博客

最近需要做人脸对齐的算法，通俗理解就是将图片人人脸姿态不太正确的给矫正过来，所以写了python版本的人脸对齐算法。也有对应的C++版本，点击

这里

。​​​​​​​基本原理是先通过MTCNN检测到人脸的五个关键点，再把原图中人脸区域外扩100%（这样做的目的是保证对齐后图片中没有黑色区域，当然这个外扩的比例是看对齐效果自己可以调节的，我这里设置的100%）。最后的人脸对齐尺寸分为两种：112X96尺寸和112X112尺寸，其中首先需要定死仿射变换后人脸在目标图上的坐标，然后直接变换。废话不多说，直接手撕代码。

# 该代码实现利用人脸的五点仿射变换实现人脸对齐
# 具体就是首先使用mtcnn检测算法检测出人脸区域，并得到lanmarks关键点坐标和检测框坐标
# 之后对人脸区域外扩60%，然后对该外扩后的区域重新得到关键点，进行五点仿射变换得到即可。
# 参考链接：https://blog.csdn.net/oTengYue/article/details/79278572
# _*_ coding:utf-8 _*_
import os
import cv2
import numpy
import logging
import tensorflow as tf
from detection.mtcnn import detect_face

logging.basicConfig(
	level=logging.DEBUG,
	format='%(asctime)s %(levelname)s: %(message)s',
	datefmt='%Y-%m-%d %H:%M:%S'
)
logger = logging.getLogger(__name__)

# 最终的人脸对齐图像尺寸分为两种：112x96和112x112，并分别对应结果图像中的两组仿射变换目标点,如下所示
imgSize1 = [112,96]
imgSize2 = [112,112]
coord5point1 = [[30.2946, 51.6963],  # 112x96的目标点
               [65.5318, 51.6963],
               [48.0252, 71.7366],
               [33.5493, 92.3655],
               [62.7299, 92.3655]]
coord5point2 = [[30.2946+8.0000, 51.6963], # 112x112的目标点
               [65.5318+8.0000, 51.6963],
               [48.0252+8.0000, 71.7366],
               [33.5493+8.0000, 92.3655],
               [62.7299+8.0000, 92.3655]]

def transformation_from_points(points1, points2):
    points1 = points1.astype(numpy.float64)
    points2 = points2.astype(numpy.float64)
    c1 = numpy.mean(points1, axis=0)
    c2 = numpy.mean(points2, axis=0)
    points1 -= c1
    points2 -= c2
    s1 = numpy.std(points1)
    s2 = numpy.std(points2)
    points1 /= s1
    points2 /= s2
    U, S, Vt = numpy.linalg.svd(points1.T * points2)
    R = (U * Vt).T
    return numpy.vstack([numpy.hstack(((s2 / s1) * R,c2.T - (s2 / s1) * R * c1.T)),numpy.matrix([0., 0., 1.])])

def warp_im(img_im, orgi_landmarks,tar_landmarks):
    pts1 = numpy.float64(numpy.matrix([[point[0], point[1]] for point in orgi_landmarks]))
    pts2 = numpy.float64(numpy.matrix([[point[0], point[1]] for point in tar_landmarks]))
    M = transformation_from_points(pts1, pts2)
    dst = cv2.warpAffine(img_im, M[:2], (img_im.shape[1], img_im.shape[0]))
    return dst

def main():
    # 对一个路径下的所有图片进行两种方式对齐，并保存
    pic_path = 'C:\\Users\\pfm\\Desktop\\test_image\\'
    #加载mtcnn参数
    with tf.Graph().as_default():
        sess = tf.Session()
        pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
    minsize = 50  # minimum size of face
    threshold = [0.6, 0.7, 0.7]  # three steps's threshold
    factor = 0.709  # scale factor

    # Size Parameter
    lower_threshold = 100
    upper_threshold = 200
    num = 0

    pic_name_list = os.listdir(pic_path)
    for every_pic_name in pic_name_list:
        img_im = cv2.imread(pic_path + every_pic_name)
        #关键点检测
        if img_im is None:
            continue
        else:
            shape = img_im.shape
            height = shape[0]
            width = shape[1]
            bounding_boxes, points = detect_face.detect_face(img_im, minsize, pnet, rnet, onet, threshold, factor)
            # 处理该张图片中的每个框
            if bounding_boxes.shape[0] > 0:
                for i in range(bounding_boxes.shape[0]):  # 根据行号得到每张图片有多少个回归框
                    x1, y1, x2, y2 = int(min(bounding_boxes[i][0], min(points[:, i][:5]))), \
                                     int(min(bounding_boxes[i][1], min(points[:, i][5:]))), \
                                     int(max(bounding_boxes[i][2], max(points[:, i][:5]))), \
                                     int(max(bounding_boxes[i][3], max(points[:, i][5:])))
                    # 外扩大100%，防止对齐后人脸出现黑边
                    new_x1 = max(int(1.50 * x1 - 0.50 * x2),0)
                    new_x2 = min(int(1.50 * x2 - 0.50 * x1),width-1)
                    new_y1 = max(int(1.50 * y1 - 0.50 * y2),0)
                    new_y2 = min(int(1.50 * y2 - 0.50 * y1),height-1)
                    # new_x1 = max(int(1.30 * x1 - 0.30 * x2),0)
                    # new_x2 = min(int(1.30 * x2 - 0.30 * x1),width-1)
                    # new_y1 = max(int(1.30 * y1 - 0.30 * y2),0)
                    # new_y2 = min(int(1.30 * y2 - 0.30 * y1),height-1)

                    # 得到原始图中关键点坐标
                    left_eye_x = points[:, i][:5][0]
                    right_eye_x = points[:, i][:5][1]
                    nose_x = points[:, i][:5][2]
                    left_mouth_x = points[:, i][:5][3]
                    right_mouth_x = points[:, i][:5][4]
                    left_eye_y = points[:, i][5:][0]
                    right_eye_y = points[:, i][5:][1]
                    nose_y = points[:, i][5:][2]
                    left_mouth_y = points[:, i][5:][3]
                    right_mouth_y = points[:, i][5:][4]

                    # 得到外扩100%后图中关键点坐标
                    new_left_eye_x = left_eye_x - new_x1
                    new_right_eye_x = right_eye_x - new_x1
                    new_nose_x = nose_x - new_x1
                    new_left_mouth_x = left_mouth_x - new_x1
                    new_right_mouth_x = right_mouth_x - new_x1
                    new_left_eye_y = left_eye_y - new_y1
                    new_right_eye_y = right_eye_y - new_y1
                    new_nose_y = nose_y - new_y1
                    new_left_mouth_y = left_mouth_y - new_y1
                    new_right_mouth_y = right_mouth_y - new_y1

                    face_landmarks = [[new_left_eye_x,new_left_eye_y], # 在扩大100%人脸图中关键点坐标
                                      [new_right_eye_x,new_right_eye_y],
                                      [new_nose_x,new_nose_y],
                                      [new_left_mouth_x,new_left_mouth_y],
                                      [new_right_mouth_x,new_right_mouth_y]]
                    face = img_im[new_y1: new_y2, new_x1: new_x2] # 扩大100%的人脸区域
                    dst1 = warp_im(face,face_landmarks,coord5point1) # 112x96对齐后尺寸
                    dst2 = warp_im(face,face_landmarks,coord5point2) # 112x112对齐后尺寸
                    crop_im1 = dst1[0:imgSize1[0],0:imgSize1[1]]
                    crop_im2 = dst2[0:imgSize2[0],0:imgSize2[1]]
                    cv2.imwrite(pic_path + every_pic_name[:-4] + '_' + str(num) + '_align_112x96.jpg',crop_im1)
                    cv2.imwrite(pic_path + every_pic_name[:-4] + '_' + str(num) + '_align_112x112.jpg',crop_im2)
                    num = num + 1

if __name__ == '__main__':
    main()
    cv2.waitKey()
    pass

接下来是运行的效果：

如果想运行这份代码，可以点击

这里

（code: djcb）下载完整的工程即可。