微表情识别的图片预处理(python版)

---

前言

微表情识别类似于表情识别，在处理时的不同之处在于微表情的关注区域是特定的(AU)，所以我们在裁剪人脸的时候也是主要获得该区域，我们实验室之前用的都是matlab代码，刚刚把它转为了python代码。

一、人脸旋转

方法就是采用dlib68模型找到双眼的位置，计算需要旋转的角度，然后做仿射变换，代码如下

import cv2
import dlib
import numpy as np


# 脸部旋转函数
def face_rotate(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y):
    center_x, center_y = int((left_eye_x + right_eye_x) / 2), int((left_eye_y + right_eye_y) // 2)
    dx, dy = right_eye_x - left_eye_x, right_eye_y - left_eye_y
    # 计算角度
    angle = np.degrees(np.arctan2(dy, dx))
    # 计算仿射矩阵
    warp_matrix = cv2.getRotationMatrix2D((center_x, center_y), angle, 1)
    # 进行仿射变换，即旋转
    rotated_img = cv2.warpAffine(img, warp_matrix, (img.shape[1], img.shape[0]))
    return rotated_img


# 脸部剪裁函数
def face_crop(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y, noise_y):
    x_min = left_eye_x - (right_eye_x - left_eye_x)
    y_min = int((left_eye_y + right_eye_y) / 2 - (noise_y - (left_eye_y + right_eye_y) / 2))
    width = 3 * (right_eye_x - left_eye_x)
    height = 3 * (noise_y - (left_eye_y + right_eye_y) // 2)
    return img[y_min:y_min + height, x_min:x_min + width]


# 加载模型
predictor_model = r'C:\Users\S\Desktop\python\model\dlib\shape_predictor_68_face_landmarks.dat'
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor(predictor_model)

# cv2读取图像
test_film_path = r"C:\Users\S\Desktop\img1024-model\0082.png"
img = cv2.imread(test_film_path)

# 取灰度，检测人脸
img_gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
rects = detector(img_gray, 0)
for i in range(len(rects)):
    landmarks = np.matrix([[p.x, p.y] for p in predictor(img, rects[i]).parts()])
    left_eye_x, left_eye_y = landmarks[39][0, 0], landmarks[39][0, 1]
    right_eye_x, right_eye_y = landmarks[42][0, 0], landmarks[42][0, 1]
    noise_x, noise_y = landmarks[33][0, 0], landmarks[33][0, 1]
    rotated_img = face_rotate(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y)
    # crop_img = face_crop(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y, noise_y)
    # cv2.imwrite("rotated_img.jpg", rotate_img)
    cv2.imshow("source",img)
    cv2.imshow("rotated_img", rotated_img)
    # cv2.imwrite("croped_img.jpg", croped_img)
    # cv2.imshow("croped_img", croped_img)
    cv2.waitKey(0)

效果：

二、人脸裁剪

方法就是根据双眼和鼻子的位置，向左右各扩展双眼的长度，向上下各扩展眼睛到鼻子的垂直距离，代码如下：

import cv2
import dlib
import numpy as np


# 脸部旋转函数
def face_rotate(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y):
    center_x, center_y = int((left_eye_x + right_eye_x) / 2), int((left_eye_y + right_eye_y) // 2)
    dx, dy = right_eye_x - left_eye_x, right_eye_y - left_eye_y
    # 计算角度
    angle = np.degrees(np.arctan2(dy, dx))
    # 计算仿射矩阵
    warp_matrix = cv2.getRotationMatrix2D((center_x, center_y), angle, 1)
    # 进行仿射变换，即旋转
    rotated_img = cv2.warpAffine(img, warp_matrix, (img.shape[1], img.shape[0]))
    return rotated_img


# 脸部剪裁函数
def face_crop(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y, noise_y):
    x_min = left_eye_x - (right_eye_x - left_eye_x)
    y_min = int((left_eye_y + right_eye_y) / 2 - (noise_y - (left_eye_y + right_eye_y) / 2))
    width = 3 * (right_eye_x - left_eye_x)
    height = 3 * (noise_y - (left_eye_y + right_eye_y) // 2)
    # 下面这个返回的是完全按照规则裁的
    # return img[y_min:y_min + height, x_min:x_min + width]
    # 下面这个，得到的是一张长宽相同的图
    croped_img = img[y_min:y_min + height, left_eye_x + (right_eye_x - left_eye_x)//2-3 * (noise_y - (left_eye_y + right_eye_y) // 2)//2:left_eye_x + (right_eye_x - left_eye_x)//2+3 * (noise_y - (left_eye_y + right_eye_y) // 2)//2]
    # 这里resize有两种方法可以得到深度学习所需要的数据，一是将后面长宽不同的图拉伸为224(一般是224)；二是将后面长宽相同的图简单resize一下，这里设的是128，有特殊用途，一般是224
    resize_img = cv2.resize(croped_img,(128,128),interpolation=cv2.INTER_AREA)
    return resize_img

# 加载模型
predictor_model = r'C:\Users\S\Desktop\python\model\dlib\shape_predictor_68_face_landmarks.dat'
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor(predictor_model)

# cv2读取图像
test_film_path = r"C:\Users\S\Desktop\train\\0720.jpg"
img = cv2.imread(test_film_path)

# 取灰度，检测人脸
img_gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
rects = detector(img_gray, 0)
for i in range(len(rects)):
    landmarks = np.matrix([[p.x, p.y] for p in predictor(img, rects[i]).parts()])
    left_eye_x, left_eye_y = landmarks[39][0, 0], landmarks[39][0, 1]
    right_eye_x, right_eye_y = landmarks[42][0, 0], landmarks[42][0, 1]
    noise_x, noise_y = landmarks[33][0, 0], landmarks[33][0, 1]
    # rotated_img = face_rotate(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y)
    croped_img = face_crop(img, left_eye_x, left_eye_y, right_eye_x, right_eye_y, noise_y)
    # cv2.imwrite("rotated_img.jpg", rotate_img)
    cv2.imshow("source",img)
    # cv2.imshow("rotated_img", rotated_img)
    #cv2.imwrite("croped_img.jpg", croped_img)
    cv2.imshow("croped_img", croped_img)
    cv2.waitKey(0)

效果：

这样就基本包含了常见的au区域，可以送进网络进行训练了