AI换脸

使用opencv-python和dlib实现的简单换脸程序

 

faceswap.py

# -*- coding: utf-8 -*-

import os
import cv2
import dlib
import numpy as np

here = os.path.dirname(os.path.abspath(__file__))

models_folder_path = os.path.join(here, 'models')  # 模型保存文件夹
faces_folder_path = os.path.join(here, 'faces')  # 人脸图片保存文件夹
predictor_path = os.path.join(models_folder_path, 'shape_predictor_68_face_landmarks.dat')  # 模型路径
image_face_path = os.path.join(faces_folder_path, 'JayChou.png')  # 人脸图片路径

detector = dlib.get_frontal_face_detector()  # dlib的正向人脸检测器
predictor = dlib.shape_predictor(predictor_path)  # dlib的人脸形状检测器


def get_image_size(image):
    """
    获取图片大小（高度,宽度）
    :param image: image
    :return: （高度,宽度）
    """
    image_size = (image.shape[0], image.shape[1])
    return image_size


def get_face_landmarks(image, face_detector, shape_predictor):
    """
    获取人脸标志，68个特征点
    :param image: image
    :param face_detector: dlib.get_frontal_face_detector
    :param shape_predictor: dlib.shape_predictor
    :return: np.array([[],[]]), 68个特征点
    """
    dets = face_detector(image, 1)
    num_faces = len(dets)
    if num_faces == 0:
        print("Sorry, there were no faces found.")
        return None
    shape = shape_predictor(image, dets[0])
    face_landmarks = np.array([[p.x, p.y] for p in shape.parts()])
    return face_landmarks


def get_face_mask(image_size, face_landmarks):
    """
    获取人脸掩模
    :param image_size: 图片大小
    :param face_landmarks: 68个特征点
    :return: image_mask, 掩模图片
    """
    mask = np.zeros(image_size, dtype=np.uint8)
    points = np.concatenate([face_landmarks[0:16], face_landmarks[26:17:-1]])
    cv2.fillPoly(img=mask, pts=[points], color=255)

    # mask = np.zeros(image_size, dtype=np.uint8)
    # points = cv2.convexHull(face_landmarks)  # 凸包
    # cv2.fillConvexPoly(mask, points, color=255)
    return mask


def get_affine_image(image1, image2, face_landmarks1, face_landmarks2):
    """
    获取图片1仿射变换后的图片
    :param image1: 图片1, 要进行仿射变换的图片
    :param image2: 图片2, 只要用来获取图片大小，生成与之大小相同的仿射变换图片
    :param face_landmarks1: 图片1的人脸特征点
    :param face_landmarks2: 图片2的人脸特征点
    :return: 仿射变换后的图片
    """
    three_points_index = [18, 8, 25]
    M = cv2.getAffineTransform(face_landmarks1[three_points_index].astype(np.float32),
                               face_landmarks2[three_points_index].astype(np.float32))
    dsize = (image2.shape[1], image2.shape[0])
    affine_image = cv2.warpAffine(image1, M, dsize)
    return affine_image.astype(np.uint8)


def get_mask_center_point(image_mask):
    """
    获取掩模的中心点坐标
    :param image_mask: 掩模图片
    :return: 掩模中心
    """
    image_mask_index = np.argwhere(image_mask > 0)
    miny, minx = np.min(image_mask_index, axis=0)
    maxy, maxx = np.max(image_mask_index, axis=0)
    center_point = ((maxx + minx) // 2, (maxy + miny) // 2)
    return center_point


def get_mask_union(mask1, mask2):
    """
    获取两个掩模掩盖部分的并集
    :param mask1: mask_image, 掩模1
    :param mask2: mask_image, 掩模2
    :return: 两个掩模掩盖部分的并集
    """
    mask = np.min([mask1, mask2], axis=0)  # 掩盖部分并集
    mask = ((cv2.blur(mask, (5, 5)) == 255) * 255).astype(np.uint8)  # 缩小掩模大小
    mask = cv2.blur(mask, (3, 3)).astype(np.uint8)  # 模糊掩模
    return mask


def skin_color_adjustment(im1, im2, mask=None):
    """
    肤色调整
    :param im1: 图片1
    :param im2: 图片2
    :param mask: 人脸 mask. 如果存在，使用人脸部分均值来求肤色变换系数；否则，使用高斯模糊来求肤色变换系数
    :return: 根据图片2的颜色调整的图片1
    """
    if mask is None:
        im1_ksize = 55
        im2_ksize = 55
        im1_factor = cv2.GaussianBlur(im1, (im1_ksize, im1_ksize), 0).astype(np.float)
        im2_factor = cv2.GaussianBlur(im2, (im2_ksize, im2_ksize), 0).astype(np.float)
    else:
        im1_face_image = cv2.bitwise_and(im1, im1, mask=mask)
        im2_face_image = cv2.bitwise_and(im2, im2, mask=mask)
        im1_factor = np.mean(im1_face_image, axis=(0, 1))
        im2_factor = np.mean(im2_face_image, axis=(0, 1))

    im1 = np.clip((im1.astype(np.float) * im2_factor / np.clip(im1_factor, 1e-6, None)), 0, 255).astype(np.uint8)
    return im1


def main():
    im1 = cv2.imread(image_face_path)  # face_image
    im1 = cv2.resize(im1, (600, im1.shape[0] * 600 // im1.shape[1]))
    landmarks1 = get_face_landmarks(im1, detector, predictor)  # 68_face_landmarks
    if landmarks1 is None:
        print('{}:检测不到人脸'.format(image_face_path))
        exit(1)
    im1_size = get_image_size(im1)  # 脸图大小
    im1_mask = get_face_mask(im1_size, landmarks1)  # 脸图人脸掩模

    cam = cv2.VideoCapture(0)
    while True:
        ret_val, im2 = cam.read()  # camera_image
        landmarks2 = get_face_landmarks(im2, detector, predictor)  # 68_face_landmarks
        if landmarks2 is not None:
            im2_size = get_image_size(im2)  # 摄像头图片大小
            im2_mask = get_face_mask(im2_size, landmarks2)  # 摄像头图片人脸掩模

            affine_im1 = get_affine_image(im1, im2, landmarks1, landmarks2)  # im1（脸图）仿射变换后的图片
            affine_im1_mask = get_affine_image(im1_mask, im2, landmarks1, landmarks2)  # im1（脸图）仿射变换后的图片的人脸掩模

            union_mask = get_mask_union(im2_mask, affine_im1_mask)  # 掩模合并

            # affine_im1_face_image = cv2.bitwise_and(affine_im1, affine_im1, mask=union_mask)  # im1（脸图）的脸
            # im2_face_image = cv2.bitwise_and(im2, im2, mask=union_mask)  # im2（摄像头图片）的脸
            # cv2.imshow('affine_im1_face_image', affine_im1_face_image)
            # cv2.imshow('im2_face_image', im2_face_image)

            affine_im1 = skin_color_adjustment(affine_im1, im2, mask=union_mask)  # 肤色调整
            point = get_mask_center_point(affine_im1_mask)  # im1（脸图）仿射变换后的图片的人脸掩模的中心点
            seamless_im = cv2.seamlessClone(affine_im1, im2, mask=union_mask, p=point, flags=cv2.NORMAL_CLONE)  # 进行泊松融合

            # cv2.imshow('affine_im1', affine_im1)
            # cv2.imshow('im2', im2)
            cv2.imshow('seamless_im', seamless_im)
        else:
            cv2.imshow('seamless_im', im2)
        if cv2.waitKey(1) == 27:  # 按Esc退出
            break
    cv2.destroyAllWindows()


if __name__ == '__main__':
    main()

实现步骤

1. 使用dlib的shape_predictor_68_face_landmarks.dat模型获取人脸图片im1和摄像头图片im2的68个人脸特征点。

2. 根据上一步获得的特征点得到两张图片的人脸掩模im1_mask和im2_mask。

3. 利用68个特征点中的3个特征点，对人脸图片im1进行仿射变换使其脸部对准摄像头图片中的脸部，得到图片affine_im1。

4. 对人脸图片的掩模im1_mask也进行相同的仿射变换得到affine_im1_mask。

5. 对掩模im2_mask和掩模affine_im1_mask的掩盖部分取并集得到union_mask。

6. 利用opencv里的seamlessClone函数对仿射变换后的affine_im1和摄像头图片im2进行泊松融合，掩模为union_mask，得到融合后的图像seamless_im。

 

需要下载dlib人脸形状检测器模型数据

[shape_predictor_68_face_landmarks.dat.bz2](http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2)，并解压在models文件夹下

 

项目目录

 

用周杰伦的照片换脸

 

 

参考：

https://github.com/Jacen789/simple_faceswap