人脸姿态校正算法 附完整C++示例代码

在一些特殊情况下，经常需要依据图像中的人脸，对图片进行倾斜矫正。

 

例如拍照角度幅度过大之类的情况，而进行人工矫正确实很叫人头大。

那是不是可以有一种算法，可以根据人脸的信息对图片进行角度的修复呢？

答案肯定是确认的。

 

再次例如，想要通过人脸的特征对人物的表情和情绪进行精准判断，

那么这个时候如果能确保人脸没有发现严重倾斜，无疑对准确率判断有一定的帮助。

 

那么假如一张图片只有一个人脸，其实很好判断，通过眼睛的位置的坐标，根据两眼的直线角度，

就可以计算出修正的角度。

然后旋转图片到对应角度即可。

但是如果，一张图片存在多张人脸的时候该怎么办？

有两种方法：

1.找到最大的那个人脸，以它为基准

2.找到频次最高的人脸角度，以频次为基准

当然在大多数情况，方法1是比较合理的。

这两个种情况就留给各位看官去实现了。

本人仅仅考虑一张人脸的情况，演示如何实现该功能。

倾斜角度计算的代码如下:

    float diffEyeX = right_eye_x - left_eye_x;
    float diffEyeY = right_eye_y - left_eye_y;

    float fAngle;
    float M_PI = 3.1415926535897932384626433832795f;
    if (fabs(diffEyeX) < 0.0000001f)
        fAngle = 0.f;
    else
        fAngle = atanf(diffEyeY / diffEyeX) * 180.0f / M_PI;

如果看不明白，需要好好补一下高中数学基础。

为了节约时间，直接复用《自动红眼移除算法 附c++完整代码》的代码。

增加函数如下:

void RotateBilinear(unsigned char *sourceData, int width, int height, int Channels, int RowBytes,
                    unsigned char *destinationData, int newWidth, int newHeight, float angle, bool keepSize = true,
                    int fillColorR = 255, int fillColorG = 255, int fillColorB = 255) {
    if (sourceData == NULL || destinationData == NULL) return;

    float oldXradius = (float) (width - 1) / 2;
    float oldYradius = (float) (height - 1) / 2;

    float newXradius = (float) (newWidth - 1) / 2;
    float newYradius = (float) (newHeight - 1) / 2;

    double MPI = 3.14159265358979323846;
    double angleRad = -angle * MPI / 180.0;
    float angleCos = (float) cos(angleRad);
    float angleSin = (float) sin(angleRad);

    int srcStride = RowBytes;
    int dstOffset = newWidth * Channels - ((Channels == 1) ? newWidth : newWidth * Channels);

    unsigned char fillR = fillColorR;
    unsigned char fillG = fillColorG;
    unsigned char fillB = fillColorB;

    unsigned char *src = (unsigned char *) sourceData;
    unsigned char *dst = (unsigned char *) destinationData;

    int ymax = height - 1;
    int xmax = width - 1;
    if (Channels == 1) {
        float cy = -newYradius;
        for (int y = 0; y < newHeight; y++) {
            float tx = angleSin * cy + oldXradius;
            float ty = angleCos * cy + oldYradius;

            float cx = -newXradius;
            for (int x = 0; x < newWidth; x++, dst++) {
                float ox = tx + angleCos * cx;
                float oy = ty - angleSin * cx;

                int ox1 = (int) ox;
                int oy1 = (int) oy;

                if ((ox1 < 0) || (oy1 < 0) || (ox1 >= width) || (oy1 >= height)) {
                    *dst = fillG;
                } else {
                    int ox2 = (ox1 == xmax) ? ox1 : ox1 + 1;
                    int oy2 = (oy1 == ymax) ? oy1 : oy1 + 1;
                    float dx1 = 0;
                    if ((dx1 = ox - (float) ox1) < 0)
                        dx1 = 0;
                    float dx2 = 1.0f - dx1;
                    float dy1 = 0;
                    if ((dy1 = oy - (float) oy1) < 0)
                        dy1 = 0;
                    float dy2 = 1.0f - dy1;

                    unsigned char *p1 = src + oy1 * srcStride;
                    unsigned char *p2 = src + oy2 * srcStride;

                    *dst = (unsigned char) (dy2 * (dx2 * p1[ox1] + dx1 * p1[ox2]) +
                                            dy1 * (dx2 * p2[ox1] + dx1 * p2[ox2]));
                }
                cx++;
            }
            cy++;
            dst += dstOffset;
        }
    } else if (Channels == 3) {
        float cy = -newYradius;
        for (int y = 0; y < newHeight; y++) {
            float tx = angleSin * cy + oldXradius;
            float ty = angleCos * cy + oldYradius;

            float cx = -newXradius;
            for (int x = 0; x < newWidth; x++, dst += Channels) {
                float ox = tx + angleCos * cx;
                float oy = ty - angleSin * cx;

                int ox1 = (int) ox;
                int oy1 = (int) oy;

                if ((ox1 < 0) || (oy1 < 0) || (ox1 >= width) || (oy1 >= height)) {
                    dst[0] = fillR;
                    dst[1] = fillG;
                    dst[2] = fillB;
                } else {
                    int ox2 = (ox1 == xmax) ? ox1 : ox1 + 1;
                    int oy2 = (oy1 == ymax) ? oy1 : oy1 + 1;

                    float dx1 = 0;
                    if ((dx1 = ox - (float) ox1) < 0)
                        dx1 = 0;
                    float dx2 = 1.0f - dx1;
                    float dy1 = 0;
                    if ((dy1 = oy - (float) oy1) < 0)
                        dy1 = 0;
                    float dy2 = 1.0f - dy1;

                    unsigned char *p1 = src + oy1 * srcStride;
                    unsigned char *p2 = p1;
                    p1 += ox1 * Channels;
                    p2 += ox2 * Channels;

                    unsigned char *p3 = src + oy2 * srcStride;
                    unsigned char *p4 = p3;
                    p3 += ox1 * Channels;
                    p4 += ox2 * Channels;

                    dst[0] = (unsigned char) (
                            dy2 * (dx2 * p1[0] + dx1 * p2[0]) +
                            dy1 * (dx2 * p3[0] + dx1 * p4[0]));

                    dst[1] = (unsigned char) (
                            dy2 * (dx2 * p1[1] + dx1 * p2[1]) +
                            dy1 * (dx2 * p3[1] + dx1 * p4[1]));

                    dst[2] = (unsigned char) (
                            dy2 * (dx2 * p1[2] + dx1 * p2[2]) +
                            dy1 * (dx2 * p3[2] + dx1 * p4[2]));
                }
                cx++;
            }
            cy++;
            dst += dstOffset;
        }
    } else if (Channels == 4) {
        float cy = -newYradius;
        for (int y = 0; y < newHeight; y++) {
            float tx = angleSin * cy + oldXradius;
            float ty = angleCos * cy + oldYradius;

            float cx = -newXradius;
            for (int x = 0; x < newWidth; x++, dst += Channels) {
                float ox = tx + angleCos * cx;
                float oy = ty - angleSin * cx;

                int ox1 = (int) ox;
                int oy1 = (int) oy;

                if ((ox1 < 0) || (oy1 < 0) || (ox1 >= width) || (oy1 >= height)) {
                    dst[0] = fillR;
                    dst[1] = fillG;
                    dst[2] = fillB;
                    dst[3] = 255;
                } else {
                    int ox2 = (ox1 == xmax) ? ox1 : ox1 + 1;
                    int oy2 = (oy1 == ymax) ? oy1 : oy1 + 1;

                    float dx1 = 0;
                    if ((dx1 = ox - (float) ox1) < 0)
                        dx1 = 0;
                    float dx2 = 1.0f - dx1;
                    float dy1 = 0;
                    if ((dy1 = oy - (float) oy1) < 0)
                        dy1 = 0;
                    float dy2 = 1.0f - dy1;

                    unsigned char *p1 = src + oy1 * srcStride;
                    unsigned char *p2 = p1;
                    p1 += ox1 * Channels;
                    p2 += ox2 * Channels;

                    unsigned char *p3 = src + oy2 * srcStride;
                    unsigned char *p4 = p3;
                    p3 += ox1 * Channels;
                    p4 += ox2 * Channels;

                    dst[0] = (unsigned char) (
                            dy2 * (dx2 * p1[0] + dx1 * p2[0]) +
                            dy1 * (dx2 * p3[0] + dx1 * p4[0]));

                    dst[1] = (unsigned char) (
                            dy2 * (dx2 * p1[1] + dx1 * p2[1]) +
                            dy1 * (dx2 * p3[1] + dx1 * p4[1]));

                    dst[2] = (unsigned char) (
                            dy2 * (dx2 * p1[2] + dx1 * p2[2]) +
                            dy1 * (dx2 * p3[2] + dx1 * p4[2]));
                    dst[3] = 255;
                }
                cx++;
            }
            cy++;
            dst += dstOffset;
        }
    }
}

void facialPoseCorrection(unsigned char *inputImage, int Width, int Height, int Channels, int left_eye_x, int left_eye_y,
                    int right_eye_x, int right_eye_y) {
    float diffEyeX = right_eye_x - left_eye_x;
    float diffEyeY = right_eye_y - left_eye_y;

    float fAngle;
    float M_PI = 3.1415926535897932384626433832795f;
    if (fabs(diffEyeX) < 0.0000001f)
        fAngle = 0.f;
    else
        fAngle = atanf(diffEyeY / diffEyeX) * 180.0f / M_PI;
    size_t numberOfPixels = Width * Height * Channels * sizeof(unsigned char);
    unsigned char *outputImage = (unsigned char *) malloc(numberOfPixels);
    if (outputImage != nullptr) {
        RotateBilinear(inputImage, Width, Height, Channels, Width * Channels, outputImage, Width, Height, fAngle);
        memcpy(inputImage, outputImage, numberOfPixels);
        free(outputImage);
    }
}

上效果图片。

原图：

红眼修复+倾斜矫正:

项目地址:

https://github.com/cpuimage/MTCNN

命令行参数:

mtcnn 模型文件路径 图片路径

例如: mtcnn ../models ../sample.jpg

 

用cmake即可进行编译示例代码，详情见CMakeLists.txt。

 

若有其他相关问题或者需求也可以邮件联系俺探讨。

邮箱地址是: 
gaozhihan@vip.qq.com