控制点 uniform float aspectRatio; // 所处理图像的宽高比 highp vec2 warpPositionToUse(vec2 centerPostion, vec2 currentPosition, float radius, float scaleRatio, float aspectRatio) { vec2 positionToUse = currentPosition; vec2 currentPositionToUse = vec2(currentPosition.x, currentPosition.y * aspectRatio + 0.5 - 0.5 * aspectRatio); vec2 centerPostionToUse = vec2(centerPostion.x, centerPostion.y * aspectRatio + 0.5 - 0.5 * aspectRatio); float r = distance(currentPositionToUse, centerPostionToUse); if(r < radius) { float alpha = 1.0 - scaleRatio * pow(r / radius - 1.0, 2.0); positionToUse = centerPostion + alpha * (currentPosition - centerPostion); } return positionToUse; } void main() { vec2 positionToUse = warpPositionToUse(leftEyeCenterPosition, textureCoordinate, radius, scaleRatio, aspectRatio); positionToUse = warpPositionToUse(rightEyeCenterPosition, positionToUse, radius, scaleRatio, aspectRatio); gl_FragColor = texture2D(inputImageTexture, positionToUse); }

图像局部平移算法

图像局部平移算法还是参见论文，多说无益，在此奉上对应Shader代码给需要的同学，可以实现瘦脸和肥脸。这里需要指定瘦脸的控制点，最多支持MAX_CONTOUR_POINT_COUNT个控制点。

 precision highp float;varying highp vec2 textureCoordinate; uniform sampler2D inputImageTexture; uniform highp float radius; uniform highp float aspectRatio; uniform float leftContourPoints[MAX_CONTOUR_POINT_COUNT*2]; uniform float rightContourPoints[MAX_CONTOUR_POINT_COUNT*2]; uniform float deltaArray[MAX_CONTOUR_POINT_COUNT]; uniform int arraySize; highp vec2 warpPositionToUse(vec2 currentPoint, vec2 contourPointA, vec2 contourPointB, float radius, float delta, float aspectRatio) { vec2 positionToUse = currentPoint; vec2 currentPointToUse = vec2(currentPoint.x, currentPoint.y * aspectRatio + 0.5 - 0.5 * aspectRatio); vec2 contourPointAToUse = vec2(contourPointA.x, contourPointA.y * aspectRatio + 0.5 - 0.5 * aspectRatio); float r = distance(currentPointToUse, contourPointAToUse); if(r < radius) { vec2 dir = normalize(contourPointB - contourPointA); float dist = radius * radius - r * r; float alpha = dist / (dist + (r-delta) * (r-delta)); alpha = alpha * alpha; positionToUse = positionToUse - alpha * delta * dir; } return positionToUse; } void main() { vec2 positionToUse = textureCoordinate; for(int i = 0; i < arraySize; i++) { positionToUse = warpPositionToUse(positionToUse, vec2(leftContourPoints[i * 2], leftContourPoints[i * 2 + 1]), vec2(rightContourPoints[i * 2], rightContourPoints[i * 2 + 1]), radius, deltaArray[i], aspectRatio); positionToUse = warpPositionToUse(positionToUse, vec2(rightContourPoints[i * 2], rightContourPoints[i * 2 + 1]), vec2(leftContourPoints[i * 2], leftContourPoints[i * 2 + 1]), radius, deltaArray[i], aspectRatio); } gl_FragColor = texture2D(inputImageTexture, positionToUse); }