dlib库标注特征点以及裁剪人脸组件

1、安装dlib

当在anaconda下使用命令 pip install dlib 安装dlib库时，提示错误安装不成功，可能是版本不匹配。在虚拟环境下查看Python版本，高版本不一定能和当前python版本兼容。可以将dlib wheel下载到本地，再手动安装。
dlib库：链接：https://pan.baidu.com/s/14FKO_wiMG85s2kgo-1aEZg 提取码：hg5m

2、标注人脸特征点

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

import numpy as np
import cv2
import dlib
from PIL import Image
import os
import time

detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
# dictory_name = "E:\code\celeba-1024"
dictory_name = './imgs'
if not os.path.exists('./eye'):
    os.mkdir('./eye')
if not os.path.exists('./nose'):
    os.mkdir('./nose')
if not os.path.exists('./mouse'):
    os.mkdir('./mouse')
# 摄像头捕捉人脸图像
def cap_face():
    cap = cv2.VideoCapture(0)

    while (cap.isOpened()):
        ret, frame = cap.read()
        frame = cv2.resize(frame, (800, 650))
        start = time.time()
        frame = cv2.flip(frame, 1)
        dets = detector(frame, 0)
        for k, point in enumerate(dets):
            shape = predictor(frame, point)
            landmarks = np.matrix([[p.x, p.y] for p in shape.parts()])
            for num in range(shape.num_parts):
                cv2.circle(frame, (shape.parts()[num].x, shape.parts()[num].y), 1, (0, 255, 0), -1)
        print("time ->", time.time() - start)
        cv2.imshow('frame', frame)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break

    cap.release()
    cv2.destroyAllWindows()
def landmark_face():
    img = cv2.imread(dictory_name + "\\" + "000085.jpg.jpg")
    # img = cv2.resize(img[:, :, ::-1],dsize=(256, 256))

    # 取灰度
    # img_gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
    # 人脸数 rects
    rects = detector(img, 0)
    for k, point in enumerate(rects):
        shape = predictor(img, point)
        landmarks = np.matrix([[p.x, p.y] for p in shape.parts()])
        for num in range(shape.num_parts):
            cv2.circle(img, (shape.parts()[num].x, shape.parts()[num].y), 3, (0, 255, 0), -1)
    cv2.imshow('image', img)
    cv2.imwrite("./imgs/landmark_face.jpg",img)
    cv2.waitKey(0)

def crop_face():
    for filename in os.listdir(dictory_name):
        # cv2读取图像
        img = cv2.imread(dictory_name + "\\" + filename)
        img = cv2.resize(img[:, :, ::-1], dsize=(256, 256))

        # 取灰度
        # img_gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
        # 人脸数 rects
        rects = detector(img, 0)
        for k, point in enumerate(rects):
            shape = predictor(img, point)
            landmarks = np.matrix([[p.x, p.y] for p in shape.parts()])
            for num in range(shape.num_parts):
                cv2.circle(img, (shape.parts()[num].x, shape.parts()[num].y), 1, (0, 255, 0), -1)
        # cv2.imshow('image', img)

        # if len(rects) == 0:

        # 眼睛+眉毛
        # x_eye_left_up = 63       # 3  23
        # y_eye_left_up = 95      # 18  18
        # x_eye_right_down = 186   # 92  50
        # y_eye_right_down = 134   # 43  43
        #
        # eye_box = (x_eye_left_up, y_eye_left_up, x_eye_right_down, y_eye_right_down)
        # eye_crop = Image.fromarray(img).crop(eye_box)
        # eye_crop.save("./eye3" + "/" + filename)

        # 鼻子
        # x_nose_left_up = 98
        # y_nose_left_up = 130
        # x_nose_right_down = 155
        # y_nose_right_down = 173
        #
        # nose_box = (x_nose_left_up, y_nose_left_up, x_nose_right_down, y_nose_right_down)
        # nose_crop = Image.fromarray(img).crop(nose_box)
        # nose_crop.save("./nose" + "/" + filename)

        # 嘴巴
        # x_mouse_left_up = 90
        # y_mouse_left_up = 173
        # x_mouse_right_down = 160
        # y_mouse_right_down = 210
        #
        # mouse_box = (x_mouse_left_up, y_mouse_left_up, x_mouse_right_down, y_mouse_right_down)
        # mouse_crop = Image.fromarray(img).crop(mouse_box)
        # mouse_crop.save("./mouse" + "/" + filename)

        # else:
        #     for i in range(len(rects)): # 超过一个人脸时
        #         landmarks = np.matrix([[p.x, p.y] for p in predictor(img, rects[i]).parts()])
        #         # print(type(landmarks))  # numpy.matrix
        #
        #         # 眼睛+眉毛
        #         x_left_up = landmarks[17, 0]
        #         y_left_up = landmarks[18, 1] - (landmarks[29, 1] - landmarks[28, 1])*2
        #         x_right_down = landmarks[26, 0]
        #         y_right_down = landmarks[28, 1] + (landmarks[29, 1] - landmarks[28, 1])
        #
        #         # 嘴巴+鼻子
        #         # x_left_up = landmarks[4, 0] + (landmarks[48, 0] - landmarks[4, 0]) / 2
        #         # y_left_up = landmarks[29, 1]
        #         # x_right_down = landmarks[54, 0] + (landmarks[12, 0] - landmarks[54, 0]) / 2
        #         # y_right_down = landmarks[57, 1] + (landmarks[8, 1] - landmarks[57, 1]) / 2
        #
        #         # 眼睛+眉毛+鼻子
        #         # x_left_up = landmarks[17, 0]
        #         # y_left_up = landmarks[18, 1] - (landmarks[29, 1] - landmarks[28, 1])*2
        #         # x_right_down = landmarks[26, 0]
        #         # y_right_down = landmarks[33, 1] + (landmarks[51, 1] - landmarks[33, 1]) / 2
        #
        #         # 眼睛+眉毛+鼻子+嘴巴
        #         # x_left_up = landmarks[17, 0]
        #         # y_left_up = landmarks[18, 1] - (landmarks[29, 1] - landmarks[28, 1]) * 2
        #         # x_right_down = landmarks[54, 0] + (landmarks[12, 0] - landmarks[54, 0])
        #         # y_right_down = landmarks[57, 1] + (landmarks[8, 1] - landmarks[57, 1]) / 2
        #
        #         box = (x_left_up, y_left_up, x_right_down, y_right_down)
        #         image_crop = Image.fromarray(img).crop(box)
        #         image_crop.save("./eye3" + "/" + filename)

if __name__ == '__main__':
    # cap_face()
    # crop_face()
    landmark_face()

调用landmark_face()，标注图像结果如下：

3、裁剪人脸组件

crop_face()方法：采用的固定大小的尺度去裁剪人脸组件（眼睛+眉毛，鼻子，嘴巴），使用PIL保存裁剪的结果，如下图所示：
眼睛：



鼻子：



嘴巴：



PIL库中的crop方法：采用两个像素点的坐标（x,y）去裁剪组件，两个点的位置分别为矩形框的左上角和右下角。注意x的方向为向右，y为向下，图像左上角的起始坐标为（0,0）。

4、后记

思考：借助dlib库检测人脸的特征点来辅助裁剪人脸组件，存在哪些问题？

• 当人脸有姿态和遮挡时，采用固定size的尺度裁剪，无法精准裁剪。优点是裁剪后大小一致，以便网络训练，以及进一步处理。
• 根据标注的特征点，利用其中的几个特征点来crop人脸组件，特征点中位置要进行加减一定的像素，但是裁剪后的图像大小不一致。当输入给神经网络时，需要resize处理，这样的裁剪好处是能将人脸组件精确裁剪下来。

由于作者水平有限，文中若有不正确的地方，欢迎大家指出，若有任何问题，请在下方讨论。